Hey everyone, some updated info. I’ve been speaking with AMD about this for the past few days and after giving them my results they now have their engineers looking into it so we should have some more info soon. AMD’s testing a wide range of games to see what impact the various memory allocation sizes have. I’ve also tested a few more games, for the most part very little difference, at most I’ve seen the average vary by up to 5% in some instances but it’s mostly 3% or less. Depending on what AMD finds (and other reviewers) I’ll follow up with more testing in a wide range of games using various quality settings/resolutions. But for now it seems like at best you’ll gain an extra 2-3 fps using 2GB’s opposed to 64MB’s, so I feel for users with 8GB’s or less memory you’ll be better off with the 64 - 512MB options. Just apply a small overclock to gain a few extra frames ;)
justin vistin i'm pretty sure gamers nexus did a test just for that not so long ago. if you are interested you should check yourself, but i think it only makes a difference with a dual titan V
+justin vistin for gaming or for stuff like raytracing? i think the 1080ti will be bottlenecked in most games (with lightly threaded nature) by the fact ryzen can't really clock over 4ghz ;) that is assuming @ 1080p, 4k is a different game, things turn back into a proper GPU bottleneck, and numbers across CPUs get much closer to each other there again. but no, as far as i've seen PCIe 3.0 x8 doesn't hold it back, it was all margin of error type stuff.
I'm really digging the way you show an almost blank graph and give us the results as you say them. Very nice improvement bro. Your videos are always informative and to the point. Keep up the good work.
This is why you are my go to source for tech videos. You're in depth, knowledgeable, and offer practical advice to your viewers. It's criminal that you're not more widely subscribed to as the other tech tubers focus on high end, or minimal benchmarking etc. I'm not even a budget gamer, far from it, but your videos are why I recommend your channel to everyone.
I’m glad you tested this, the comments section of every review has been full of people saying “MAKE SURE YOU SET THE VRAM TO 2gb IN THE BIOS!!!1!1!!!” Great work Steve.
yeah it all started from Jay's channel, he's such a big youtuber but his testings aren't really great, but since he's huge, everyone believe him I saw his video and literally my reaction was just "this is bullshit, it's the same memory, even if you don't allocate more to the gpu, it'll just spill to normal memory" and I got shit on for it by his fans
It may be useful to know that the theoretical peak bandwidth for Dual Channel (128bit) DDR4 3200 is actually 51.2GB/s, and while the effective speed will be a good deal lower than that, the advertised bandwidth for a dedicated GPUs memory bandwidth is also a theoretical peak. For the mathematically inclined of you, the theoretical bandwidth can be calculated by multiplying the effective clock speed in Gbps (clock in GHz) by the bus width in bits (there are 64bits per channel) and dividing by 8 to convert from Gb/s to GB/s.
Preach it. Windows 10 dynamically manages iGPU video memory (VidMm), so allocating system memory in the Bios largely isn’t worth it in most cases. Let windows do its thang 😄
You can also increase the limit of the gpu memory if you have more than enough RAM-16gb or more.The fact thats a budget build doesnt exclude that option. Very good and informative video.
Nikola Vasilev doubt a gamer with an apu has 16 gig of ram with this CRAZY prices that means he paid more than his cpu(2400g) or double the price of his cpu(2200g)
This is one of the only channels where I regularly learn something new and/or interesting about PC building and configuration. All the other PC channels just feel soft by comparison. Best PC tech channel is best.
Hey Steve, have you done any overclock testing? Funny thing is I just was setting my frame buffer size about 5 minutes before I started watching your video. 64mb should be enough anyways :) My 2400g is 4ghz stable with a decent cooler at 1.39 v and LLC set to auto. seems like I did ok. Video soon!
That's done it, I've been following you for months but this is the last straw for me. You spend too much dedication on all your benchs and for this reason I will subscribe to this channel. I hope you happy now.
My first try with my new ryzen 2400G ... I set to 2g ... FPS did not improve ... but did cut usable mem for other programs... set to 512mb, frame rate is at between 36 - 60+ and I can play most of the games I like... except those killer games that only run on Windbloz.... Great Video!!!
I think you are the best reviewer on RUclips right now. Since where every other reviewer is like allocating the 2Gb of Ram for Video Memory. You are the one who came first to address this Rumour.
