That Lasso app, may be a good work around for older games with DRM that flakes out when bounced between P-cores and E-cores, just a thought. Great vid der8auer and crew. B)
Unfortunately, that will almost certainly not work because DRM is typically at the level of kernel modules (otherwise you could just run some user level software that fakes the runtime environment to fool it)
You should try disabling the P cores and seeing how far you can overclock the E cores by themselves. How close can an overclocked E core get to the performance of a P core when it has the thermal and voltage headroom?
I do use PL to set all background tasks to the lowest cores in my 5900X leaving the upper faster cores for gaming and productivity apps. You don't have to go options > add to list.. Just right click the selected apps and choose an "always" affinity. At the end you can backup all your settings. Problem with CB is that it adjusts the affinity by itself if I recall correctly.
@@ThunderingRoar I haven't done that since one core is too little for some of them such as iCUE I leave all my light background (non time critical) tasks such as TaskManager, HWInfo, RainMeter, Walpaper Engine (I only play a video with it), Oculus tray tool, Skype, Discord etc... to the last 6 logical cores. (18-23) Things like Chrome go on 8-17 More time critical gaming related background tasks such as Virtual Desktop / Oculus wireless streamer services / apps go to 6-23 so they can use the first CCD while gaming (lower inter CCD latency) Games and heavy productivity apps go to 0-23 so they can use all cores and most importantly, run the heaviest threads on the first cores. I have tested with and without PL. While the Windows scheduler in Win10 does a great job you can eliminate any microstutter from games and chrome by assigning chunks of your CPU cores so they don't conflict. Most of the windows tasks and services are set to use all cores. Some of them aren't supposed to have affinity. I can say the Task Manager, Print Spooler, Gaming services can have a custom affinity without issues. Not that I encounter any when I tried to set affinity to a lot of windows tasks anyway. Things like DWM need to be snappy so I think is better to leave alone. Anything system critical should be left alone IMHO. Also "ProBalance" off or anything related to Power managment. Those things are for really slow machines. You wanna use PL exclusively to manage affinities.
@@SaccoBelmonte I'd love to see some channels do some extensive testing on win10/11 scheduler performance differences vs manual affinities regarding latency, stutter etc in all kinds of scenarios-- so far I'm having a hard time finding any in win11 on my 5900x build
@@interlace84 There are none. The scheduler is a rather obscure process running behind scene and is hard to know (if even possible) what's going on. I think the current CPPC mask (and CPPC preferred cores) in Zen3 is incorrect in Win11 because it exposes a strange behavior in two of the lowest cores on my 5900X. I have them both disabled for now and rely on PL to set affinities around. PL's latest update added CPU Sets....I need to dig a bit deeper into that. Sounds interesting though.
So, it is just as I imagined - disabling the P-cores basically turns this into a i7-7700k, though with a 65w TDP, essentially confirming what they are - efficient.
7700k is four cores. There are eight economy cores on this CPU. He said single economy core performance is about the same as one core of 7700k. ... Surely that makes the eight E-cores like two 7700k stuck together.
@@yancgc5098 Not really as it took 8 e cores to even match and yet still be behind a 7700K and that is double core count. And no hyper threading is not anywhere near the same as actual real cores. Hyper threading smoothes out multi tasking, it does not seem to give more raw performance rather more efficient utilization of processes that do not fully utilize a CPU core and in theory if 2 threads can use 50% or less of one CPU core, it can divide the threads for better use., But that is no where near the performance of an actual core so no a good comparison.
Interesting observation! So if Intel would make e-core only CPU it would have about much higher multicore performance. Or maybe 4 p-cores + 24 e-cores :)
@@dmitrylozenko7123 I’ve been wondering if they’ll do this for mobile class CPUs. Maybe 2 or 4 P-Cores but stack up and have 8 e-cores. Have more efficient cores to run most tasks but having powerful cores available for heavier tasks. I wonder if this could be a leap forward in battery life.
maybe for 7-15watt chips it's good, but 2 powerful cores and 8 slow cores are not really beneficial especially since modern ultrabooks can actually game
@Jesse Schultz AMD doesnt come close to dominate, what are you smoking? in terms of best cpu for that (and which "that"? different segments of mobile exist)? They certainly are not the sales leader. YOu have to do more than beat intel for on year or so - you have to start being truly competitive in ALL ways for many years.
Would love to see a single core benchmark 4GHz p-core vs 4GHz e-core. Another awesome video that doesnt mirror the same garbage other tech youtubers keep showing, THANK YOU so much for good content.
@@saricubra2867 you mean lowest-tier 6000-series without hyperthreading? and only in the 12900k? Get a 12700k and prepare to be very disappointed. They won't clock much further than base either..
I always maintained that the Gracemont cores are a brilliant piece of engineering. All 8 of them only consumes around 50W under full utilization and are as fast as a first generation 8 Core Ryzen.
For me it's kind of disappointing since the IPC is lower than of skylake and just marginally faster than haswell. that's 2013-2014 territory there. if you compare a stock 12900k and compare the stock E and p cores, the P cores performance per core is twice as fast but the P cores dont even consume twice the energy, so in that regard the p cores are more efficient than the e cores
@@Nemesis1ism a processor is first of all not only the cores, there are a lot more components than cores, but 1 e core consumes 25watts of energy on its own already.
I agree. The whole point of the E-cores was to give smaller programs that won't really see a benefit to the bigger E-cores a place to run, like Discord, browsers, mail clients and office applications, MP3 players, programs for LEDs and other ancillary hardware (like iCue from Corsair), etc. None of those uses put any demanding strain on the CPU's. I could see an argument for more P-cores for use cases like CAD/CAM, media editing, modeling software (things like fluid dynamics, crash tests, strain loading, etc.), etc. But even then, just to chuck the processes for the various Windows systems onto the E-cores would be a godsend. Plus, it's a way to add a lot more cores without really increasing the price. (Looking at you, Threadripper.)
A random thought I just had was that it would also decrease the likelyhood of a random process freezing one of your main cores and leading to unresponsive OS or IO
@@Kyrator88 We have multicore processors now, and that logic would apply to those as well. I don't think that's something an e-core would necessarily prevent. But it's a good though.
@@interlace84 That's still 8 threads in total. Although the point is not to compare with existing skus but seeing the opportunity on use case scenarios where efficiency is a major part of the constraints. Intel could totally make a chip only populated with ecores that can crash a M1pro and even beefier iterations in workstation/server applications and still keeping a small form factor or a passive cooling design.
Damn, that really puts into perspective just how old and slow my CPU is! I thought my 4790K was still fast, but it only gets about 410 Single Core and 2000 Multi Core in Cinebench R20! Just the E Cores in this chip are faster than my 4790k!
@@kendokaaa if I can play 1080p 60 with my setup, any upgrade will do better, so it's imo still a good example. Of course if you want 170hz+ hyper mega ultra settings full rtx that's not relevant, and you probably will want the latest CPU. Of course an old cpu like that will be a bottleneck for a 3080, but for a 1650 it probably won't.
@@kendokaaa I just replaced my oc'd 3770k with a 5900k-- both systems had equal 4k performance with my rtx2070 at 4k maxed out in several titles.. both playable so the bottlenecks aren't always CPUs. Lower resolutions got major increases though. @Daniel You're comparing your quadcore vs 8 E-cores that are only in the 12900k-- the 12700k ones are instantly halved already. Don't sweat over it.
Very good video. People forget that E-core is comparable to Skylake or Zen, Zen+ . It might be hard to believe - but it is true. Calling it Atom was a bit of mistake IMO, but otherwise ingenious move.
I love that you always manage to have interesting and unique content like this! It always keeps me coming back. Plus you do almost everything in English and German.
Interesting piece Roman. Why would people think the e-cores are so shite, that's what I don't get. As you said, makes perfect sense to have CPUs like this to assign the appropriate type of core to the appropriate type of process/workload.
Sounds like unverified rumors... Just people randomly spouting things and repeating it without testing it first.. I figure the names "performance" and "efficiency" probably add a bias without knowing how they work.
I saw another reviewer who did similar tests and his conclusion was very different. Maybe it's the software he used to separate the cores but he has much less then half the fps in games and multi threading capability was awful. This review gives me renewed hope in those little e cores.
E cores only make sense in a device that runs on a battery/small device. In a desktop that already sucks up hundreds of watts it doesnt make sense. a 10core 20 thread p core only chip would be even plenty faster than a 16 core 24thread 8p+8e core config.... It's intel just not being able to get a better efficiency chip that's why they had to introduce this technology to somehow keep their chips from melting themselves.
