"These arrive just after my son arrived, which is 2 weeks, so..." You look EXTREMELY well rested for someone who has a 2 week old baby. :) Congratulations!!!!
@@ServeTheHomeVideoeHomeVideo So literally the super computer that just went on sale can be replaced CPU for for CPU core with 500 dualsocket blade of this thing. Assuming (generously) that each dual socket take up 4U that will be like 2000U or 50 racks... If in 2U blades they will be 25 racks, or barely even 1 ailes. It is insane.
These compare to xeon scalable 2 if I'm remembering right, so anyone who hasn't upgraded inn the last 5 years or so can probably do massive consolidation with these.
@huy1k995 Moore’s law describes a doubling of transistor density every 2 years. The consolidation he’s describing replaces silicon that is almost 10 years old. How is that an example of Moore’s Law being alive and well?
Honestly after watching that video I once again think that some leakers are only good at leaking sometimes the right stuff but lack the ability to really understand technical stuff and market positioning. Keep up that great content!
Congrats again on the baby. Hope mom and him are both doing well! We just had our second 3 months ago and I was a bit stressed about them until it was all over and we knew all was good.
@@ServeTheHomeVideo I definitely feel that one! I try to remind myself I’ll long for him to be little like this before I know it and cherish the time, even at 3am but good lord am I tired.
I'll wait for independent benchmarks from Anandtech or something. Right now, we have to take this video content with a grain of salt, considering it is an Intel-sponsored video. Patrick ain't gonna bash Intel Xeon too much, considering Intel's paying his lunch :)
Is there anywhere an average person could even purchase a CXL memory module at the moment? Just about every vendor in the space seems to be "contact us" which typically means "$10,000+" in my mind, but it would be really nice to be able to get a hold of modules of that type in a not infeasible package. Even $1,000-5,000 isn't impossible, if painful, but it seems like CXL is basically limited to enterprise large volume customers at the moment.
There is a slide in this presentation while I am talking about it. You are right. No AVX-512 and AMX on E cores, but there is an AVX2 with things like VNNI.
the ultimate test will be to run a benchmark test on a only one core (rest of cpu's in idle state) that has to traverse its data across cache and PCI to off chip memory .......get the benchmark results and then duplicate this whole process in parallel across all cores using all cashes and PCI lanes and see how much slower the original benchmark is affected by IO bottlenecks. MY GUESS IS THAT IT WILL BE BETWEEN 30-50 percent. In the end buss access is still all about the resources divided by the number of CPUs
I just hope they didn't compromise on security too much for the sake of performance, only time will tell. The efficiency tho is mighty impressive, they finally have something for the competition and not just brand bs marketing.
What sockets do the chips use? You have not actually stated THAT, in the video or ON the web site. On Page 2 of the article linked in the reply. Thanks.
@@ServeTheHomeVideo Page 2 Paragraph 10? I guess you can call that a paragraph. I'll edit my other post. Thank you. Still would have liked to see this mentioned in the video.
We talk about it in the video and have a section of the main site article. This is much faster and uses DDR5/ PCIe gen 5, has accelerators, and 250W/ 144C. Of course Ampere can be cheaper and if you want to develop on Arm without going NVIDIA then it is still the best option out there. Perf/W this is a big upgrade
I love geeking out over the latest and greatest but I'm always curious to see what trickles down for the single bare metal server or up to 3 VMs for the SMB
@@ServeTheHomeVideotoday Intel announced that Skymont E-Cores are 1.7x faster at the same power draw than Crestmont E-Cores. This CPU uses Crestmont E-Cores of course 😂😂😂 Calling it a beefed up n100 is harsh but fair.
As someone who spent the last two years writing firmware for the Birch Stream platform that these CPUs run on, this couldn't be farther from the truth. Aside from the microarchitecture and process node being completely different, the platform itself is completely different. You do not get full RAS capabilities and self-boot on the CPU with an N100. Also, the floorplan of the silicon is completely different. The N100 is monolithic while these CPUs have the compute tile in the middle with IO tiles flanking it on either side. By your logic, an AMD Epyc Genoa is just a beefed up AMD Ryzen 7000 (which is ironically closer to the truth than your statement).
@@AlexSchendel So it's a beefed up n100 with some server features that are present in all silicon but not fused off in the server part. Thanks for confirming this.
More than these chips, I am excited about the ripple effect this release will cause in the second hand market when Cascade lake Xeons start flooding eBay at a much lower price than they currently are.
Spectre and meltdown exploited inherent properties of speculative execution and out-of-order execution. Certain variants might be eliminated by having SMT-free cores, but I don't think it eliminates all (or even most) of the surface area for those vulnerabilities.
I see what you are saying now. Yeah STH is very wrong here. There's essentially no change in vulnerability to side channel attacks due to eliminating SMT.
My mind keeps coming back to Xeon Phi with those, and its atrocious 1T performance which made them annoying to work with for mixed HPC workloads. Back then, I thought they'd really do well with a single P-Core and I honestly don't see the difference with Sierra Forest. Care to explain? :)
The Atom core has evolved significantly since Airmont. The one on Sierra Forest is about as fast as a Skylake core, nothing groundbreaking but certainly fast enough for running housekeeping tasks. Note that despite their apparent similarities, Sierra Forest doesn't target the same market as Knights Landing. It's more of hosting/cloud thing. A lot of its potential customers (especially those running vSphere) will not appreciate the scheduling complexity introduced by the addition of a P-core.
Finally, an interesting Intel Xeon product! However, with Turin 5c on TSMC N3 right around the corner, I don't think Intel products will bode well. AMD is going to have the performance lead with their 192c chip and I guess the perf/watt lead as well.
@@ServeTheHomeVideo Xeon 6900 will be the same arch and process but with more core. At the moment, 64c 128t Epyc 4th gen is roughly equivalent to the Xeon 144c. Zen 5c will be on a new process, new arch and will have more core. It is clear to me that the Epyc 192c/384t will be superior, but I'm curious to see what the result will be at thread parity !
@@matthope_qc Well - TBF I think 288 cores will have more memory channels than 144 cores one. And more I/O in general probably. Yes, it's likely that Intel will still be behind AMD in performance. performance per Watt is still a bit of an open question (heard ton of good things about Intel3, some bad, some good things about N3). I certainly can't wait for benchmarks and various comparisons.
I think reading the Phoronix benchmarks against the top of the line EPYC skus the comparison (if you’re not comparing floating point & avx) is actually quite decent, so I definitely have a bit of hope that Intel and AMD will be pretty close for Turin Dense vs SF-AP.
@@ServeTheHomeVideo Considering Turin is Genoa SP5 compatible, I think AMD has a big advantage there over Intel, basically anyone on EPYC will remain on EPYC.
more cores. I dig it as long as I am not payin the power bill (yup, I know they sip power). What is the audience/workload for a machine with that many (x2?) cores? virtual host? not crypto.
I complained last time about the title and thumbnail. I must say that I love this title and thumbnail. They are informative but also raise my curiosity. And since there is a lot of information that is in the video itself, there was no reason to be stingy on the former. Thanks Patrick. And congratulations on the new member of the family! Sorry I was so harsh the previous video around the birth of your child. I did not know.
sponsored or not, this is a terrific chip. I have a two racks of 6 core Xeons, the bottom end of these chips with 2 servers would replace a bunch of servers.