HWU notification > like the video > watch it. Don't even have to question whether or not it's gonna be of the very best quality. Great work again HWU! You da real MVP!
Thanks for doing more tests on whether or not, and how, you can get more performance from Raven Ridge. Can you test 2 single rank, vs 2 dual rank, vs 4 single rank DIMMs next? That seems like the biggest thing that needs tested. Ideally at 2133, 2666, and 2933MHz for each of the 3 configurations, same timings, and a handful of games. I've only seen dual rank vs single rank on Raven Ridge tested in 2 games, and not in much detail. Top end Crucial Ballistix Sport would seem ideal to use as the dual rank RAM to test it.
Thank you so much for this, pretty much all the other benchmarkers tested exclusively with 16GB RAM despite it being ridiculously expensive while the CPUs are for budget builds. Can finally do a raven ridge build for my little bro.
Great video Steve! Results are around what I would expect, but there could be slightly higher CPU usage with lower VRAM as the operating system doesnt differentiate between dedicated VRAM and this APU-style partitioned VRAM, so it uses the same mechanisms and when you start spilling into system memory, it will still have to copy in and out of VRAM. Couple small things I cought, Vega's HBM does not run at 1.25 GHz, the default clock is around 950 MHz or around there and because HBM is DDR memory, the effective clock is double of that, so around 1.9 GHz, similar goes with GDDR5/x, which is QDR, so a base clock of 2.5 GHz would mean an effective clock of 10 GHz (AMD and NVidia cards report it differently, AMD would report 2.5 GHz and nvidia would report 5 GHz, but Im not aware of any GPU actually reporting the 10 GHz effective clock) PCI-E based GPUs arent limited by system RAM speed, 35 GB/s is way more than a x16 PCI-E 3.0 connection, which caps at 16 GB/s + regular draw calls and all that, so in the end you might only have around 10 GB/s for GPUsystem RAM communication. Lastly 35 GB/s isnt really accurate for 3200 MHz RAM, real world performance will vary based on a lot of things, such as how heavily youre hitting your CPU, how many cores it has, what programs are running, etc. Theoretical maximum performance of DDR4 3200 MHz is 51.2 GB/s in dualchannel mode and 25.6 GB/s in singlechannel.
This pretty much answers my two biggest questions regarding the Zen APUs: If 8gb dual channel is enough; and if is necessary to sacrifice 2gb of RAM for the iGPU, which would hurt a lot a system with only 8gb of RAM.
Steve your killing it on getting info out to the world. All the big tech tubers are busy deliding and comparing APU's to dedicated cards that cost more than the APU alone.
Nice work mate, really thorough benches. You could even say the 1-2 FPS delta are within the margin of error, so realistically, the difference is not statistically significant so it's basically the same. Also worth mentioning, is that APU gaming is on low/med settings where it's low quality assets in play in memory so it doesn't need much capacity. You could game perfectly fine with an 8GB budget build, with 2x 4GB sticks of 2933 DDR4 to save on money.
Fantastic review, I was not aware that the game will use the system memory as needed, whether you allocate 64 Mb or 1 Gb. I have always set the reserved memory to 2 Gb on my APU Systems that have 16 Gb of Ram.
This makes perfect sense. As always, do remember to use the highest speed ram you can grab, I'm personally seeing DDR4 3000 for sale around these parts for a decent (so to speak...) price. Great to know how to set this little beastie (I'm going for the R3) up beforehand, very appreciated! Thanks Steve :-)
Momi RAM is so "cheap" in your country, here 8gb 2400mhz (4gb x2) would cost 230 usd. But 3200mhz RAM would cost 247 usd, so no reason to not buy it for an APU. Actually is not RAM being cheap there but everything being ridiculously overpriced here, I spend 800 usd on a build with a 1050 ti, R3 1200, B350 and 8gb 2400mhz.