@@SweatyFeetGirl before I watched this video I would've agreed with you, but after looking at his data I'd prefer the 8+8 combo over a 10 core in basically any situation.
people just think its bad because of a handful of cases where the thread director decides to put the load on e cores instead of an unused pcore optimally, leading to a potential drop of performance. It's the same idea and situation as when hyperthreading was introduced, and games back then during sandy bridge which used 4 threads or less ideally loaded it onto the physical core rather than the hyper threads, and would see a performance game by doing so/turning hyperthreading off. Of course though, over time, that sentiment ages badly
I wonder if project lasso could be used to fix alder lake compatibility issues with certain apps on Windows 10. If those issues were caused by jumps to the E-cores, disabling E-cores for that task could solve the issue.
I haven't used mine much but I took a win 10 installation that has seen like 10 chip sets, and it used the cores on my 12600k very well a few performance cores were running during gaming. And all other cores at 0% useage. And it used much less power than my 5900x about 50-70% less,(70-100w for the 5900x and about +/-35W for the 12600k. And despite much worse cooling that the 5900x has the 12600k wouldn't go over 70c gaming and 30c idle while the 5900x idles at 60-70 and goes up to 90 while doing light gaming. It was also getting more fps in nfs heat with a worse ram(not much worse both 3600mhz but ryzen had dual rank ram with better timings). So all things considered very happy with Intel's new CPU.
This video of yours is pure gold! For that Lasso Pro tool alone already. I used a BIOS feature to park all E-Cores, on a key-press, so my x265 processes could run full speed. Now, with Lasso, I can just permanently set affinity to just all P-Cores, and still have my E-Cores active for everything else. :)
Comparing just ST to ST, E-cores have about 60% of the performance of a P-core. Especially given the substantial clockspeed difference, that's pretty nuts! Very cool video.
A way I explain hybrid cores... it's like optane. You have low cost (in silicon) e-cores to do less interesting tasks. P cores = cache for high volume tasks. If you want 100% p-cores, you get a lot less compute/storage for your dollar/silicon. 100% e-cores will suffer in certain task models.
My concern has never been with the processing capability of the E cores. I don't have much confidence in schedulers doing the "right thing" with processes. We've been living in an SMP world for so long. Yes, ARM asymmetric SoCs have been around for a long time, but that's in mobile workloads, not desktop. We shall see, though.
Sure ARM improved that tech during a decade but you seem to neglect that's not just mobile workloads but embedded workloads as well with much higher time constraints and those worloads run on more slower processors and memory/cache than laptop/desktop so they need all the resources available as well as the optimizations. The reason ARM heterogeneous arch works well without any special scheduler is because they mainly use the Linux kernel which was long time ago tailored for NUMA multi CPU servers and doesn't have the tendency to randomly switch process execution from a thread to another. The scheduler just need to be as simple as the Pstates frequency governor : prioritizing performance or efficiency by launching a new processus on one type of core according to the processus priority then moving it from the other type depending on its resource usage. And It just works.
I totally agree. Just not a good design decision. We have been in an SMP world at the low level Windows NT flavors and even Linux and the like schedulers for so long in the X86 ecosystem that Intel's approach to this kind of change just does not make sense. And thus I have no confidence in schedulers doing right things with processes designed for same type of CPU core count and type as well as a massive library of software in the X86 ecosystem. This is a much bigger and more radical change that can cause much more issues than any of the other changes we have seen the last 20 years or maybe more. I mean chipset changes, CPU arch changes like pipelines and IPC and core counts, just new drivers and such at low level. And scheduler changes in Windows not an issue as SMP has been around since like 1995 with Pentium Pro. Switch to dual core CPUs. Not at all an issue as a dual core CPU was no different to Windows NT and Linux scheduler than a dual socket CPU system as all dual core was was 2 CPUs in one chip. Likewise Quad cores 4 CPUs in one chip as opposed to quad socket single core CPU mobos. Not an issue for SMP world. But a hybrid arch like this and right away not good. Nothing wrong with trying something different, but Intel making a mistake locking SKUs to 8 P cores only and not offering separate ones with no e cores that have 8 or more. Cause it is going to be a long long long time if ever hybrid arch catches on or is good for the desktop.
Well if you are gonna make this argument then you should try running the game on just 2 P-cores and see if the result is better than running on all the E-cores.
Great video and great thing that you showed Process Lasso! I am a proud pro user since 2014, it gave a new dimension to process management on Windows, and helped me plenty of times with pro audio related stuff and games. Great little piece of software.
@Roman, a reason for P cores instead of 80 e-cores would be that at some point, the interconnects either take too much space, or become too slow. It's a balancing game.
Interesting suggesting 6P + 16E. I came to the same idea previously. Also, process lasso could be used to fix compatibility issues with programs that don't work on hybrid architecture without disabling E cores in bios, or waiting for scroll lock hack which likely does same thing on the fly.
Re: Process Lasso. I've got a 5800X, I use it for gaming, so I've got two sort of "extra" cores in there (Not all games can take advantage of eight cores and sixteen threads), so I used Process Lasso to put all my background software on my two slowest cores. Backup software, iCUE, KeePass, Razer Synapse, all that extra -stuff- hangs out on cores five and seven, which leaves the rest of the cores wide open for whatever. I don't know if it provides any real benefit, but it's fun to play with!
It might help or it might hinder, I'd benchmark it. I tried this both on my 2700X and my 3900X and performance usually ended up worse by a small amount
Man. A die full of E cores would make a heckuva virtualization lab, especially without hyperthreading. Here's hoping we do see the 6x16 arrangement you mentioned. 👍️
I'd love to see this pushed to the absolute max, a CPU with just one or two P-core of the fastest possible design for running single-threaded tasks as quickly as possible, and the rest as E-cores.
Yeah, I agree. This seems like the way to go. People run highly parallelizable codes on GPU's but some codes are really difficult to implement to GPU. This would be the perfect solution. Although I think I'd prefer something like 4 P cores and the rest E cores just because I have some codes that I want to run with 4 cores super fast but most of my codes are as many cores as possible.
This is super cool, I can definitely see some situations like power saving on a laptop device while gaming which would preserve battery life but still offer very decent performance.
9:26 Atom cores from 2016 were good, you just didnt get many of them and they were usually clocked REALLY low, and that's why almost all atom processors were terrible. Like, if you had gotten 8+ Atom cores clocked to 3.2Ghz instead of 2-4 cores at 1.8Ghz, they'd still be pretty efficient but way better performance. because below 3.2Ghz there is a diminishing return on power savings compared to performance loss on most modern silicon. This holds true on both my i7-8750H, my 3950x, and my 4650G. Oh, and thats probably why the maximum clocks for the M1 Mac Mini is 3.2Ghz
it's generally a problem with nodes. as a reference, i can run my 3700X at 1.175 Volts at 4.1ghz but for 4.2ghz i need 1.25Volts, that's 6% more voltage for 2% more speed - it might seem low, but the temperature spike and wattage increases a lot for diminishing returns.
Being able to hit high clock speeds is part of a processor architecture's design. Atom cores could not and it's absolutely valid to criticize them for that.
Lol no, intel atom is really bad even the clock is higher, intel atom Denverton clocked at 2,2 ghz, and geekbench score is only 300-400, intel xeon from sandy bridge with same clock score around 500-600, and skylake still score around 700 with same clock speed, intel inside skylake mobile processor is not atom processor
@@adiirfan01 Thats sort my point though, that Denverton Atom C3858 2Ghz 12 core at 25w used closed to 1/3rd the power of an E3-1245 V6 75w(kabylake both 14nm so neither had a node advantage) but the atom was closer to half the performance instead of 1/3rd. I dont remember which model, but there was an atom that had 2 or 4 more cores, and ran at... 2.5Ghz and only went up to 30w. This is what i was talking about there being a diminishing return of power savings They added 2-4 more of these same atom cores(15-30% more), bumped up the clocks by 25%, but it only used 20% more power? It should have used 45% more power if this stuff scaled linearly. There seems to be a bell curve with power per watt peaking at a certain point in the silicon's range. With Apple M1, and AMD Zen2 it seems to be 3-3.3Ghz(zen2 seems to like around 3, where apple might peak around 2.8Ghz. If i had to guess, Denverton's curve peaked at 2.8Ghz as well, if it could have reached 2.8, but who knows, In tel never offered higher clocked atoms, so 2.5 might have been the peak of the efficiency curve for that arc. The atoms main problem from my memory was the small cache, yes adding cache would have drove up power draw, but it should have increased performance by more. IIRC Denverton had a total of 1MB of cache per core, where KabyLake had 2MB of just L3 cache per core. Also Intel just needs more cache for everything. Ever wonder why the i7-5575C would sometimes beat everything but Alderlake and non APU Zen3 even at its tiny ~60W power draw? It was because of that massive L4 cache(128MB of eDRAM), allowing that 3.7Ghz processor to beat 5.3Ghz rocketlake in several tasks at sometimes 1/4 the power(definately not most tasks, but enough to ask why didnt we continue using this and improved the eDRAM for lower latency and higher bandwidth/size). If intel had added one or more of those eDRAM chiplets to all of their top end processors, with half the latency of DDR3, and 1/3 the latency of DDR4 AMD have had a much harder time climbing to the top. I'm really interested to know how something like KAbyLake G would have performed if the CPU were allowed to use that 204GB/s HBM with slightly higher latency ~100ns, instead of the 30-40GB DDR4 with ~80ns of latency. Is bandwidth king for efficiency, or latency, or both
I must say I was definitely sceptical of e cores. Can easily see the value in having them on a laptop or such like with a battery but figured they'd be a waste on desktop. But its nice to see they're performing better than I expected. And its really good to see this kind of demonstration 🙂
Waste on desktop? E-cores are there for higher MT performance. You would get less performance if all cores were P-cores with same silicon space and power envelopment. Raptor lake will double number of E-cores and add zero P-cores.