Also, this is so exciting. I’m really hoping Intel can execute both on the fab side as well as design. I think Gelsinger is a good pick for the role and I’m rooting for him and the entire org. The more truly viable entrants we have in the market the better. I cheered for AMD when they started to mount a comeback and now I’m doing the same for Intel. On the consumer side, we have repeatedly seen what happens when there is no real competition. Very minor bumps gen on gen.
The NVIDIA Grace Superchip has soldered LPDDR5X memory and can hit 144 cores on the dual 72 core chip module. Those 144 cores with memory use around 500W versus Intel at 250W excluding memory but with the ability to scale to dual sockets at 288 cores. Very different CPU segments
So I think you are wrong on the 6700 series, I suspect it will be the primary sales driver. It’s a great replacement for data centers that have aging equipment that had lower power draws. I see this as the primary upgrade path for an enormous number of 5-6yr old racks/pods. As always, imho :)
I don't have any interest about server hardware at all but just like to watch due to your enthusiasm and excitement of talking about this stuff. Keep on Serving.
This is less of the AI chip. Granite Rapids (P-core) is better for that with AMX. Think of this as taking VMs off of legacy chips from many servers, and putting lots of those VMs and microservice onto a single server instead. That is why this is an exciting launch. Intel is, for the first time, really putting forward different domain specific CPU architectures.
I think the real intent (other than a move favorable comparison) to that choice was the TDP target of 250w. The Epyc 9534 default TDP is 280w and configurable down to 240W. The 128 Core ranges 320-400w for which it better be faster. On that basis the 84 core Epyc 9634 would've been the better comparison with the same TDP range as the 64 core.
Congrats on the kiddo, STH+ST_Crib, I'm still running my home nas from before we had kids and the oldest is still years away from having independent screen time, but I'm hoping to see some ideas for engaging kids with real tech stuff early on. So far 3D printing is leading the way more than compute type stuff.
@@ServeTheHomeVideo By that time, I think that was the idea, but they lacked acknowledgement or technology. On my point of view now they perfected that old idea. As many core as possible , on low consumption and a kind of weird configuration !
Nice. Congrats to Intel on delivering and beyond. I already went with AMD Siena, it has a lot of PCIe, and scales down to 8 cores, while keeping all io and memory channels. For crazy workloads, of course Bergamo. I can see Intel leaping ahead, and also having built in accelerators (compression, encryption), being a nice bonus, but I think long term these things will move to NICs (already available, just not super well integrated, and bit awkward to use it for compression for storage, on a NIC). I will stick to AMD, their more smaller tiles solution I think is more cost efficient, plus Intel screwed us many times in the past. But time will tell. In general I really like the trend, of more smaller (more efficient) cores, because a lot of workloads I do, scales very well, and ultimately I can do more in same power budget (compared to less but faster cores).
Wow this plus Lunar Lake information has me Jazzed. I'm hoping the next Gen E core server CPUs might do clusters of E cores so they can turn off unused clusters like on Lunar Lake. Would be good for cloud folks and even better for us second hand home lab buyers down the road.
This is going to redefine servers? Others have long had the same or more optimizations in this direction. Seems more like Intel is catching up at best.
@@ServeTheHomeVideo"Ahead" meaning what in terms of "redefining"? Even if the first true, if it is just about being a good product, then good for them and for us to have it as an option, but nothing that changes how servers are used. CXL has been long in the making and far from specific to this, chiplets are great, but hardly their innovation, more cores, each more efficient rather than performant, great but again just their version of what AMD and ARM have been pushing already, same goes for removing the PCH. Good to see competition doing what it is supposed to of course, but nothing that seems structurally new, other than being new for Intel. And on that note, it's easy to look good, if you compare yourself to the competitions last gen. 5th gen Epyc has up to 192 cores & 384 threads. If the efficiency cores of Intel are about on par with last gens Epyc cores in performance, win some loose some, and a bit better in efficiency, that isn't that impressive, since 5th gen Epyc, especially the -c version will be on better nodes than their previous generation as well of course, while adding some improvements beyond passing on TSMCs I'm sure. As for Neoverse N1, it's from 2019, it better be defeated easily by a 2024 product. Granted Nvidia isn't realeasing their next gen yet, unlike AMD, but Grace is not new, and also has 144 cores and they aren't even from the lots of core N line, but the V one, maybe a little physically bigger as a package based on the less chiplet approach, but since that package includes memory, it's not actually large if you include that for the others.
Yes there is. In data centres the two most important things are space (limiting the number of racks needed) and energy (which effects cooling, power costs, etc). If a single 2u server can do the same work as a 24u rack, it can be worth it because you can expand without needing more square footage.
@@SpoonHurler you always need redundancy, for failover + at least one more DC. I cannot imagine an enterprise which would require several such systems. I mean i cannot imagine this planet having so many such enteprises to justify the creation of this CPU. Only may be IaaS providers.
The enterprise market is shrinking and now represents less than 50% of worldwide server shipments and this number is falling by the year. Cloud service providers are the majority share of server volume and this targets the primary workload used by the CSP customer base. General purpose P-core designs will become increasingly niche over time and this is represented in the server CPU roadmaps.
Intel 3 is 5-7nm tsmc equivalent? Never expected to have cpus with as many cores as GPUs. Its not there yet but the gap is closing. Just crazy considering the number of cores just 10 years ago.
@@timmy85211 Intel has density.. tsmc has efficiency. Intel can pack transistors closer but requires more power than the "equivalent" tsmc node. Power usage is a large part of the benefit of the node so a huge downside to Intel. "Same node" but twice the power means... Not an equivalent node imo
Before watching the video: This is the "infamous" Sierra Forest, right? I kind of wish they did 288 core version, but they were afraid of being too late on the market. Sadly from what I've heard they are still too late (by at least a quarter) and now they don't have their "flagship" product. Not that too many companies would buy 288 core version in big quantities. Also word on the "street" is that this is 144 E-cores. I really hope those are the updated Skymonts not the older ones. Anyway, this is going to be interesting. I wonder how it will stack against Turin Dense (192 Zen5c cores). Also I do wonder how regular Turin will look. EDIT: After watching the video - I'm surprised they will have 288 core part. While I think the biggest product was above 300 cores, that was scrapped (I would have to look back on leaked road maps), but it's nice that they kept working on 288 core variant. Intel was always fantastic in innovation, just very variable in actually doing something with said innovation. And it is great to know that E-cores are on Intel 3. That should make them real efficient and it will be useful in places where power to performance is the most important (but you don't have to pay "per core" costs, so no VMWare). Sadly for Intel, this is still a bit too late (12-18 months) for them to retain large portion of their customers. Some went with Arm, others with AMD - depending on what they needed. If Emerald Rapids didn't come to the rescue in the fall of 2023, Intel would be looking at abysmal future. For now it just looks bad and they will have to leverage their Foundries heavily. And also stoked to see that P-core versions will be available this year… I think. Intel's road maps have tendency to be very fluid, and I don't even remember all their code names anymore.