(I've asked a *ton* of questions in this post, please do look at it =) 1:44 The problem with a lot of tests of low end GPUs with little VRAM is that the system itself have a lot of RAM. When all the data needed by the GPU isn't located in the VRAM you end up with that data having to be stored in the system memory... And low end systems often have little system memory too... And trust me, when paging happens (especially to a HD and not SSD) the performance suffers... 6:25 Ok, granted I don't have experience with newer AMD APUs... But the old way that APUs from both Intel and AMD was done was that the GPU could *not* access the system memory directly, instead the GPU had to tell the CPU that it can't find the data it needs in the parts of the memory allocated to it, then the CPU has to move that data from the parts of the system memory it uses to the parts used by the GPU and only *then* can the GPU itself access that data... In *theory* with HSA the new AMD APUs *should* be able to access that data directly. But does it in fact do that without any special coding involved? What kind of performance penalty is there for using the minimum memory allocation when exceeding that allocated VRAM with the game assets? I'm curious about those results. 11:22 Alright, nice to see the results of your tests. =) However is it possible to do some kind of latency testing too? If the GPU has to request data from the CPU and can't get it directly from the system memory outside the alocated parts itself it might have to wait a little longer for the data even if the *amount* of data is just fine leading to higher latency. And another thing that could be fun testing is the impact on the GPUs bandwidth when the CPU is also loaded. When both are constrained, what is prioritized for instance. I'm guessing the CPU, but still would be a interesting test I think. And what about the frame rate impact in titles where *both* the CPU and GPU is loaded to the max? If the GPU has to wait for a lot of CPU tasks to finish before the data can be shuffled... Also, what about a higher priority CPU task of some kind running in the background, like a demanding driver? If the GPU drivers are high priority a even higher priority piece of software might cause issues... All in all there's a ton of things I'm curious about. Oh, and another thing. Does the APU have HSA enabled? AMD has advertised HSA capability in earlier APUs but not with this one, curiously... Is there any HSA compatible software you can test on the APUs?
Wow! Thank you, Mr. Your Host Steve! I did not know this. From what I've loosely picked up from claims on different fora, along with a lack of knowledge that made it seem "like common sense", I though more was best to a far bigger degree than your benchmarks show here. Of course, I try not to blindly follow one singe source of benchmarks :P but this was very interesting to say the least. Excellent information :D Thanks!
You are just about the best Tech Tuber out there.... And I know I give you some shit from time to time, BUT you are.. No one does the amount of testing that you do.. No other Tech tuber knew what you knew about the memory allocation not mattering... Everyone else said it had to be set at 2 GIG's for max performance.. But now we know it doesn't matter and it really doesn't matter to me because, I already have a 1700/3.9 in a Taichi motherboard with an Rx 480, ( because GPU's are way to fucking expensive right now) but anyway, alot of PPL are wanting to know if we should flash to the latest BIOS even tho we have Gen 1 Ryzen chips, Like the 1600 and 1700.. So can we update our BIOS to the latest Bios without hurting performance and stability with our Gen 1 chips in order to be more prepared for the Gen 2 chips in April... I know with Asrock there's a bridge BIO's and another Bios after that in order to use the Vega APU's, But seriously fuck those APU's, Most PPL only care about, weather or not they are going to be able to keep their X370 motherboards and get the same performance out of overclocking as the x470's that are coming out???
Nice job on the video and excellent explanation. I agree with you wholeheartedly about the reason that AMD would want to tout 4GB. In this era of 8GB video cards, more must be better. A paltry 512MB or 64MB sounds like we're back to the Diamond Viper days of the 1990s. Again, well done!
Great video. Never heard of the channel I was already sold 3 minutes in. Channels that briefly provide comparisons and spend the most time explaining and discussing their testing are fantastic but so rare. I'm interested in the system on the left, has he made a build video of it?
Steve I'll take you back to 2010 when I got my first 'gaming' PC, it was based of an ECS (Your favorite brand) AM3 board. It had an integrated HD 3300 ATI chipset, and I tweaked anything possible to make gaming experiences better, overclocking, setting dedicated Ram etc, and I remember increasing the memory allocation gave a big boost. Perhaps all is not what it seems with this RR allocation, with early firmware and all it could be that the settings did in fact not take place/work as intended.
Honestly I'd have to install Windows 7 and check, haven't used it properly in a very long time. I don't think there is a Win 7 Raven Ridge APU driver though.
the statement that you made about games that require a certain amount of vram to even launch very interesting point and that could be a reason for the higher vram settings so they do not lock some people out of there games.
Thank you so much for this. Answered all my remaining questions about Raven Ridge. And I've gotta say that Vega 8 GPU is the real star in these APUs. Overclocked it looks like it maxes out what an APU can do, and it would appear that the system itself can't really push past 8GB of RAM. It's the perfect APU for a really small budget.