@@kognak6640 I know, thats why I said that its nice to see them performing better than I expected. I meant before they launched I thought they would be a waste.
back in the core 2 quad days there was a minor penalty for running cross die. So I would always tell games that had the lowest performance to run exclusively on the 2nd die. This way that penalty would be eliminated. Back then games usually only had 1 or 2 threads. Lasso would have been nice for this since I had to do it every time.
If I heard correctly, E-cores are about as fast as the original skylake generation, the 6700K and alike. And having 8 of them.. That's quite a bit of power to have. Definitely worth having!
Excelent video! I really would like to see more E cores too. Few important tasks demands a good single thread performance. I don't know the reason why intel didn't go for more E cores. I would prefer 4 P and 24 or even 16 E cores.
Performance headlines. If they’d used more E-cores in the 12900K and less P-cores, it wouldn’t show up very well in benchmarks. Techspot did some interesting research and showed that for gaming, it would be best to only have P-cores. So Intel balanced it out to take the important ‘gaming crown’ with the P-cores yet put enough E-cores in there to also be competitive in heavily multi-threaded tasks.
They were meant to be used exactply that way, that's why raptor lake will have 16 of them while keeping only 8 normal ones - to compensate MT performance and surpass Ryzen with them for dirt cheap in production.
thank you for this, love this. Ive been doing thread pinning for so long. glad you tried it. maybe graph out power usage comparing P to E core runs....??
when it comes to multithreading, more lower power cores is better than few high power cores, and this idea of P and E cores allows you to have as close as the best of both worlds, high power low threaded apps and the highly threaded apps. Also apps can be made to use specific cores, so it might have the most intensive part on the P and the less intensive on E cores.
Hell yeah, exactly what I thought is was. Thanks for the testing and Lasso recommendation. 6P cores can me pretty fine, I'm glad the E cores are treated like bonus performance in pricing too
but having a low-end die with only P-cores is not ideal for the business workstation and the SFF markets. 5p (10t) +4e on that die could have allowed a better versatility.
This is really neat. Hmm.... Seems like it might be a bit of a waste if you only use the E-cores nearly exclusively, but the whole being able to bind specific applications to specific cores with Process Lasso on a permanent basis is certainly food for thought. Thank you for this video.
Thanks derb8auer for this highly intriguing and informative video. It is amazing how Intel made such efficient cores for the Alder Lake architecture. I am probably wrong but, last time I checked, I think both the P cores and the E cores are on Intel's 10nm process which helps contribute to the hybrid design. It is definitely a step in the right direction since Apple also has a hybrid CPU design too.
Well now this makes more sense. I got that the e cores are skylake but more efficient cores from the get go (leaks?). The fact that they're a 4th-5th of the size of a p core yet are about 30% powerful was new and made more sense.
Basically E-Cores seem like cores that miss AVX-512 instructions that can be used only for some kind of tasks that are likely optimized to run even better on GPU anyways (such as image/video processing) and they do not have hyperthreading. I saw some interesting talk about if it's better to have these features and just to have much more rather simple cores, cause in real world most applications will benefit from the second approach. On the other hand as a software developer I have to say, that scaling to many cores is somewhat problematic: if you have task running 48ms on single core, you distribute it to 16 threads on something like Ryzen 2700x, you may expect that it will be done in 3ms. Sadly OS decides it needs one thread for itself and two for another apps - so 13 threads will complete task and remaining 3 task will complete afterwards. In the end you may see that task lasts from 5 to 8ms and CPU usage is 60%. For tasks that run longer, it's not and issue, cause they will complete roughly at the same time.
They should've just made adjustable threading and c-stepping where it decides which applications need more power and which can be divided up among cores so while gaming u could use all cores max turbo or if your running browser or streaming in background it divides that among one or two cores using threads maybe stepping down to save power
I would appreciate it if the separation of the two types of cores were better though. They still share the same L3 cache, which means that the ring ratio has to be set lower unless you disable e-cores. The ring ratio affects the frequency of all cache including L1 and L2 cache which are not shared between the two types of cores. Ideally, the L3 cache is separated with the option to have different ring ratios. Or, there's the separate option to set a different frequency for the L1 & L2 caches of each type of core and one for the shared L3 cache. Then there's the whole AVX-512 fiasco. The P-cores have the ability to execute AVX-512, which physically takes up space on the CPU doing nothing (other than maybe improving thermals, but then it is literally a glorified and suboptimal heat sink). It definitely is possible to have the P-cores support AVX-512 while having the E-cores running. Intel worked hard on its Thread Director controller and working with Microsoft on Windows Thread Scheduler, so why stop there? Why is a unified common instruction set necessary? These two problems illustrate the problems with e-cores, which is that they handicap the performance and features of the p-cores due to their lack of true separation, forcing Intel to handicap everything they share (things like the L3 cache and the common instruction set of the CPU).
Ian Cutress over at AnandTech said that the E-cores have up to 50% the performance of the P-cores. Also, Golden Cove has a 43% IPC uplift from Skylake. Add higher clocks and you roughly have a 50% performance boost over SkyLake. Is it a coincident, that the Gracemont cores also have 50% the performance of Golden Cove. It makes sense because he tested a i7 6700K/7700K at 4.2GHz and found that the 8 E-cores on the i9 were basically the same performance. So, intel was right about the new Gracemont cores having the same IPC of Skylake. Using just CBR20 to get a performance percentage isn't going to show the whole picture. Check out AnandTech's review on the 12900K. He's done extensive testing on this, including integer and floating point.
I spent some time going through old submissions at HWBot to try to get a better idea on where the E cores sit simply in terms of performance and you are right about CBR20 not showing the whole picture. While it's true older Intel cpu's were not very good at Cinebench performance in general it's not like they were terrible. Since Roman used CBR20 just taking a look at those results his 8 E cores scored 3203 in multi core. Looking at the 7700k (Sure it's not a perfect comparison with 4c 8t vs 8c E ) it needs to be clocked at roughly 5925MHz to match them. Looking at the 7600k ( 4c 4t ) at it's base clock of 3800MHz you get a score of around 1600 cb so putting 2 of them together to get 8 cores makes it equal to Romans example using 8 E cores, both scoring 3200 cb. Pretty wild huh
I think the reason for low E core CPUs is because its new and software is not use to E cores yet and having the higher P core, the CPU will work mostly on their own. As software and such gets use to the E cores then they will introduce more of them. Need to find out the full spectrum difference between both types, not just performance in games but other aspects. Finding the right balance between them might be more dependent on what you are doing than just more of them. We are starting to get to that wall of cooling these down vs the power and performance and finding the balance for a workload will be more ideal than trying to cool off more of P and E cores in bigger CPUs
4 e cores are the size of 1 pcore, so it would be better to have a 10core 20 thread processor with only p cores instead of 8 p cores and 8 e cores and 24 threads
@@jayden974 e cores are just a waste. Ipc of ivy bridge, nuff said. They're not useful for anything. Intel should just focus on more efficient designs with high performance cores, just like AMD does. They can have 5 million e cores, it still won't matter as they're useless for gaming. And alder lake is advertised as gaming processor
I can understand why some people would have the preconceived notion that the e-cores have poor performance; as someone who has used netbooks and NUCs equipped with atom cores its encouraging to see that its not true, and that they are perfectly usable. I myself hope that hyperthreading can be added to the e-cores in the future as from what I understand its then able to use parts of the core for other tasks at a given moment that would otherwise sit dormant during a processor cycle. Thank you for covering this aspect of Alder Lake.
Adding HT would add performance, but also power and size, partially negating their benefits. To add overall performance, I like the 6P + 16E core idea better. A side note, rumor is that Zen5 will use Zen4 Dense cores as it's "little" cores. I forget if they will have SMT, but I would guess they would. Bit of a different comparison though, Zen 3 cores are about half the size of a P core, though not that much less IPC if you normalize for clockspeed.