@@ServeTheHomeVideo That's good. I think it is still behind original schedule, but at least it is before AMD releases their new Architecture. Personally I dislike when those companies release stuff in "reactive" timing, something AMD and Intel did before - so usually waiting what other will announce and announce something better "just quarter later". But in this case, AMD is a quarter late and if we don't count really old Intel road maps, then this is more or less on time (though I think Sierra Forest was supposed to be Q1 even in last years road maps). There will e competition and the most important thing - there will be lots of SKU's that will fit most niches.
This looks very exciting, but it does have me wondering if you consolidate like 12 servers to a single machine, are you not creating a much bigger single point of failure in your infrastructure?
I’m so excited for the future of Intel … the r/d have really did alot of breakthroughs .. powervia and ribbonfet are coming and 2027 will be produced in the new high N.A. machines !!
Hard to unpack what is just "marketing speak" when it comes to E and P cores...seems like E cores are possibly just 'P' cores that can't handle higher clock speeds rather than a fundamentally different architecture.
They are fundamentally different. They are about 1/4 the size for starters, and then also the differences in instructions. They're even designed by different teams at Intel.
Turin is Q4. So the 128 core Turin in Q4 will actually see Granite Rapids-AP launch first with 128 cores. AMD will be behind next quarter unless they move Turin up. Turin Dense at 192 cores will be competing with 288 core Sierra Forest. 288 physical cores versus 192C/ 384T is not a straight win. We go into this in the video.
@@dota2tournamentss Yes, 2nd half '24, which could be as early as 1 month away, or as late as 6 months. AMD did fast-forward the Zen 5 Strix Point processors to next month, they may also have pushed forward Turin, but there's no way to know.
@@diamondlion47 It appears that Intel is dropping HT (hyper threading) in favour of using a single thread on each p-core, with the e-cores presumably taking on the role of what used to be a thread, sort of but not really, since e-cores are also not fully ISA compatible with the p-cores. AMD's "c" cores have the full ISA and threading, so there's no need for a much more complex and messy scheduling system to make their solution work.
All E cores would be super if could combine with an all P core Xeon. That would be a step forward. Not 2 new sockets in the same generation, and you get stuck on the same E or P core CPU
I think after intel showing how well skymont performs compared to cresmont-lp, I wouldn't be too keen on these chips and would rather wait a bit for skymont.
Wow. That is a wild statement. This is an absolutely crazy change for server CPUs. Easily the biggest for Intel in two decades or more. Sorry. After using these (and you can see our Genoa/ Genoa-X/ Bergamo videos as well) I have no way to reconcile that comment with the servers we are showing.
@@ServeTheHomeVideo Perhaps for their specific use? Though it's possible that said person didn't watch the video. Or that change isn't for them. I agree that this is a huge change, just not for everyone. And other CPU's should come later. If there are no delays. For some it will be too late.
@@jannegrey I also look at it like by making the E-core CPU, it lets Intel do more on the P-core side since it does not have to focus on the lower performance market.
It's rather simple reasoning. If more, weaker, cores is your thing, AMD has a better solution. If less, more powerful single cores is your thing, AMD has a better solution. Why would I go Intel when either way, AMD has a better solution in their newly announced product stack? That Epyc 9534 processor has been out for 2 years now. There is no compelling reason to switch back to Intel. This might be "cool" for enthusiast reasons but it's not a viable option when compared to the competition.
@@ServeTheHomeVideo It is a lot of cores and a very good power target - Intel explicitly market the E-Cores for some sub-markets like web servers and CDNs. It could well be possible that these chips will not so hot for other tasks where Intel explicitly markets the P-Cores (databases, simulations, etc.). In the end the ARM servers chips with many cores and good efficiency suffered from the same issue and only limited pickup.
All the intel hate is wild in the comments. There is a reason not every company has replaced their intel servers with AMD. Intel has a TON of integrated features that AMD just doesnt have. Intel also has way more, more task optimized CPUs
@@AlpineTheHusky You're wasting your time reasoning with some of the people man. They simplistically think "Intel quad cores for 7 years straight on desktop bad, AMD good" and apply that to the server space where It's not always all about core counts and PCIe lanes. Intel has decades of software optimizations and working with the ecosystem to make sure the latest SW runs smoothly on Xeon, plus they have their proprietary accelerators, their AMX AI engine in the cores.... that's not something the typical "moar cores moar better" people will easily pick up on or understand. They'll just see more cores or see that their 7800X3D on desktop wins gaming and then use that to say Intel is washed lol.
It's not out yet. As of this moment this is the highest core count CPU.. at least for x86 (IDK what Ampere or the custom ARM silicon that AWS, Google, etc. have as far as core count goes).
They announced a 192 core part currently slated for Q4. Intel 288 core is slated for Q1. AMD will also lose the P-core max core count in Q3 with Granite AP at 128 cores before 128c Turin is out.
Who cares 90% of clients won't touch AMD and their issues, hassles, shit drivers, updates, inferior designs so it's moot. Maybe one day clients will let me sell them cheap amd stuff but not now, IT buyers won't buy AMD, which is why Intel CPU's account for 80% of marketshare even after essentially not even trying to compete for a generation or three lol.
It is hard to say that. Think roughly 2 E-cores = 2 Emerald Rapids P-core threads. Or if you have an Intel Xeon E5-2650 V4, just assume core for core replacement. They are slower than mainstream cores, but that is also the point to hit power savings targets.
The decade old Xeon Phi with many cores on a dedicated card is a better approach... can you question Intel why do that approach instead of this weaker main cpu core approach
Not sure what you mean? We reviewed Knights Landing and such many years ago (our box is the one LTT ended up using.) This is a much better approach. Also as a "weaker main CPU" it has higher integer performance at 250W than the 64 core top-end 5th Gen Xeon at 350W. Faster and 100W less TDP is hard to call weaker.
@@ServeTheHomeVideo Your article on 500+ cores per U in blade system for Clearwater Forest makes sense then for power savings // I was thinking Home Server- Small business which this chip series is not the target market... I can dream of a PCIE 6.0 /7.0 card with 512 cores someday though- Thanks
Intel makes their own 14900k look stupid with their own cpu. Silly. 144core at 250 watts. verses over 300 watts. smh. Intel needs to slow down and do things right the first time.
Okay - "Key Lessons Learned' for this video - you can't make a full review in 2 weeks, while there is a newborn in said 2 weeks. So much work but congratulations. For now I will look at this more as a pre-review and I will check your website if there isn't anything more there (updates). Especially 3d party benchmarks. Take care! (YT deleted my comment, when I tried to put an emoji and Ctrl-Z didn't work, which is why it looks a bit weird). Plus all the chaos at home might cause you to find out that you named your son by accident to "Tiny Mini Micro" or something like that 🤣
I feel like this is the level of scaling that's difficult to visualize so I started writing things down. You're talking literally thousands of cores per rack now, right? For some reason the first emotion that comes to me is...fear?
This actually seems like a sweet homelab part for people who want to play around with virtualization but don't need tons of single core perf. Basically this lets people get the performance of buying an old craigslist dell server without heating their entire home with it. Can even probably squeeze into 1U without fans that your neighbors will file noise complaints about.