Hi Steve. It might be worth noting that while you don't see any real performance change using lowest possible dedicated VRAM setting in most games, you might see significant reduced performance in workload test with the highest possible VRAM setting as effectively reducing total memory for the CPU from 8GB to 6GB can have quite a large negative effect. Of course, if you have 16GB RAM it wouldn't hardly matter but I figure that most people buying a budget system will try and get the fastest 8GB RAM kit they can find. Another test of possible negative impact of highest possible VRAM settings on a 8GB system, perhaps?
I have just ordered parts to build an AMD apu system, thanks for the great video as this will help setting up the PC for best performance for the price
Even if windows managed the system/VRAM dynamically I feel like the purpose of manually setting a reservation is sort of lost on people. A few things to think about. Near Capacity - If your video is more important and you don't want stability issues of the system fighting back and forth to share that memory when your near memory capacity... Set it manually to slightly higher than the minimum you need to ensure your video performs smoothly. Latency - When your system memory is used for an IGP dynamically, the location of those assets are constantly being shifted around. This adds latency because of the slow system memory that has to refresh much more often. Your memory's latency (timings) are going to be very important. When set manually, that 2GB portion is reserved and registered to the IGP which then knows EXACTLY where to store it's assets, no waiting for a refresh(s) on the memory. Stability - In rare cases, certain applications would use memory space that overlapped with one another and can cause a program crash. While this is much more uncommon these days thanks to modern OS memory management, it can still happen with a badly designed piece of software. Overall it is best for the system to have a dedicated portion of memory to store and work out of for video resources. I'm curious to see your results if you test again and pay attention to the 99th percentile frame rate latency and load times on things like maps or load areas in games. Also check to see if you notice a difference between the high and low frames, not the average. Granted, these things may not be very noticable to most users and be of minimal benefit for the average user. The obvious best option if you are worried about performance and using more than 8gb for both system and video is to buy 16GB. You would also gain a marginal boost in performance from having more dense DIMMs assuming the same latency timings. Price is of course a limiting factor thanks to RAM prices these days, but what is new.
Long time ago, before most of ya'll were born. Setting the frame buffer size determined which resolutions were available to you in Windows, Linux, OS/2.
very interesting work Steve. i found one limitation of the low capacity framebuffer already: i tried to start ETH mining on my new 2200g for the heck of it. and it seems like the max 2GB VRAM does prevent the miner from placing the DAG file into the framebuffer. i suppose you would need to allocate 3GB to archive that. a very theoretical problem though ;)
I would like to see the difference between 3200mhz ram and say 2400mhz or 2666mhz as high speed ram is so expensive right now. Also dual channel vs single channel.
Allocating 2G memory for older APUs (A-Series) actually helps a lot in boosting the iGPU performance. If I leave that setting to "Auto" which is 1G, my PC would have a hard time running a lot of games. Setting it to 2G gives me decent frame rates but of course some of those memory go to waste when I don't need the GPU's power. So seeing these results is actually good.
Thanks updated bios to UMA FIXED 64m for my 5600g. Its mostly used for hypervisor use so will appreciate the extra ram. As it has 32gig ram it auto sized to a bit 2 gig.
This was very informative. when the first load of benchmarks came out and everyone was testing with 16 gb kits. i did start to wonder if 8 gb kits would comprise performance with the shared ram. turns out that isnt true. However i think a 1200 + gtx 1050 + 8gb ram offers more bang for your buck gaming wise than a 2200g + 16gb ram
D-GPU processes 1GB data then has to move it to the CPU or CPU processes 1GB data then moves it to D-GPU in each case over the available bandwidth. APU has UMA, I-GPU processes 1GB data leaves it where it is (in system memory) and passes a pointer giving the location of the data to the CPU or vice versa. So in an APU if operating system and drivers are working properly a lot of data transfers should never need to occur. Are AMD delivering on the promise of UMA or are they forcing the APU to mimic the more traditional CPU/D-GPU setup?
I understand that this is not the main topic of the video, it's just a small comment on what you say around 2:25, because in reality the hit should be even harder for the 550 than 30GB/s once it's out of onboard ram, because of PCIE limitations, it should be more like the read/copy on Aida speed I guess, anyway, I understand that the relevance of this is very limited for this video, since the topic is the IGP, and in that case yes, the memory access over the "dedicated" amount should be just as fast as the dedicated portion, as your video proves, anyway, not trying to be annoying, good job with the video showing the fixed amount of ram on Vega IGP is not really relevant for graphics performance, which lots of people seem to believe it is.