@@xeridea An old Intel white paper claimed that adding SMT only increased die space by 5%. This was back in 2002, and the number might have changed substantially in the two decades since. But with the 2023 Meteor Lake projected to be Intel’s second consumer MCM in recent years (after Lakefield), and the adoption of a substantially denser node, Intel should have a lot more flexibility when it comes to core complexity. I don’t think Intel will back down from 8P any time soon. Current gen consoles have 8 decently performant cores, so high end desktop CPUs should have at least the same amount of full performance cores. I don’t think there’s official confirmation on SMT support for Zen4c. But if it is based on the lean cache APU version then it probably does. If measured by core area alone (sans MLC, LLC, interconnect agent), Golden Cove (5.65sqmm) is almost exactly twice the size of Zen2 (2.83sqmm), or around 1.74x the size of a Zen3 core. I don’t have any numbers for Zen cores with cache (cache takes up a substantial amount of space on the CCD), but at the same frequency there is a ~5% difference between the ST performance of APU version and the MCM version (5600X ST is around 5% better than the 5900HX, both have a specified boost frequency of 4.6GHz). Actually, adjusted for frequency, a Golden Cove core does still have a substantial performance lead over a MCM Zen3 core. Both the 12600K and 5950X have a specified boost frequency of 4.9GHz, the 12600K has a 18% lead in Cinebench R23 ST; 17% lead in Cinebench R20 ST.
@@DGao-zz5vq Yeah, I know HT doesn't take up a lot of space, but it is some, and may adversely affect the chip layout. Future nodes would alleviate this. There is another drawback, in that it hurts ST performance. This may be less of an issue given the efficiency nature, and background tasks, but it could abnormally slowdown some tasks if not properly written for many different speeds of cores. For instance, often MT is done in code by simply splitting up the screen into tiles. If you just do one large tile per core, slow cores, especially if HT/SMT would be a big hinderance. Using more tiles mitigates this, but for some quick tasks, the overhead of splitting up further may not always make sense. This may affect only a small number of apps though, I have no idea, and likely it would be improved in time. I was thinking about how historically, SMT has been a way to get more performance out of big cores, while still allowing to oversize certain parts of core to maximize ST. Just speculation though. My reference to Zen 5 is from Moore's Law is Dead video. IIRC he was referencing Zen 4D, not 4C, apparently they are different? Still speculation, 1-2 years out. I was thinking around 10-15%, assuming a minor clock difference, I thought 5950X boosted around 4.6, since it has 16 cores and lower power budget. If it boosts 4.9, different story, I didn't look up any numbers, was from memory. I was referring in general to diminishing returns on IPC for increasing core size, which is why Intel has made the little cores. At some point, which I would say we have reached, ST is fast enough for most apps, and apps are increasingly multithreaded. I think 6P + 16E would be a great production focused CPU.
Most of the "space" in a core is used by AVX instructions AVX256 and AVX512 takes up soooo much space and use soo much power. The e-core is just a P-core without these heavy instructions, just basic x86 instructions with some of the older extensions which arent heavy.
so I can have AMD 16 cores 32 threads, or Intel 16 cores 24 threads with some slower cores if i want to compile something... and it makes perfect sense for any program to run on any core so we dont have to worry about choosing what belongs where or what might suddenly need more powerful computing. the compamies are cheaping out by not giving us the power to scale freely.
The "leaks"/etc. for 13 series (from "i5 13600K -> i9 13900L) have [6/12 + 8], [8/16 + 8] and [8/16 + 16] core setups where [P/Thread + E]. I think the i7-13700K will have a nice sweet spot for P cores, P threads and E cores.
So it's less about efficiency (on PC), more about the core count, since the system with both types of cores can rely on P-cores to maximize single-threaded performance and E-cores to add more workers in multi-threaded tasks. Since power efficiency on PC makes no sense, but versatility does.
The reason I'm not sold on the E-core concept is that most games I play just can't use that many cores. Let's say a certain game only supports 8 cores (afaik, current gen consoles only got 8 cores, so to my intuition it makes sense a lot of games will be coded accordingly); Why would I want 8 strong cores with some weaker cores on the same die when I could have had 8 slightly beefier cores instead? Let's say that (hypothetically) without the E-cores, I perhaps could have had 8 cores at 5.3 GHz instead of 4.9 GHz. I can live with 2 small cores for background processing I guess.
Yeah, what about a 50-50 silicon space split - 4 p-cores and 24 e-cores, this could be an interesting option as well given how much performance to power they offer...
I am so grateful for the fact that you decided to make an english channel
Yeah, I love his input on topics like this.
That Lasso app, may be a good work around for older games with DRM that flakes out when bounced between P-cores and E-cores, just a thought. Great vid der8auer and crew. B)
WHAT'S UP ZDAWG
Would love to see someone test this
Yea was thinking the same 👍
Unfortunately, that will almost certainly not work because DRM is typically at the level of kernel modules (otherwise you could just run some user level software that fakes the runtime environment to fool it)
only for user mode aplications, but anti-cheats based around kernel mode stuff, won't care about process lasso.
You should try disabling the P cores and seeing how far you can overclock the E cores by themselves. How close can an overclocked E core get to the performance of a P core when it has the thermal and voltage headroom?
Afaik you need at least one p core for cpu to function
Intel says 300mgz on e-core
Some instruction set extension are missing on the E cores so it won't be a perfect comparison, but it would be interesting.
You get like an i7-9700K inside a 12900K, or a 9900K without hyperthreading.
@@saricubra2867 more like an 8-core atom cpu that's nowhere near i7 ipc performance but still decently punching above its weight.
I do use PL to set all background tasks to the lowest cores in my 5900X leaving the upper faster cores for gaming and productivity apps. You don't have to go options > add to list.. Just right click the selected apps and choose an "always" affinity. At the end you can backup all your settings.
Problem with CB is that it adjusts the affinity by itself if I recall correctly.
did you measure any difference by isolating background tasks to a single core?
@@ThunderingRoar I haven't done that since one core is too little for some of them such as iCUE
I leave all my light background (non time critical) tasks such as TaskManager, HWInfo, RainMeter, Walpaper Engine (I only play a video with it), Oculus tray tool, Skype, Discord etc... to the last 6 logical cores. (18-23)
Things like Chrome go on 8-17
More time critical gaming related background tasks such as Virtual Desktop / Oculus wireless streamer services / apps go to 6-23 so they can use the first CCD while gaming (lower inter CCD latency)
Games and heavy productivity apps go to 0-23 so they can use all cores and most importantly, run the heaviest threads on the first cores.
I have tested with and without PL. While the Windows scheduler in Win10 does a great job you can eliminate any microstutter from games and chrome by assigning chunks of your CPU cores so they don't conflict.
Most of the windows tasks and services are set to use all cores. Some of them aren't supposed to have affinity.
I can say the Task Manager, Print Spooler, Gaming services can have a custom affinity without issues. Not that I encounter any when I tried to set affinity to a lot of windows tasks anyway. Things like DWM need to be snappy so I think is better to leave alone. Anything system critical should be left alone IMHO.
Also "ProBalance" off or anything related to Power managment. Those things are for really slow machines. You wanna use PL exclusively to manage affinities.
Thanks!
@@SaccoBelmonte I'd love to see some channels do some extensive testing on win10/11 scheduler performance differences vs manual affinities regarding latency, stutter etc in all kinds of scenarios-- so far I'm having a hard time finding any in win11 on my 5900x build
@@interlace84 There are none. The scheduler is a rather obscure process running behind scene and is hard to know (if even possible) what's going on.
I think the current CPPC mask (and CPPC preferred cores) in Zen3 is incorrect in Win11 because it exposes a strange behavior in two of the lowest cores on my 5900X. I have them both disabled for now and rely on PL to set affinities around.
PL's latest update added CPU Sets....I need to dig a bit deeper into that. Sounds interesting though.
So, it is just as I imagined - disabling the P-cores basically turns this into a i7-7700k, though with a 65w TDP, essentially confirming what they are - efficient.
7700k is four cores. There are eight economy cores on this CPU. He said single economy core performance is about the same as one core of 7700k. ... Surely that makes the eight E-cores like two 7700k stuck together.
that's actually quite impressive
@@threecats8219 More like two 7600Ks stuck together since E-cores don’t have hyper-threading
@@yancgc5098 Not really as it took 8 e cores to even match and yet still be behind a 7700K and that is double core count. And no hyper threading is not anywhere near the same as actual real cores. Hyper threading smoothes out multi tasking, it does not seem to give more raw performance rather more efficient utilization of processes that do not fully utilize a CPU core and in theory if 2 threads can use 50% or less of one CPU core, it can divide the threads for better use., But that is no where near the performance of an actual core so no a good comparison.