@@ServeTheHomeVideo Yeah I agree I think for homelab the lower count parts are the sweet spot. And it's also a compelling part compared to something like Ampere M128-26.
How does Patrick know anything about 'efficiency' when he drives a 6500lb+ vehicle that only gets one mile per kWh? I'm not sure we can take his reporting at face value after the Cybertruck revelations.
The low efficiency is just a downside you need to live with if you want to show your support for Elon‘s weird and fascist politics by buying a Cybertruck.
Those 144 e-core chips aren't competing with Bergamo for the reasons mentioned in the video. They (and also the 288 core parts) are also not competing with Bergamo, because Bergamo is AMD's last gen part. Their actual competitor is going be Turin-Dense.
"MOST Server cores EVER Intel Xeon 6" ServeTheHome you are technically correct, the best kind of correct. My meaning being is that AMD's 192C 384T Zen 5 chip is coming soon, but it's not out yet, so technically correct.
We covered this recently since AMD pre-announced Turin Dense VERY early. Turin is ~SC24 timeline, so about 5 months away which puts Turin Dense more in the timeframe of Sierra Forest AP with 288 cores. Here you go: www.servethehome.com/server-core-counts-going-supernova-by-q1-2025-intel-amd-arm-nvidia-ampere/
That efficiency is crazy. Doing some back of the envelope math servers with these new chips could pay for themselves in pretty short order. Intel is back!
Why sad? This is a 250W CPU with 144 cores that has faster integer performance than the top-end 64 P-core Emerald Rapids Xeon at 350W. That is a crazy win. We have some charts versus Bergamo which is probably the closest competitor, but it is a higher power platform.
You have to remember the relative performance compared to the last generation of Xeons, in which this is a 3nm process and the last gen Xeons used the 10nm process. If you compare an E core of this generation to the P core from last generation, the E core from this generation would have a slight lead from 10nm P core as the E cores are usually around 44% the performance of a P core. This means that we have more performance, higher power efficiency, and more cores which is amazing.
@@ghty-kw7hm I don't really need to remember anything. I care about what cpu has the better performance, and currently the AMD Epyc lineup is better. Nice for intel thqt they had an uplift compared to the previous generation, but in the end, how they match up to Epyc is what matters. I am also curious about the price, because performance to purchase price has been lacking too the last years...
This may have been the case over the past few generations, but with this and the new Granite Rapids-AP, that is no longer clearly the case. The Turin generation is going to be closer. It is getting a lot more nuanced starting with Sierra Forest.
Huh? We had benchmarks and power in there. Price was not released until after this but we have it on the main site and we talk about platform costs here. Very strange comment
This cpu would have been scheduled for launch in 2077 if AMD had not stepped up their game
I think it is not just AMD, but also cloud provider custom silicon.
@@ServeTheHomeVideo good point 👍
@@ServeTheHomeVideoplease do a video on amd 192 cores eypc CPU
@@ServeTheHomeVideointel is never going to best arm in efficiency thus neither in raw power
On 14+(^25) node
"These arrive just after my son arrived, which is 2 weeks, so..."
You look EXTREMELY well rested for someone who has a 2 week old baby. :)
Congratulations!!!!
In that "so tired they don't know they are tired" stage 😂
Feels like that.
Certainly tired.
Facts.
Let’s all welcome the new SFF Patrick!
This means more decommissioned servers for me to buy 😂
Data center that's still using V4 ish Xeon is literally going to be consolated to a few rack. Holy hell. Moore's Law is alive and well.
Totally. 12-16 cores was common back then with chips like the E5-2650 V4. Imagine doing 12:1 consolidation of those midrange SKUs.
That would be like consolidating my entire E5v4/1st gen SP homelab into a single 1S server.
@@ServeTheHomeVideoeHomeVideo So literally the super computer that just went on sale can be replaced CPU for for CPU core with 500 dualsocket blade of this thing. Assuming (generously) that each dual socket take up 4U that will be like 2000U or 50 racks... If in 2U blades they will be 25 racks, or barely even 1 ailes. It is insane.
These compare to xeon scalable 2 if I'm remembering right, so anyone who hasn't upgraded inn the last 5 years or so can probably do massive consolidation with these.
@huy1k995 Moore’s law describes a doubling of transistor density every 2 years. The consolidation he’s describing replaces silicon that is almost 10 years old. How is that an example of Moore’s Law being alive and well?
Huge Congratulations Patrick!! Bring on Serve the Crib
Baby STH got his own post on STH
@@ServeTheHomeVideoCongrats for your lil son, future Kennedy. ❤🥰😅
Honestly after watching that video I once again think that some leakers are only good at leaking sometimes the right stuff but lack the ability to really understand technical stuff and market positioning. Keep up that great content!
There is a lot of that out there. Even details like the PCH platform change is a huge one that we showed in 2023.
that's in large part because the market positioning can be changed literally at any moment, so its not really wise to leak those.
Let’s all welcome the new SFF - small form factor Patrick !
Congrats again on the baby. Hope mom and him are both doing well! We just had our second 3 months ago and I was a bit stressed about them until it was all over and we knew all was good.
I need more sleep!
@@ServeTheHomeVideo I definitely feel that one! I try to remind myself I’ll long for him to be little like this before I know it and cherish the time, even at 3am but good lord am I tired.
The power efficiency with that many cores and that much RAM is impressive.
Yes. It is wild.
I'll wait for independent benchmarks from Anandtech or something. Right now, we have to take this video content with a grain of salt, considering it is an Intel-sponsored video. Patrick ain't gonna bash Intel Xeon too much, considering Intel's paying his lunch :)
Can't wait till these are $100 on ebay and I can build a 288-core homelab server
Remind me in 2032!
12:45 - THAT is a lot of power savings for Intel Platforms.
Yea, the power is wild.
Is there anywhere an average person could even purchase a CXL memory module at the moment? Just about every vendor in the space seems to be "contact us" which typically means "$10,000+" in my mind, but it would be really nice to be able to get a hold of modules of that type in a not infeasible package. Even $1,000-5,000 isn't impossible, if painful, but it seems like CXL is basically limited to enterprise large volume customers at the moment.
Hmu
do you have a video that explains the difference between P and E cores? what's missing from E cores? complex instructions like AVX?
There is a slide in this presentation while I am talking about it. You are right. No AVX-512 and AMX on E cores, but there is an AVX2 with things like VNNI.
What cli tool was used to show the e core clocks?
@11:06 pretty sure this is s-tui
the ultimate test will be to run a benchmark test on a only one core (rest of cpu's in idle state) that has to traverse its data across cache and PCI to off chip memory .......get the benchmark results and then duplicate this whole process in parallel across all cores using all cashes and PCI lanes and see how much slower the original benchmark is affected by IO bottlenecks. MY GUESS IS THAT IT WILL BE BETWEEN 30-50 percent. In the end buss access is still all about the resources divided by the number of CPUs
I just hope they didn't compromise on security too much for the sake of performance, only time will tell. The efficiency tho is mighty impressive, they finally have something for the competition and not just brand bs marketing.
Is this Out Already to buy? Where can I order or when is it coming out?
It released on the 6th of June 2024, so its been quite a while meaning its probably out to buy. IDK where to buy these enterprise server chips though.