Sorry you are completely right, I misunderstood your comment. Yes, I overlooked the obvious here, you are indeed limited by the PCIe bus before the system memory bandwidth. Thank you for pointing that out, I'll try not to make that mistake again.
Yes, that is the point I tried to make earlier in a conversation w/ you Steve. The important thing your quoted test numbers revealed, & you seemed to have overlooked, that the alleged 8 lane gpu link on the Fabric bus apu, in fact performs like a ~24 lane link to system ram, as I would expect from a miniaturised new gen bus. Ie., on occasions when either discrete or igp DO have to link to system RAM, as BOTH do, the apu igp is in fact at a great ADVANTAGE over discrete. In other situations, the apu igp is not at as big disadvantage as commonly perceived. Needless to say, this fasterFabric apu bus also applies to cpugpu links, - a further apu edge.
it's a good point, I noticed on the Digital Foundry 2200G video, he was pushing high textures in games that normally kills the performance for 2GB discrete cards just fine, basically the game was using over 2GB even if he had the IGP on the bios set to allocate that amount, and since there is no PCIE bottleneck it handles it OK with the dynamic allocation, it's a fairly interesting point, now I look at the GT 1030 and it uses a x4 link only, it's kind of something that could perhaps be interesting to explore.
I think it's basically the same problem as with CVTs on cars; a true optimized CVT will hold the engine at the best performance rpm for any given situation (economy, torque or power) so that you're always getting the most out of your engine, but almost nobody would buy a car with such a CVT because according to customers they sound boring/broken. The actual performance is better than with any other transmission, but it's much easier to add virtual gears than convince customers that the car isn't broken and explain why it has the best acceleration at a constant engine speed.
Thank you very much, great video. If I had access to your array of hardware it would be interesting to see if the rx550 suffered worse than the vega 8 in superposition if it were forced to use say 4-5 gigs of vram on a 2 gig card. Basically determine if the vega 8 can access system memory faster than the rx550 can?
Hey Steve, did you try setting everything to auto and seeing what the performance was? See how good the AMD drivers are at allocating the memory for game use?
That's not what the auto setting does mate. On auto the motherboard just detects how much RAM you have and sets accordingly. So it you have 8GB's most boards will set it to somewhere between 512MB and 1GB.
You might want to talk to JayzTwoCents and GamersNexus and help them get their Ryzen APU setups working, they've been having so many issues they can't even benchmark anything.
I think the difference between the theoretical peak of DC DDR4-3200 and the measured 30GB/s comes from the limitations of the PCI-E interface. PCI-E 3.0 can deliver up to 16GB/s per direction, which may explain the 30GB cap. Correct me if I'm wrong.
I searched in vain for a chart, but I have seen the results for openCL on these apuS, and they are outstanding, as you would expect from such tightly integrated cpu/gpu.
I have a guess. If someone's CPU is completely saturated and have spare RAM they might get a noticeable boost with a higher buffer setting. Maybe it's DMA so there is no CPU involvement because the GPU is handling it itself. Even on high end systems, if there are stutters do to overwhelming CPU usage, there's a very slight chance that making that buffer bigger might help.
but there are lots of problems with adobe collection.. premiere cannot enable openCL by itself, illustrator cc greyed out GPU acceleration, and photoshop always show us a notification about less of vram at launch.
![Felix "Unfair" | [Stray Kids : SKZ-PLAYER]](http://i.ytimg.com/vi/Oswujxm2Ag0/mqdefault.jpg)
Hey everyone, some updated info. I’ve been speaking with AMD about this for the past few days and after giving them my results they now have their engineers looking into it so we should have some more info soon. AMD’s testing a wide range of games to see what impact the various memory allocation sizes have. I’ve also tested a few more games, for the most part very little difference, at most I’ve seen the average vary by up to 5% in some instances but it’s mostly 3% or less.