Would be interesting to see how the performance changes when using: ALL Cores VS P-Cores ONLY VS E-Cores ONLY
Anandtech has that for some benchmarks IIRC.
Interesting observation! So if Intel would make e-core only CPU it would have about much higher multicore performance.
Or maybe 4 p-cores + 24 e-cores :)
@@dmitrylozenko7123 I’ve been wondering if they’ll do this for mobile class CPUs. Maybe 2 or 4 P-Cores but stack up and have 8 e-cores. Have more efficient cores to run most tasks but having powerful cores available for heavier tasks. I wonder if this could be a leap forward in battery life.
I think PC World touched on that about a week ago. Gordon is wearing a Detroit Police t-shirt in the vid.
2P +8E for ultra book seem like it will be good.
maybe for 7-15watt chips it's good, but 2 powerful cores and 8 slow cores are not really beneficial especially since modern ultrabooks can actually game
@@stalememe6407 I assume Jesse is a bot??
@Jesse Schultz dont compare amd 5000 series to intel 9 gen
@Jesse Schultz AMD doesnt come close to dominate, what are you smoking? in terms of best cpu for that (and which "that"? different segments of mobile exist)? They certainly are not the sales leader. YOu have to do more than beat intel for on year or so - you have to start being truly competitive in ALL ways for many years.
Being 80% the performance while only using 1/3 the watts is crazy. Why dont they just use like ALL E cores on a cpu? Like 24 E cores haha.
Would love to see a single core benchmark 4GHz p-core vs 4GHz e-core. Another awesome video that doesnt mirror the same garbage other tech youtubers keep showing, THANK YOU so much for good content.
This is something I'd like to see too. What is the actual IPC difference at the same clocks.
CORE WARS
P cores would be 40% faster, the e-cores are basically Skylake.
@@saricubra2867 you mean lowest-tier 6000-series without hyperthreading? and only in the 12900k? Get a 12700k and prepare to be very disappointed. They won't clock much further than base either..
@@interlace84 I watched that the real multithread perfomance of the e cores on the 12900K are comparable to a Ryzen 7 1700.
I always maintained that the Gracemont cores are a brilliant piece of engineering. All 8 of them only consumes around 50W under full utilization and are as fast as a first generation 8 Core Ryzen.
Faster actually.
And because they lack hyperthreading, each efficiency core only has one thread to work with. Which will not hurt performance as much.
For me it's kind of disappointing since the IPC is lower than of skylake and just marginally faster than haswell. that's 2013-2014 territory there. if you compare a stock 12900k and compare the stock E and p cores, the P cores performance per core is twice as fast but the P cores dont even consume twice the energy, so in that regard the p cores are more efficient than the e cores
@@SweatyFeetGirl Complete lie you just told. Stock intel setting is PL2 241 watts. That is stock.
@@Nemesis1ism a processor is first of all not only the cores, there are a lot more components than cores, but 1 e core consumes 25watts of energy on its own already.
I agree. The whole point of the E-cores was to give smaller programs that won't really see a benefit to the bigger E-cores a place to run, like Discord, browsers, mail clients and office applications, MP3 players, programs for LEDs and other ancillary hardware (like iCue from Corsair), etc. None of those uses put any demanding strain on the CPU's. I could see an argument for more P-cores for use cases like CAD/CAM, media editing, modeling software (things like fluid dynamics, crash tests, strain loading, etc.), etc. But even then, just to chuck the processes for the various Windows systems onto the E-cores would be a godsend. Plus, it's a way to add a lot more cores without really increasing the price. (Looking at you, Threadripper.)
A random thought I just had was that it would also decrease the likelyhood of a random process freezing one of your main cores and leading to unresponsive OS or IO
@@Kyrator88 We have multicore processors now, and that logic would apply to those as well. I don't think that's something an e-core would necessarily prevent.
But it's a good though.
These E-Cores perform like my i7-6700k, but they are much smaller, and much more efficient
yeah that's why it's unbelievable that Intel chose to not have any e-cores on the low-end die of alder lake S.
Yes. That is very dissapointing
You mean those 8 12900k E-cores perform just like your old quadcore.. people on the 12700k side of things won't be so lucky.
@@interlace84 That's still 8 threads in total. Although the point is not to compare with existing skus but seeing the opportunity on use case scenarios where efficiency is a major part of the constraints.
Intel could totally make a chip only populated with ecores that can crash a M1pro and even beefier iterations in workstation/server applications and still keeping a small form factor or a passive cooling design.
that is how technological advancement should work. We all have AMD to thank for that.
Nice video. The concept of having smaller cores with "low" IPC is a very old one - GPUs are a prime example, packing 1000s of cores on a small die.
Damn, that really puts into perspective just how old and slow my CPU is! I thought my 4790K was still fast, but it only gets about 410 Single Core and 2000 Multi Core in Cinebench R20! Just the E Cores in this chip are faster than my 4790k!
Still, in most cases it doesn't really make a big difference, I'm still on a 3930k and 7970 and my only bottleneck is the 3GB vram
@@satibel Your use case isn't a good example of it not mattering though, unless you assume everyone uses GPUs from 10 years ago as you do
@@kendokaaa if I can play 1080p 60 with my setup, any upgrade will do better, so it's imo still a good example.
Of course if you want 170hz+ hyper mega ultra settings full rtx that's not relevant, and you probably will want the latest CPU.
Of course an old cpu like that will be a bottleneck for a 3080, but for a 1650 it probably won't.
@@kendokaaa I just replaced my oc'd 3770k with a 5900k-- both systems had equal 4k performance with my rtx2070 at 4k maxed out in several titles.. both playable so the bottlenecks aren't always CPUs. Lower resolutions got major increases though.
@Daniel You're comparing your quadcore vs 8 E-cores that are only in the 12900k-- the 12700k ones are instantly halved already. Don't sweat over it.
Very good video. People forget that E-core is comparable to Skylake or Zen, Zen+ . It might be hard to believe - but it is true. Calling it Atom was a bit of mistake IMO, but otherwise ingenious move.
And all 8 of them combined uses less than 50 watts under full AVX2 load. Kudos to the Gracemont team!
I want a 16 core version with only e cores
the e core is rather comparable to haswell/zen than to skylake
@@SweatyFeetGirl Depends in what task and what benchmark.
Having a little buddy by your side is great.
I love that you always manage to have interesting and unique content like this! It always keeps me coming back. Plus you do almost everything in English and German.
Interesting piece Roman. Why would people think the e-cores are so shite, that's what I don't get. As you said, makes perfect sense to have CPUs like this to assign the appropriate type of core to the appropriate type of process/workload.
Sounds like unverified rumors... Just people randomly spouting things and repeating it without testing it first.. I figure the names "performance" and "efficiency" probably add a bias without knowing how they work.
I saw another reviewer who did similar tests and his conclusion was very different. Maybe it's the software he used to separate the cores but he has much less then half the fps in games and multi threading capability was awful. This review gives me renewed hope in those little e cores.
E cores only make sense in a device that runs on a battery/small device. In a desktop that already sucks up hundreds of watts it doesnt make sense. a 10core 20 thread p core only chip would be even plenty faster than a 16 core 24thread 8p+8e core config.... It's intel just not being able to get a better efficiency chip that's why they had to introduce this technology to somehow keep their chips from melting themselves.
@@SweatyFeetGirl before I watched this video I would've agreed with you, but after looking at his data I'd prefer the 8+8 combo over a 10 core in basically any situation.
people just think its bad because of a handful of cases where the thread director decides to put the load on e cores instead of an unused pcore optimally, leading to a potential drop of performance. It's the same idea and situation as when hyperthreading was introduced, and games back then during sandy bridge which used 4 threads or less ideally loaded it onto the physical core rather than the hyper threads, and would see a performance game by doing so/turning hyperthreading off. Of course though, over time, that sentiment ages badly
I wasn't sure about Process Lasso previously, it's really nice to see a channel I watch use it
I wonder if project lasso could be used to fix alder lake compatibility issues with certain apps on Windows 10. If those issues were caused by jumps to the E-cores, disabling E-cores for that task could solve the issue.
I haven't used mine much but I took a win 10 installation that has seen like 10 chip sets, and it used the cores on my 12600k very well a few performance cores were running during gaming. And all other cores at 0% useage.
And it used much less power than my 5900x about 50-70% less,(70-100w for the 5900x and about +/-35W for the 12600k.
And despite much worse cooling that the 5900x has the 12600k wouldn't go over 70c gaming and 30c idle while the 5900x idles at 60-70 and goes up to 90 while doing light gaming.
It was also getting more fps in nfs heat with a worse ram(not much worse both 3600mhz but ryzen had dual rank ram with better timings).
So all things considered very happy with Intel's new CPU.