What sockets do the chips use? You have not actually stated THAT, in the video or ON the web site.
On Page 2 of the article linked in the reply. Thanks.
LGA4710 www.servethehome.com/intel-xeon-6-6700e-sierra-forest-shatters-xeon-expectations/2/
@@ServeTheHomeVideo Page 2 Paragraph 10? I guess you can call that a paragraph. I'll edit my other post. Thank you. Still would have liked to see this mentioned in the video.
14:10 So how efficient does this compare to Ampere Altra Arm server CPU?
We talk about it in the video and have a section of the main site article. This is much faster and uses DDR5/ PCIe gen 5, has accelerators, and 250W/ 144C. Of course Ampere can be cheaper and if you want to develop on Arm without going NVIDIA then it is still the best option out there. Perf/W this is a big upgrade
@@ServeTheHomeVideo Yeah, It has faster IO but then still. There is no number no graph to talk about their comparison in efficiency?
I love geeking out over the latest and greatest but I'm always curious to see what trickles down for the single bare metal server or up to 3 VMs for the SMB
So this is a beefed up n100? lol
A generation newer E-core, no iGPU, huge accelerators with things like QAT.
I keep hoping for a n100 but with like 64 pcie4 lanes with bifurcation support.
@@ServeTheHomeVideotoday Intel announced that Skymont E-Cores are 1.7x faster at the same power draw than Crestmont E-Cores. This CPU uses Crestmont E-Cores of course 😂😂😂
Calling it a beefed up n100 is harsh but fair.
As someone who spent the last two years writing firmware for the Birch Stream platform that these CPUs run on, this couldn't be farther from the truth. Aside from the microarchitecture and process node being completely different, the platform itself is completely different. You do not get full RAS capabilities and self-boot on the CPU with an N100. Also, the floorplan of the silicon is completely different. The N100 is monolithic while these CPUs have the compute tile in the middle with IO tiles flanking it on either side.
By your logic, an AMD Epyc Genoa is just a beefed up AMD Ryzen 7000 (which is ironically closer to the truth than your statement).
@@AlexSchendel So it's a beefed up n100 with some server features that are present in all silicon but not fused off in the server part. Thanks for confirming this.
Memory plugged into PCI slots! Takes me back to old PC-XT’s and microchannel PS2’s.
do we have any idea on pricing on these we run a decent amount of Hyper V VMs so this might be interesting to us!
More than these chips, I am excited about the ripple effect this release will cause in the second hand market when Cascade lake Xeons start flooding eBay at a much lower price than they currently are.
Spectre and meltdown exploited inherent properties of speculative execution and out-of-order execution. Certain variants might be eliminated by having SMT-free cores, but I don't think it eliminates all (or even most) of the surface area for those vulnerabilities.
Speculative execution and out of order execution are not related to hyper threading.
I see what you are saying now. Yeah STH is very wrong here. There's essentially no change in vulnerability to side channel attacks due to eliminating SMT.
My mind keeps coming back to Xeon Phi with those, and its atrocious 1T performance which made them annoying to work with for mixed HPC workloads. Back then, I thought they'd really do well with a single P-Core and I honestly don't see the difference with Sierra Forest. Care to explain? :)
The Atom core has evolved significantly since Airmont. The one on Sierra Forest is about as fast as a Skylake core, nothing groundbreaking but certainly fast enough for running housekeeping tasks.
Note that despite their apparent similarities, Sierra Forest doesn't target the same market as Knights Landing. It's more of hosting/cloud thing. A lot of its potential customers (especially those running vSphere) will not appreciate the scheduling complexity introduced by the addition of a P-core.
Finally, an interesting Intel Xeon product! However, with Turin 5c on TSMC N3 right around the corner, I don't think Intel products will bode well. AMD is going to have the performance lead with their 192c chip and I guess the perf/watt lead as well.
Turin Dense is Q4. Sierra Forest-AP is 288 cores in Q1. That will be the bigger battle.
@@ServeTheHomeVideo Xeon 6900 will be the same arch and process but with more core. At the moment, 64c 128t Epyc 4th gen is roughly equivalent to the Xeon 144c. Zen 5c will be on a new process, new arch and will have more core. It is clear to me that the Epyc 192c/384t will be superior, but I'm curious to see what the result will be at thread parity !
@@matthope_qc Well - TBF I think 288 cores will have more memory channels than 144 cores one. And more I/O in general probably. Yes, it's likely that Intel will still be behind AMD in performance. performance per Watt is still a bit of an open question (heard ton of good things about Intel3, some bad, some good things about N3). I certainly can't wait for benchmarks and various comparisons.
I think reading the Phoronix benchmarks against the top of the line EPYC skus the comparison (if you’re not comparing floating point & avx) is actually quite decent, so I definitely have a bit of hope that Intel and AMD will be pretty close for Turin Dense vs SF-AP.
@@ServeTheHomeVideo Considering Turin is Genoa SP5 compatible, I think AMD has a big advantage there over Intel, basically anyone on EPYC will remain on EPYC.
*CONGRATULATIONS DAD!!*
Thanks!
more cores. I dig it as long as I am not payin the power bill (yup, I know they sip power).
What is the audience/workload for a machine with that many (x2?) cores?
virtual host?
not crypto.
VM host, microservices, video transcoding (a lot still happens on CPU) and so forth.
That's a lotta cores!
THAT'LL BE 4 BUCKS BABY.
I complained last time about the title and thumbnail. I must say that I love this title and thumbnail. They are informative but also raise my curiosity. And since there is a lot of information that is in the video itself, there was no reason to be stingy on the former. Thanks Patrick.
And congratulations on the new member of the family! Sorry I was so harsh the previous video around the birth of your child. I did not know.
@1:35 Everyone needs to know, this is an Intel-sponsored video. If Patrick completely bashes their CPU, he won't get more $$, simple.
sponsored or not, this is a terrific chip. I have a two racks of 6 core Xeons, the bottom end of these chips with 2 servers would replace a bunch of servers.
Also, this is so exciting. I’m really hoping Intel can execute both on the fab side as well as design. I think Gelsinger is a good pick for the role and I’m rooting for him and the entire org. The more truly viable entrants we have in the market the better. I cheered for AMD when they started to mount a comeback and now I’m doing the same for Intel.
On the consumer side, we have repeatedly seen what happens when there is no real competition. Very minor bumps gen on gen.
what is the FPU performance like?
Will this complete with nvidia?
The NVIDIA Grace Superchip has soldered LPDDR5X memory and can hit 144 cores on the dual 72 core chip module. Those 144 cores with memory use around 500W versus Intel at 250W excluding memory but with the ability to scale to dual sockets at 288 cores. Very different CPU segments
So I think you are wrong on the 6700 series, I suspect it will be the primary sales driver. It’s a great replacement for data centers that have aging equipment that had lower power draws. I see this as the primary upgrade path for an enormous number of 5-6yr old racks/pods. As always, imho :)
I don't have any interest about server hardware at all but just like to watch due to your enthusiasm and excitement of talking about this stuff. Keep on Serving.
Ha thanks!
Is this compatible with w790 sage
No new platform.
Q: With more than 10X the core count of a Snapdragon X Elite laptop, what can it do in the TOPS count? Are desktops & servers still the #AI boss?