E-cores performance is not a joke indeed considering how efficient they are
Process lasso seems like a very useful tool. Thanks for sharing
Process Lasso is so insane stable and it works like you want it to work.
Top 10 must have software to make Windows run more balanced.
This video of yours is pure gold! For that Lasso Pro tool alone already. I used a BIOS feature to park all E-Cores, on a key-press, so my x265 processes could run full speed. Now, with Lasso, I can just permanently set affinity to just all P-Cores, and still have my E-Cores active for everything else. :)
Comparing just ST to ST, E-cores have about 60% of the performance of a P-core. Especially given the substantial clockspeed difference, that's pretty nuts! Very cool video.
Let's not forget the difference in thermals and power consumption.
@@gurjindersingh3843 And 2 threads on a P-core, which, in the end, makes it ~2,2x productive.
A purrffect demonstration. :)
That's a very shiek comment.
A way I explain hybrid cores... it's like optane. You have low cost (in silicon) e-cores to do less interesting tasks. P cores = cache for high volume tasks. If you want 100% p-cores, you get a lot less compute/storage for your dollar/silicon. 100% e-cores will suffer in certain task models.
I mostly am concerned about scheduling. It would be great to see a video on proper use and how to take advantage of the E-Cores.
Props for mentioning Process Lasso Pro.. it's not mentioned enough by tech channels. I'm surprised that so many act like it doesn't exist.
My concern has never been with the processing capability of the E cores. I don't have much confidence in schedulers doing the "right thing" with processes. We've been living in an SMP world for so long.
Yes, ARM asymmetric SoCs have been around for a long time, but that's in mobile workloads, not desktop.
We shall see, though.
Sure ARM improved that tech during a decade but you seem to neglect that's not just mobile workloads but embedded workloads as well with much higher time constraints and those worloads run on more slower processors and memory/cache than laptop/desktop so they need all the resources available as well as the optimizations.
The reason ARM heterogeneous arch works well without any special scheduler is because they mainly use the Linux kernel which was long time ago tailored for NUMA multi CPU servers and doesn't have the tendency to randomly switch process execution from a thread to another.
The scheduler just need to be as simple as the Pstates frequency governor : prioritizing performance or efficiency by launching a new processus on one type of core according to the processus priority then moving it from the other type depending on its resource usage. And It just works.
I totally agree. Just not a good design decision. We have been in an SMP world at the low level Windows NT flavors and even Linux and the like schedulers for so long in the X86 ecosystem that Intel's approach to this kind of change just does not make sense. And thus I have no confidence in schedulers doing right things with processes designed for same type of CPU core count and type as well as a massive library of software in the X86 ecosystem. This is a much bigger and more radical change that can cause much more issues than any of the other changes we have seen the last 20 years or maybe more. I mean chipset changes, CPU arch changes like pipelines and IPC and core counts, just new drivers and such at low level. And scheduler changes in Windows not an issue as SMP has been around since like 1995 with Pentium Pro. Switch to dual core CPUs. Not at all an issue as a dual core CPU was no different to Windows NT and Linux scheduler than a dual socket CPU system as all dual core was was 2 CPUs in one chip. Likewise Quad cores 4 CPUs in one chip as opposed to quad socket single core CPU mobos. Not an issue for SMP world. But a hybrid arch like this and right away not good. Nothing wrong with trying something different, but Intel making a mistake locking SKUs to 8 P cores only and not offering separate ones with no e cores that have 8 or more. Cause it is going to be a long long long time if ever hybrid arch catches on or is good for the desktop.
Well if you are gonna make this argument then you should try running the game on just 2 P-cores and see if the result is better than running on all the E-cores.
Great video and great thing that you showed Process Lasso! I am a proud pro user since 2014, it gave a new dimension to process management on Windows, and helped me plenty of times with pro audio related stuff and games. Great little piece of software.
@Roman, a reason for P cores instead of 80 e-cores would be that at some point, the interconnects either take too much space, or become too slow. It's a balancing game.
Actually you know the reverse would be cool too, all E-cores for the HEDT platform, that'd be really interesting
That's what they woud've done if they didn't need to ship 'em with high-end AIOs to handle the insane power draw...
Interesting suggesting 6P + 16E. I came to the same idea previously. Also, process lasso could be used to fix compatibility issues with programs that don't work on hybrid architecture without disabling E cores in bios, or waiting for scroll lock hack which likely does same thing on the fly.
Good thing the new `13th gen will have 8p cores 16e cores for 24 cores 32 threads. GONNA BE SO SICK
How about using P core for gaming while you use E cores for streaming? I think would be interesting to see and experiment.
Re: Process Lasso. I've got a 5800X, I use it for gaming, so I've got two sort of "extra" cores in there (Not all games can take advantage of eight cores and sixteen threads), so I used Process Lasso to put all my background software on my two slowest cores. Backup software, iCUE, KeePass, Razer Synapse, all that extra -stuff- hangs out on cores five and seven, which leaves the rest of the cores wide open for whatever. I don't know if it provides any real benefit, but it's fun to play with!
:) It is fun to play with :)
At the very least it could help marginally. It doesn't really hurt anything anyway.
@@MLWJ1993 Yeah, it helps avoiding microstutter.
It might help or it might hinder, I'd benchmark it. I tried this both on my 2700X and my 3900X and performance usually ended up worse by a small amount
I know we are getting 2P + 8E Alder Lake for laptops. Pretty excited to see how that'll perform!
Thank you for the process lasso demonstration, you are a genius my sir. Can't wait to put my 12th gen system together and try this out.
The rectangular shape and the gaps between the P-Cores makes it resemble the T-800 chip from the Terminator.
Man. A die full of E cores would make a heckuva virtualization lab, especially without hyperthreading. Here's hoping we do see the 6x16 arrangement you mentioned. 👍️
Luckily raptor lake (13th gen Intel ) is adding another 8 e cores to the I9. So 8p 8e now, to 8p 16e cores next gen.
Nice video. Because 'size' is a bit ambiguous, I recommend specifying 'area' or 'length' (node size for example) so there's less confusion.
This is great, I've been waiting for someone to try this.
This is very interesting. I was actually really curious about how much performance the e-cores offer and that's quite a bit more than what I imagined.
Virgin Gamers: I NEED 10 P CORES NOW.
Chad Budget Server Enthusiasts: I NEED 40 E CORES NOW.
just buy AMD and stop complaining. No one cares what CPU you use . You want a workstation build a workstation. Stop whining.
Inb4 Intel Xeon Phi is back
@@der8auer-en An now 2 years later we have 144 of them in one socket, with 288 on the way! Phi rises!
Wow amazing discovery. Thanks for it and good job Intel!
Very good video Roman! Totally agree with you..
I'd love to see this pushed to the absolute max, a CPU with just one or two P-core of the fastest possible design for running single-threaded tasks as quickly as possible, and the rest as E-cores.
Yeah, I agree. This seems like the way to go. People run highly parallelizable codes on GPU's but some codes are really difficult to implement to GPU. This would be the perfect solution. Although I think I'd prefer something like 4 P cores and the rest E cores just because I have some codes that I want to run with 4 cores super fast but most of my codes are as many cores as possible.
9:19 THIS IS THE CONTENT I SUBSCRIBE FOR
Can you push the e-cores to the max just to see what the upper bound is, without the p-cores working of course?
1 P-core has to be active at all times sadly
Thanks man. So many people underestimate the E-cores.
This is super cool, I can definitely see some situations like power saving on a laptop device while gaming which would preserve battery life but still offer very decent performance.
9:26 Atom cores from 2016 were good, you just didnt get many of them and they were usually clocked REALLY low, and that's why almost all atom processors were terrible. Like, if you had gotten 8+ Atom cores clocked to 3.2Ghz instead of 2-4 cores at 1.8Ghz, they'd still be pretty efficient but way better performance. because below 3.2Ghz there is a diminishing return on power savings compared to performance loss on most modern silicon. This holds true on both my i7-8750H, my 3950x, and my 4650G. Oh, and thats probably why the maximum clocks for the M1 Mac Mini is 3.2Ghz
it's generally a problem with nodes. as a reference, i can run my 3700X at 1.175 Volts at 4.1ghz but for 4.2ghz i need 1.25Volts, that's 6% more voltage for 2% more speed - it might seem low, but the temperature spike and wattage increases a lot for diminishing returns.
Being able to hit high clock speeds is part of a processor architecture's design. Atom cores could not and it's absolutely valid to criticize them for that.
Lol no, intel atom is really bad even the clock is higher, intel atom Denverton clocked at 2,2 ghz, and geekbench score is only 300-400, intel xeon from sandy bridge with same clock score around 500-600, and skylake still score around 700 with same clock speed, intel inside skylake mobile processor is not atom processor
@@adiirfan01 Thats sort my point though, that Denverton Atom C3858 2Ghz 12 core at 25w used closed to 1/3rd the power of an E3-1245 V6 75w(kabylake both 14nm so neither had a node advantage) but the atom was closer to half the performance instead of 1/3rd. I dont remember which model, but there was an atom that had 2 or 4 more cores, and ran at... 2.5Ghz and only went up to 30w. This is what i was talking about there being a diminishing return of power savings They added 2-4 more of these same atom cores(15-30% more), bumped up the clocks by 25%, but it only used 20% more power? It should have used 45% more power if this stuff scaled linearly. There seems to be a bell curve with power per watt peaking at a certain point in the silicon's range. With Apple M1, and AMD Zen2 it seems to be 3-3.3Ghz(zen2 seems to like around 3, where apple might peak around 2.8Ghz. If i had to guess, Denverton's curve peaked at 2.8Ghz as well, if it could have reached 2.8, but who knows, In tel never offered higher clocked atoms, so 2.5 might have been the peak of the efficiency curve for that arc.
The atoms main problem from my memory was the small cache, yes adding cache would have drove up power draw, but it should have increased performance by more. IIRC Denverton had a total of 1MB of cache per core, where KabyLake had 2MB of just L3 cache per core. Also Intel just needs more cache for everything.
Ever wonder why the i7-5575C would sometimes beat everything but Alderlake and non APU Zen3 even at its tiny ~60W power draw? It was because of that massive L4 cache(128MB of eDRAM), allowing that 3.7Ghz processor to beat 5.3Ghz rocketlake in several tasks at sometimes 1/4 the power(definately not most tasks, but enough to ask why didnt we continue using this and improved the eDRAM for lower latency and higher bandwidth/size).
If intel had added one or more of those eDRAM chiplets to all of their top end processors, with half the latency of DDR3, and 1/3 the latency of DDR4 AMD have had a much harder time climbing to the top.
I'm really interested to know how something like KAbyLake G would have performed if the CPU were allowed to use that 204GB/s HBM with slightly higher latency ~100ns, instead of the 30-40GB DDR4 with ~80ns of latency. Is bandwidth king for efficiency, or latency, or both
@@denvera1g1 Wow I kinda hate you now, you brought up a lot of things that I now want answers to but know I never will lol
I clicked on this after spotting those lit hardware wall frames. Cool! I want .. but where to find?
I must say I was definitely sceptical of e cores. Can easily see the value in having them on a laptop or such like with a battery but figured they'd be a waste on desktop. But its nice to see they're performing better than I expected. And its really good to see this kind of demonstration 🙂
Waste on desktop? E-cores are there for higher MT performance. You would get less performance if all cores were P-cores with same silicon space and power envelopment. Raptor lake will double number of E-cores and add zero P-cores.
@@kognak6640 I know, thats why I said that its nice to see them performing better than I expected. I meant before they launched I thought they would be a waste.
back in the core 2 quad days there was a minor penalty for running cross die. So I would always tell games that had the lowest performance to run exclusively on the 2nd die. This way that penalty would be eliminated. Back then games usually only had 1 or 2 threads. Lasso would have been nice for this since I had to do it every time.
I would appreciate large charts to summarize the results & benchmarks.
If I heard correctly, E-cores are about as fast as the original skylake generation, the 6700K and alike. And having 8 of them.. That's quite a bit of power to have. Definitely worth having!
Forgot to add - without hyperthreading. So in reality, more like a 6600K. Not a slouch at all.
About the same IPC. Lower clocks by default.
full E core die with or without 1-4 P cores would be incredibly powerful for certain workloads at amazingly better power usage as well
Excelent video! I really would like to see more E cores too. Few important tasks demands a good single thread performance.
I don't know the reason why intel didn't go for more E cores. I would prefer 4 P and 24 or even 16 E cores.
Power, thermals and socket size, considering how power hungry and hot this generation gets
Performance headlines. If they’d used more E-cores in the 12900K and less P-cores, it wouldn’t show up very well in benchmarks. Techspot did some interesting research and showed that for gaming, it would be best to only have P-cores. So Intel balanced it out to take the important ‘gaming crown’ with the P-cores yet put enough E-cores in there to also be competitive in heavily multi-threaded tasks.
Great content, as usual. I loved the real world test of the e-cores.
They were meant to be used exactply that way, that's why raptor lake will have 16 of them while keeping only 8 normal ones - to compensate MT performance and surpass Ryzen with them for dirt cheap in production.
How can intel make an 8P+16E CPU in the same die space?
thank you for this, love this. Ive been doing thread pinning for so long. glad you tried it. maybe graph out power usage comparing P to E core runs....??
when it comes to multithreading, more lower power cores is better than few high power cores, and this idea of P and E cores allows you to have as close as the best of both worlds, high power low threaded apps and the highly threaded apps. Also apps can be made to use specific cores, so it might have the most intensive part on the P and the less intensive on E cores.
Hell yeah, exactly what I thought is was. Thanks for the testing and Lasso recommendation. 6P cores can me pretty fine, I'm glad the E cores are treated like bonus performance in pricing too
but having a low-end die with only P-cores is not ideal for the business workstation and the SFF markets. 5p (10t) +4e on that die could have allowed a better versatility.
Knowledge is power, so is efficiency!
This is really neat.
Hmm....
Seems like it might be a bit of a waste if you only use the E-cores nearly exclusively, but the whole being able to bind specific applications to specific cores with Process Lasso on a permanent basis is certainly food for thought.
Thank you for this video.
Eagerly awaiting your OC guide on the MSI board, there is a lack of good info out there for Alder Lake and MSI boards.
Thanks derb8auer for this highly intriguing and informative video. It is amazing how Intel made such efficient cores for the Alder Lake architecture. I am probably wrong but, last time I checked, I think both the P cores and the E cores are on Intel's 10nm process which helps contribute to the hybrid design. It is definitely a step in the right direction since Apple also has a hybrid CPU design too.
Well now this makes more sense. I got that the e cores are skylake but more efficient cores from the get go (leaks?). The fact that they're a 4th-5th of the size of a p core yet are about 30% powerful was new and made more sense.
So useful for playing older games. Wow.
I love how so many techtubers have cats. 😊
Cats are obviously the elegant pet for the discerning nerd!
The E cores are pretty close to skylake IPC and use one quarter of the die space of the P cores as far as I'm aware.
Basically E-Cores seem like cores that miss AVX-512 instructions that can be used only for some kind of tasks that are likely optimized to run even better on GPU anyways (such as image/video processing) and they do not have hyperthreading. I saw some interesting talk about if it's better to have these features and just to have much more rather simple cores, cause in real world most applications will benefit from the second approach.
On the other hand as a software developer I have to say, that scaling to many cores is somewhat problematic: if you have task running 48ms on single core, you distribute it to 16 threads on something like Ryzen 2700x, you may expect that it will be done in 3ms. Sadly OS decides it needs one thread for itself and two for another apps - so 13 threads will complete task and remaining 3 task will complete afterwards. In the end you may see that task lasts from 5 to 8ms and CPU usage is 60%. For tasks that run longer, it's not and issue, cause they will complete roughly at the same time.
For my codes threads don't optimize performance (I use MPI) and so E-Cores seem like the way to go.
would have been nice with 40 e-core version - seems to fit right in
the e cores are almost as fast as my overclocked 9700k
i remember using process lasso when i had my 1950x
They should've just made adjustable threading and c-stepping where it decides which applications need more power and which can be divided up among cores so while gaming u could use all cores max turbo or if your running browser or streaming in background it divides that among one or two cores using threads maybe stepping down to save power
I would appreciate it if the separation of the two types of cores were better though. They still share the same L3 cache, which means that the ring ratio has to be set lower unless you disable e-cores. The ring ratio affects the frequency of all cache including L1 and L2 cache which are not shared between the two types of cores. Ideally, the L3 cache is separated with the option to have different ring ratios. Or, there's the separate option to set a different frequency for the L1 & L2 caches of each type of core and one for the shared L3 cache. Then there's the whole AVX-512 fiasco. The P-cores have the ability to execute AVX-512, which physically takes up space on the CPU doing nothing (other than maybe improving thermals, but then it is literally a glorified and suboptimal heat sink). It definitely is possible to have the P-cores support AVX-512 while having the E-cores running. Intel worked hard on its Thread Director controller and working with Microsoft on Windows Thread Scheduler, so why stop there? Why is a unified common instruction set necessary? These two problems illustrate the problems with e-cores, which is that they handicap the performance and features of the p-cores due to their lack of true separation, forcing Intel to handicap everything they share (things like the L3 cache and the common instruction set of the CPU).