This is less of the AI chip. Granite Rapids (P-core) is better for that with AMX. Think of this as taking VMs off of legacy chips from many servers, and putting lots of those VMs and microservice onto a single server instead. That is why this is an exciting launch. Intel is, for the first time, really putting forward different domain specific CPU architectures.
ARM / Ampere killer ... go Intel !
Crazy
AMD Epyc Turin: Am I a joke to you?
@@aninditabasak7694 not a joke, but it's not out yet
Video begins at 2:28
9:52 No they're comparing this CPU to 64core 128 thread Epyc 9534 not 128 core Epyc.
I think the real intent (other than a move favorable comparison) to that choice was the TDP target of 250w. The Epyc 9534 default TDP is 280w and configurable down to 240W. The 128 Core ranges 320-400w for which it better be faster. On that basis the 84 core Epyc 9634 would've been the better comparison with the same TDP range as the 64 core.
Totally correct. Also you need to be at 32c for Genoa to be at 250W default TDP or less
Yes. Note added on that slide as I caught that just before uploading this morning
Congrats on the kiddo, STH+ST_Crib, I'm still running my home nas from before we had kids and the oldest is still years away from having independent screen time, but I'm hoping to see some ideas for engaging kids with real tech stuff early on. So far 3D printing is leading the way more than compute type stuff.
Kind the 7200 series Xeon Phi, from years ago.
Also somewhat the opposite of that! I remember doing the Knights Landing review. Our system went to LTT after.
@@ServeTheHomeVideo By that time, I think that was the idea, but they lacked acknowledgement or technology.
On my point of view now they perfected that old idea.
As many core as possible , on low consumption and a kind of weird configuration !
Nice. Congrats to Intel on delivering and beyond. I already went with AMD Siena, it has a lot of PCIe, and scales down to 8 cores, while keeping all io and memory channels. For crazy workloads, of course Bergamo. I can see Intel leaping ahead, and also having built in accelerators (compression, encryption), being a nice bonus, but I think long term these things will move to NICs (already available, just not super well integrated, and bit awkward to use it for compression for storage, on a NIC). I will stick to AMD, their more smaller tiles solution I think is more cost efficient, plus Intel screwed us many times in the past. But time will tell.
In general I really like the trend, of more smaller (more efficient) cores, because a lot of workloads I do, scales very well, and ultimately I can do more in same power budget (compared to less but faster cores).
Totally
Awesome overview! Looks like Intel is truly trying to catch up
Wow this plus Lunar Lake information has me Jazzed. I'm hoping the next Gen E core server CPUs might do clusters of E cores so they can turn off unused clusters like on Lunar Lake. Would be good for cloud folks and even better for us second hand home lab buyers down the road.
Congrats on the baby.
Thanks!
Lpe xeon when?
Interesting idea.
This is going to redefine servers? Others have long had the same or more optimizations in this direction. Seems more like Intel is catching up at best.
This puts Intel ahead of current Arm solutions and AMD. Perhaps catch up too, but they are now squarely ahead
@@ServeTheHomeVideo"Ahead" meaning what in terms of "redefining"? Even if the first true, if it is just about being a good product, then good for them and for us to have it as an option, but nothing that changes how servers are used. CXL has been long in the making and far from specific to this, chiplets are great, but hardly their innovation, more cores, each more efficient rather than performant, great but again just their version of what AMD and ARM have been pushing already, same goes for removing the PCH. Good to see competition doing what it is supposed to of course, but nothing that seems structurally new, other than being new for Intel.
And on that note, it's easy to look good, if you compare yourself to the competitions last gen. 5th gen Epyc has up to 192 cores & 384 threads. If the efficiency cores of Intel are about on par with last gens Epyc cores in performance, win some loose some, and a bit better in efficiency, that isn't that impressive, since 5th gen Epyc, especially the -c version will be on better nodes than their previous generation as well of course, while adding some improvements beyond passing on TSMCs I'm sure. As for Neoverse N1, it's from 2019, it better be defeated easily by a 2024 product. Granted Nvidia isn't realeasing their next gen yet, unlike AMD, but Grace is not new, and also has 144 cores and they aren't even from the lots of core N line, but the V one, maybe a little physically bigger as a package based on the less chiplet approach, but since that package includes memory, it's not actually large if you include that for the others.
All E Cores? that's great for web app hosting or VPS
That is one of the target markets.
Can I use E-Core as AI inferencing?
A compilation of compute competition
AMD announces 3nm EPYC Turin with 192 cores (384 threads) couple days ago.
Sure. Turin Q4 is after 128 P-core Granite Rapids-AP in Q3 and before Q1 288 core Sierra Forest.
Is there even demand for such CPU's out there?
Yes there is. In data centres the two most important things are space (limiting the number of racks needed) and energy (which effects cooling, power costs, etc). If a single 2u server can do the same work as a 24u rack, it can be worth it because you can expand without needing more square footage.
@@SpoonHurler you always need redundancy, for failover + at least one more DC. I cannot imagine an enterprise which would require several such systems. I mean i cannot imagine this planet having so many such enteprises to justify the creation of this CPU. Only may be IaaS providers.
The enterprise market is shrinking and now represents less than 50% of worldwide server shipments and this number is falling by the year. Cloud service providers are the majority share of server volume and this targets the primary workload used by the CSP customer base. General purpose P-core designs will become increasingly niche over time and this is represented in the server CPU roadmaps.
These days it is much more about power than space with all of the AI build-outs.
Intel 3 is 5-7nm tsmc equivalent?
Never expected to have cpus with as many cores as GPUs. Its not there yet but the gap is closing. Just crazy considering the number of cores just 10 years ago.
No Intel 3 has a higher real density than tsmc 5nm， Intel 3 is designed to compete tsmc n4 n3, not n5 n7.
@@timmy85211 Intel has density.. tsmc has efficiency. Intel can pack transistors closer but requires more power than the "equivalent" tsmc node. Power usage is a large part of the benefit of the node so a huge downside to Intel. "Same node" but twice the power means... Not an equivalent node imo
Wee need Sierra forest xeon-D parts for home lab.
Yes! One can argue that is the 64C Sierra.
No cinebench = no benchmark
Cinebench is Windows... eeek! c-ray is actually very close.
I'm just joking mate 😅
@ServeTheHomeVideo
that's a lot
Tons of cores
Before watching the video:
This is the "infamous" Sierra Forest, right?
I kind of wish they did 288 core version, but they were afraid of being too late on the market. Sadly from what I've heard they are still too late (by at least a quarter) and now they don't have their "flagship" product. Not that too many companies would buy 288 core version in big quantities. Also word on the "street" is that this is 144 E-cores. I really hope those are the updated Skymonts not the older ones.
Anyway, this is going to be interesting. I wonder how it will stack against Turin Dense (192 Zen5c cores). Also I do wonder how regular Turin will look.