I've been running Proc lasso a few years it's very useful even on a 8700k
Ian Cutress over at AnandTech said that the E-cores have up to 50% the performance of the P-cores. Also, Golden Cove has a 43% IPC uplift from Skylake. Add higher clocks and you roughly have a 50% performance boost over SkyLake. Is it a coincident, that the Gracemont cores also have 50% the performance of Golden Cove.
It makes sense because he tested a i7 6700K/7700K at 4.2GHz and found that the 8 E-cores on the i9 were basically the same performance. So, intel was right about the new Gracemont cores having the same IPC of Skylake. Using just CBR20 to get a performance percentage isn't going to show the whole picture. Check out AnandTech's review on the 12900K. He's done extensive testing on this, including integer and floating point.
I spent some time going through old submissions at HWBot to try to get a better idea on where the E cores sit simply in terms of performance and you are right about CBR20 not showing the whole picture. While it's true older Intel cpu's were not very good at Cinebench performance in general it's not like they were terrible. Since Roman used CBR20 just taking a look at those results his 8 E cores scored 3203 in multi core. Looking at the 7700k (Sure it's not a perfect comparison with 4c 8t vs 8c E ) it needs to be clocked at roughly 5925MHz to match them. Looking at the 7600k ( 4c 4t ) at it's base clock of 3800MHz you get a score of around 1600 cb so putting 2 of them together to get 8 cores makes it equal to Romans example using 8 E cores, both scoring 3200 cb. Pretty wild huh
I think the reason for low E core CPUs is because its new and software is not use to E cores yet and having the higher P core, the CPU will work mostly on their own. As software and such gets use to the E cores then they will introduce more of them. Need to find out the full spectrum difference between both types, not just performance in games but other aspects. Finding the right balance between them might be more dependent on what you are doing than just more of them. We are starting to get to that wall of cooling these down vs the power and performance and finding the balance for a workload will be more ideal than trying to cool off more of P and E cores in bigger CPUs
The E-cores take up a small part of the die compared to the P-cores.
You wouldn't get one P-core per E-core if they did swap them out.
4 e cores are the size of 1 pcore, so it would be better to have a 10core 20 thread processor with only p cores instead of 8 p cores and 8 e cores and 24 threads
@@jayden974 not in gaming though
@@jayden974 e cores are just a waste. Ipc of ivy bridge, nuff said. They're not useful for anything. Intel should just focus on more efficient designs with high performance cores, just like AMD does. They can have 5 million e cores, it still won't matter as they're useless for gaming. And alder lake is advertised as gaming processor
@@jayden974 performance even regresses with e cores enabled, just for your information
so the E-cores have Zen3 energy efficient and Zen+ processing power?
Interesting statement. Makes the e cores sound kinda weak.
Still need to consider die space.
nice! i was wondering if anyone would make a video on this. Cheers!
I can understand why some people would have the preconceived notion that the e-cores have poor performance; as someone who has used netbooks and NUCs equipped with atom cores its encouraging to see that its not true, and that they are perfectly usable.
I myself hope that hyperthreading can be added to the e-cores in the future as from what I understand its then able to use parts of the core for other tasks at a given moment that would otherwise sit dormant during a processor cycle.
Thank you for covering this aspect of Alder Lake.
Adding HT would add performance, but also power and size, partially negating their benefits. To add overall performance, I like the 6P + 16E core idea better. A side note, rumor is that Zen5 will use Zen4 Dense cores as it's "little" cores. I forget if they will have SMT, but I would guess they would. Bit of a different comparison though, Zen 3 cores are about half the size of a P core, though not that much less IPC if you normalize for clockspeed.
@@xeridea
An old Intel white paper claimed that adding SMT only increased die space by 5%. This was back in 2002, and the number might have changed substantially in the two decades since. But with the 2023 Meteor Lake projected to be Intel’s second consumer MCM in recent years (after Lakefield), and the adoption of a substantially denser node, Intel should have a lot more flexibility when it comes to core complexity.
I don’t think Intel will back down from 8P any time soon. Current gen consoles have 8 decently performant cores, so high end desktop CPUs should have at least the same amount of full performance cores.
I don’t think there’s official confirmation on SMT support for Zen4c. But if it is based on the lean cache APU version then it probably does.
If measured by core area alone (sans MLC, LLC, interconnect agent), Golden Cove (5.65sqmm) is almost exactly twice the size of Zen2 (2.83sqmm), or around 1.74x the size of a Zen3 core. I don’t have any numbers for Zen cores with cache (cache takes up a substantial amount of space on the CCD), but at the same frequency there is a ~5% difference between the ST performance of APU version and the MCM version (5600X ST is around 5% better than the 5900HX, both have a specified boost frequency of 4.6GHz).
Actually, adjusted for frequency, a Golden Cove core does still have a substantial performance lead over a MCM Zen3 core. Both the 12600K and 5950X have a specified boost frequency of 4.9GHz, the 12600K has a 18% lead in Cinebench R23 ST; 17% lead in Cinebench R20 ST.
@@DGao-zz5vq Yeah, I know HT doesn't take up a lot of space, but it is some, and may adversely affect the chip layout. Future nodes would alleviate this.
There is another drawback, in that it hurts ST performance. This may be less of an issue given the efficiency nature, and background tasks, but it could abnormally slowdown some tasks if not properly written for many different speeds of cores. For instance, often MT is done in code by simply splitting up the screen into tiles. If you just do one large tile per core, slow cores, especially if HT/SMT would be a big hinderance. Using more tiles mitigates this, but for some quick tasks, the overhead of splitting up further may not always make sense. This may affect only a small number of apps though, I have no idea, and likely it would be improved in time.
I was thinking about how historically, SMT has been a way to get more performance out of big cores, while still allowing to oversize certain parts of core to maximize ST. Just speculation though.
My reference to Zen 5 is from Moore's Law is Dead video. IIRC he was referencing Zen 4D, not 4C, apparently they are different? Still speculation, 1-2 years out.
I was thinking around 10-15%, assuming a minor clock difference, I thought 5950X boosted around 4.6, since it has 16 cores and lower power budget. If it boosts 4.9, different story, I didn't look up any numbers, was from memory. I was referring in general to diminishing returns on IPC for increasing core size, which is why Intel has made the little cores. At some point, which I would say we have reached, ST is fast enough for most apps, and apps are increasingly multithreaded. I think 6P + 16E would be a great production focused CPU.
Now I am excited to see how they work on Laptops.
Thank you for this.
Most of the "space" in a core is used by AVX instructions AVX256 and AVX512 takes up soooo much space and use soo much power.
The e-core is just a P-core without these heavy instructions, just basic x86 instructions with some of the older extensions which arent heavy.
Manually linking type of cores with specific app - i knew it will end like this.
so I can have AMD 16 cores 32 threads, or Intel 16 cores 24 threads with some slower cores if i want to compile something... and it makes perfect sense for any program to run on any core so we dont have to worry about choosing what belongs where or what might suddenly need more powerful computing. the compamies are cheaping out by not giving us the power to scale freely.
The "leaks"/etc. for 13 series (from "i5 13600K -> i9 13900L) have [6/12 + 8], [8/16 + 8] and [8/16 + 16] core setups where [P/Thread + E].
I think the i7-13700K will have a nice sweet spot for P cores, P threads and E cores.
The e cores are like a 9700K running at 4.1GHz... that's crazy. Basically identical IPC comparing to skylake but at much lower power consumption.
It's finally a new process node so I hope it's more efficient
That AIO is huge, also is it supposed to be some certain design on the display?
The performance of the E-cores lines up exactly with prelaunch leak expectations. Basically staple a 7700k to higher clocked tigerlake.
11900KB is 65 watts for PL1 and performs like a desktop i9-11900K.
So it's less about efficiency (on PC), more about the core count, since the system with both types of cores can rely on P-cores to maximize single-threaded performance and E-cores to add more workers in multi-threaded tasks.
Since power efficiency on PC makes no sense, but versatility does.
Perfomance per watt and area on the die.
The reason I'm not sold on the E-core concept is that most games I play just can't use that many cores. Let's say a certain game only supports 8 cores (afaik, current gen consoles only got 8 cores, so to my intuition it makes sense a lot of games will be coded accordingly); Why would I want 8 strong cores with some weaker cores on the same die when I could have had 8 slightly beefier cores instead? Let's say that (hypothetically) without the E-cores, I perhaps could have had 8 cores at 5.3 GHz instead of 4.9 GHz. I can live with 2 small cores for background processing I guess.
Top reasons to watch Roman:
1. He keeps hair product companies in business
2. His cat
🤣🤣
Very informative video!
That's exactly why I went with a 12900k. I want to split tasks between the P cores + rtx3080 and E cores + igpu.
Yeah, what about a 50-50 silicon space split - 4 p-cores and 24 e-cores, this could be an interesting option as well given how much performance to power they offer...