EDIT:
After watching the video - I'm surprised they will have 288 core part. While I think the biggest product was above 300 cores, that was scrapped (I would have to look back on leaked road maps), but it's nice that they kept working on 288 core variant. Intel was always fantastic in innovation, just very variable in actually doing something with said innovation. And it is great to know that E-cores are on Intel 3. That should make them real efficient and it will be useful in places where power to performance is the most important (but you don't have to pay "per core" costs, so no VMWare). Sadly for Intel, this is still a bit too late (12-18 months) for them to retain large portion of their customers. Some went with Arm, others with AMD - depending on what they needed. If Emerald Rapids didn't come to the rescue in the fall of 2023, Intel would be looking at abysmal future. For now it just looks bad and they will have to leverage their Foundries heavily.
And also stoked to see that P-core versions will be available this year… I think. Intel's road maps have tendency to be very fluid, and I don't even remember all their code names anymore.
Granite Rapids-AP at 128 cores will be competing with Genoa 96 cores in Q3 of this year.
@@ServeTheHomeVideo That's good. I think it is still behind original schedule, but at least it is before AMD releases their new Architecture. Personally I dislike when those companies release stuff in "reactive" timing, something AMD and Intel did before - so usually waiting what other will announce and announce something better "just quarter later". But in this case, AMD is a quarter late and if we don't count really old Intel road maps, then this is more or less on time (though I think Sierra Forest was supposed to be Q1 even in last years road maps). There will e competition and the most important thing - there will be lots of SKU's that will fit most niches.
If only i could pick to disable my p cores on laptop automatically when on battery power.
You could try to force something close with processlaso.
Dam I'd still go epyc imo
This looks very exciting, but it does have me wondering if you consolidate like 12 servers to a single machine, are you not creating a much bigger single point of failure in your infrastructure?
I’m so excited for the future of Intel … the r/d have really did alot of breakthroughs .. powervia and ribbonfet are coming and 2027 will be produced in the new high N.A. machines !!
Hard to unpack what is just "marketing speak" when it comes to E and P cores...seems like E cores are possibly just 'P' cores that can't handle higher clock speeds rather than a fundamentally different architecture.
They are different. For example, they do not have AVX-512 and AMX, but have things like AVX2 VNNI
@@ServeTheHomeVideo Thanks, I should have looked it up rather than just make an assumption, ha ha
They are fundamentally different. They are about 1/4 the size for starters, and then also the differences in instructions. They're even designed by different teams at Intel.
Nice to see that Intel is stepping up on all areas! Was there a small Steve Jobs moment when Pat Gelsinger re-entered the company?...
A company like Intel doesn't work unless there's an engineer at the helm. It's time for the world to rid itself of the MBA mindset.
I wanna see the 288 core version against the all 196 all zen c core CPU
Yes but those are Q4 and Q1 products so still many months away.
Me patting my aging 6700k in the chasis,
"Maybe one day."
Actually.... that might still be the better CPU for gaming lol. These things aren't designed for gaming AT ALL (if you care about gaming).
@@MrHav1k Oh I know, my 6700k is in my home server.
In the mean time AMD announced Epyc with 192 cores and 384 threads lol
Turin is Q4. So the 128 core Turin in Q4 will actually see Granite Rapids-AP launch first with 128 cores. AMD will be behind next quarter unless they move Turin up. Turin Dense at 192 cores will be competing with 288 core Sierra Forest. 288 physical cores versus 192C/ 384T is not a straight win. We go into this in the video.
@@ServeTheHomeVideo Isn't Turin second half of 2024? Which basically means it can be launched next month
@@dota2tournamentss Yes, 2nd half '24, which could be as early as 1 month away, or as late as 6 months. AMD did fast-forward the Zen 5 Strix Point processors to next month, they may also have pushed forward Turin, but there's no way to know.
Also those are actual cores, not gimped "e" cores. Only benefit is from Intel not using Intel fabs this time lol
@@diamondlion47 It appears that Intel is dropping HT (hyper threading) in favour of using a single thread on each p-core, with the e-cores presumably taking on the role of what used to be a thread, sort of but not really, since e-cores are also not fully ISA compatible with the p-cores. AMD's "c" cores have the full ISA and threading, so there's no need for a much more complex and messy scheduling system to make their solution work.
All E cores would be super if could combine with an all P core Xeon.
That would be a step forward.
Not 2 new sockets in the same generation, and you get stuck on the same E or P core CPU
I think after intel showing how well skymont performs compared to cresmont-lp, I wouldn't be too keen on these chips and would rather wait a bit for skymont.
That perf per watt is CRAZY. Intel is back!!!
wait till it's benchmarked.
Already showed benchmarks in here.
Now I want a ram drive in my PC please ... Like the ones with ddr1 and 2 back in the day. Gigabyte had one 😅
We stopped ordering Intel servers. This isn't much of a change so not likely to start again.
Wow. That is a wild statement. This is an absolutely crazy change for server CPUs. Easily the biggest for Intel in two decades or more. Sorry. After using these (and you can see our Genoa/ Genoa-X/ Bergamo videos as well) I have no way to reconcile that comment with the servers we are showing.
@@ServeTheHomeVideo Perhaps for their specific use? Though it's possible that said person didn't watch the video. Or that change isn't for them. I agree that this is a huge change, just not for everyone. And other CPU's should come later. If there are no delays. For some it will be too late.
@@jannegrey I also look at it like by making the E-core CPU, it lets Intel do more on the P-core side since it does not have to focus on the lower performance market.
It's rather simple reasoning. If more, weaker, cores is your thing, AMD has a better solution. If less, more powerful single cores is your thing, AMD has a better solution. Why would I go Intel when either way, AMD has a better solution in their newly announced product stack? That Epyc 9534 processor has been out for 2 years now.
There is no compelling reason to switch back to Intel.
This might be "cool" for enthusiast reasons but it's not a viable option when compared to the competition.
@@ServeTheHomeVideo It is a lot of cores and a very good power target - Intel explicitly market the E-Cores for some sub-markets like web servers and CDNs. It could well be possible that these chips will not so hot for other tasks where Intel explicitly markets the P-Cores (databases, simulations, etc.). In the end the ARM servers chips with many cores and good efficiency suffered from the same issue and only limited pickup.
All the intel hate is wild in the comments. There is a reason not every company has replaced their intel servers with AMD. Intel has a TON of integrated features that AMD just doesnt have. Intel also has way more, more task optimized CPUs
Nope these days AMD supports all kinds of features and workloads supported by Intel, including AVX 512.
@@aninditabasak7694 That is one of a ton of features. From cryptographic engines, quicksync to traffic encryption and network offload
@@AlpineTheHusky You're wasting your time reasoning with some of the people man. They simplistically think "Intel quad cores for 7 years straight on desktop bad, AMD good" and apply that to the server space where It's not always all about core counts and PCIe lanes. Intel has decades of software optimizations and working with the ecosystem to make sure the latest SW runs smoothly on Xeon, plus they have their proprietary accelerators, their AMX AI engine in the cores.... that's not something the typical "moar cores moar better" people will easily pick up on or understand. They'll just see more cores or see that their 7800X3D on desktop wins gaming and then use that to say Intel is washed lol.
There is no way I would store any personal or business info on a cloud service. We have our own servers, and they dont run windows.
Wait.. but didn't AMD announced yesterday and EPYC CPU with 192 cores?
It's not out yet. As of this moment this is the highest core count CPU.. at least for x86 (IDK what Ampere or the custom ARM silicon that AWS, Google, etc. have as far as core count goes).
They announced a 192 core part currently slated for Q4. Intel 288 core is slated for Q1. AMD will also lose the P-core max core count in Q3 with Granite AP at 128 cores before 128c Turin is out.
Who cares 90% of clients won't touch AMD and their issues, hassles, shit drivers, updates, inferior designs so it's moot. Maybe one day clients will let me sell them cheap amd stuff but not now, IT buyers won't buy AMD, which is why Intel CPU's account for 80% of marketshare even after essentially not even trying to compete for a generation or three lol.
I'm sorry how many??? Oh.. 'efficient cores'. Sounds like those slower weaker cores used for medial tasks in BIGlittle chips?
It is hard to say that. Think roughly 2 E-cores = 2 Emerald Rapids P-core threads. Or if you have an Intel Xeon E5-2650 V4, just assume core for core replacement. They are slower than mainstream cores, but that is also the point to hit power savings targets.
The decade old Xeon Phi with many cores on a dedicated card is a better approach... can you question Intel why do that approach instead of this weaker main cpu core approach
Not sure what you mean? We reviewed Knights Landing and such many years ago (our box is the one LTT ended up using.) This is a much better approach. Also as a "weaker main CPU" it has higher integer performance at 250W than the 64 core top-end 5th Gen Xeon at 350W. Faster and 100W less TDP is hard to call weaker.
@@ServeTheHomeVideo Your article on 500+ cores per U in blade system for Clearwater Forest makes sense then for power savings // I was thinking Home Server- Small business which this chip series is not the target market... I can dream of a PCIE 6.0 /7.0 card with 512 cores someday though- Thanks
No AVX isn't going to appeal to the scientific compute folks
It is not intended for that market, but it does have an AVX2 with VNNI and such
@@ServeTheHomeVideo Ah does it now, I stand corrected - thanks!
Thank you....
sierra phirest?
im in
Whaaaaaaaaat!? Intel: Oh you want flexibility? I got ya fam.
Intel makes their own 14900k look stupid with their own cpu. Silly. 144core at 250 watts. verses over 300 watts. smh. Intel needs to slow down and do things right the first time.
That is the difference of pushing maximum performance per core versus maximum efficiency
Okay - "Key Lessons Learned' for this video - you can't make a full review in 2 weeks, while there is a newborn in said 2 weeks. So much work but congratulations. For now I will look at this more as a pre-review and I will check your website if there isn't anything more there (updates). Especially 3d party benchmarks. Take care! (YT deleted my comment, when I tried to put an emoji and Ctrl-Z didn't work, which is why it looks a bit weird).
Plus all the chaos at home might cause you to find out that you named your son by accident to "Tiny Mini Micro" or something like that 🤣
Spare some thread ripper sir...
TR is fewer cores
Too little too late, the damage has already been done, Epyc will crush the remaining Intel line.
I feel like this is the level of scaling that's difficult to visualize so I started writing things down. You're talking literally thousands of cores per rack now, right? For some reason the first emotion that comes to me is...fear?
Go another step. 1152 physical cores in a 2U 4-node design with only 2kW of CPU TDP
@@ServeTheHomeVideo 😵💫
More cores more better
So it’s what… running at 500MHz? Kidding kidding, I’ll watch the video now.
2.7GHz-3GHz easily
This actually seems like a sweet homelab part for people who want to play around with virtualization but don't need tons of single core perf. Basically this lets people get the performance of buying an old craigslist dell server without heating their entire home with it. Can even probably squeeze into 1U without fans that your neighbors will file noise complaints about.
I think the 96 and 64 core parts might be super for this as well. Also, these days 2U makes a lot of sense and lets you use quieter fans
@@ServeTheHomeVideo Yeah I agree I think for homelab the lower count parts are the sweet spot. And it's also a compelling part compared to something like Ampere M128-26.
They are really stepping up there game now.
Yea. Totally.
How does Patrick know anything about 'efficiency' when he drives a 6500lb+ vehicle that only gets one mile per kWh? I'm not sure we can take his reporting at face value after the Cybertruck revelations.
The low efficiency is just a downside you need to live with if you want to show your support for Elon‘s weird and fascist politics by buying a Cybertruck.
Closer to 3mi/ kWh (just over 330/mi). I put the screenshot online some time back. So the efficiency is REALLY good actually. Very strange comment.
Those 144 e-core chips aren't competing with Bergamo for the reasons mentioned in the video. They (and also the 288 core parts) are also not competing with Bergamo, because Bergamo is AMD's last gen part. Their actual competitor is going be Turin-Dense.
Correct, but Turin Dense is also 5+ months off at minimum.
@@ServeTheHomeVideo Also correct :) The 288 core parts won't be out much sooner than Turin-Dense though.
"MOST Server cores EVER Intel Xeon 6"
ServeTheHome you are technically correct, the best kind of correct.
My meaning being is that AMD's 192C 384T Zen 5 chip is coming soon, but it's not out yet, so technically correct.
We covered this recently since AMD pre-announced Turin Dense VERY early. Turin is ~SC24 timeline, so about 5 months away which puts Turin Dense more in the timeframe of Sierra Forest AP with 288 cores. Here you go: www.servethehome.com/server-core-counts-going-supernova-by-q1-2025-intel-amd-arm-nvidia-ampere/
That efficiency is crazy. Doing some back of the envelope math servers with these new chips could pay for themselves in pretty short order. Intel is back!
Yes
Intel looks desperate... E-cores only is kinda sad. Curious to it do a head 2 head with AMD's Epyc lineup
Why sad? This is a 250W CPU with 144 cores that has faster integer performance than the top-end 64 P-core Emerald Rapids Xeon at 350W. That is a crazy win. We have some charts versus Bergamo which is probably the closest competitor, but it is a higher power platform.
You have to remember the relative performance compared to the last generation of Xeons, in which this is a 3nm process and the last gen Xeons used the 10nm process. If you compare an E core of this generation to the P core from last generation, the E core from this generation would have a slight lead from 10nm P core as the E cores are usually around 44% the performance of a P core. This means that we have more performance, higher power efficiency, and more cores which is amazing.
AMD does the exact same thing except they dont say they are cut down cores
@@AlpineTheHusky amd has hyperthreading on each core... Threadcount is much higher.
@@ghty-kw7hm I don't really need to remember anything. I care about what cpu has the better performance, and currently the AMD Epyc lineup is better. Nice for intel thqt they had an uplift compared to the previous generation, but in the end, how they match up to Epyc is what matters. I am also curious about the price, because performance to purchase price has been lacking too the last years...
"Everything you can do I can do better" - AMD to Intel
This may have been the case over the past few generations, but with this and the new Granite Rapids-AP, that is no longer clearly the case. The Turin generation is going to be closer. It is getting a lot more nuanced starting with Sierra Forest.
Maybe if you ignore integration and accelerators.
Here i am running a 22 core X99 V4 chip
You have not compared AMD vs Intel in terms of of price, consumption, and benchmarks. What you have mentioned is Intel’s own opinion.
Huh? We had benchmarks and power in there. Price was not released until after this but we have it on the main site and we talk about platform costs here. Very strange comment