Depending on what AMD finds (and other reviewers) I’ll follow up with more testing in a wide range of games using various quality settings/resolutions. But for now it seems like at best you’ll gain an extra 2-3 fps using 2GB’s opposed to 64MB’s, so I feel for users with 8GB’s or less memory you’ll be better off with the 64 - 512MB options. Just apply a small overclock to gain a few extra frames ;)
awesome bro thanks
justin vistin
nope x8 doesn't limits 1080ti bcoz PCI spec is way ahead of it's time
justin vistin
i'm pretty sure gamers nexus did a test just for that not so long ago. if you are interested you should check yourself, but i think it only makes a difference with a dual titan V
+justin vistin
for gaming or for stuff like raytracing?
i think the 1080ti will be bottlenecked in most games (with lightly threaded nature) by the fact ryzen can't really clock over 4ghz ;)
that is assuming @ 1080p,
4k is a different game, things turn back into a proper GPU bottleneck, and numbers across CPUs get much closer to each other there again.
but no, as far as i've seen PCIe 3.0 x8 doesn't hold it back, it was all margin of error type stuff.
I am very curious aboout what the "auto" setting for the frambuffer does.
I'm really digging the way you show an almost blank graph and give us the results as you say them. Very nice improvement bro. Your videos are always informative and to the point. Keep up the good work.
Jeff Green agreed. Hope everyone does this as it is easier for me to connect what you talking about with the results of your finding.
This is why you are my go to source for tech videos. You're in depth, knowledgeable, and offer practical advice to your viewers. It's criminal that you're not more widely subscribed to as the other tech tubers focus on high end, or minimal benchmarking etc. I'm not even a budget gamer, far from it, but your videos are why I recommend your channel to everyone.
You're so close to 200K subscribers! Good on you Steve!
I’m glad you tested this, the comments section of every review has been full of people saying “MAKE SURE YOU SET THE VRAM TO 2gb IN THE BIOS!!!1!1!!!” Great work Steve.
yeah it all started from Jay's channel, he's such a big youtuber but his testings aren't really great, but since he's huge, everyone believe him
I saw his video and literally my reaction was just "this is bullshit, it's the same memory, even if you don't allocate more to the gpu, it'll just spill to normal memory" and I got shit on for it by his fans
Vegetables are good for you doesn’t surprise me unfortunately people always jump to conclusions especially in the pc gaming biz
Douglas m if I may ask what is the general settings to make that happen on the apu.( I know bios varies on motherboard).
Well every bios will be a little different if you see in the video here on a gigabyte board it is uma frame buffer size
Douglas m thanks for the reply!!!
I have to say you are my favorite youtuber when it comes to benchmarking. Your tests go much more in depth than other channels.
It may be useful to know that the theoretical peak bandwidth for Dual Channel (128bit) DDR4 3200 is actually 51.2GB/s, and while the effective speed will be a good deal lower than that, the advertised bandwidth for a dedicated GPUs memory bandwidth is also a theoretical peak. For the mathematically inclined of you, the theoretical bandwidth can be calculated by multiplying the effective clock speed in Gbps (clock in GHz) by the bus width in bits (there are 64bits per channel) and dividing by 8 to convert from Gb/s to GB/s.
Preach it. Windows 10 dynamically manages iGPU video memory (VidMm), so allocating system memory in the Bios largely isn’t worth it in most cases. Let windows do its thang 😄
F2F Tech dynamic memory is depend on the motherboard chips and bios.
It is up to the UEFI. Even UEFIs have a set limit on how much it can allocate. Also dynamic allocation causes latency spikes btw
:) That's just in theory, in practice it's allocating just the minimum.
F2F Tech did you think that not everyone uses windows?
re FLIPd Of course. I use flavors of Linux myself, but for gaming Windows is still the majority.
You can also increase the limit of the gpu memory if you have more than enough RAM-16gb or more.The fact thats a budget build doesnt exclude that option. Very good and informative video.
Nikola Vasilev doubt a gamer with an apu has 16 gig of ram with this CRAZY prices that means he paid more than his cpu(2400g) or double the price of his cpu(2200g)
Maybe people already have the fast ddr4 ram from before the pricing apocalypse.
@nikola vasilev then is it worth the 16gb investment?, im planning on building with the 2200g and I normally have 2 or 3 Windows opened while playing
Wow it's been a while since I've seen a genuinely interesting video
Great work guys
You can actually learn alot of things on this channel that you might not have a decent knowledge of. Thanks Steve!
Another excellent video from the best PC + Tech channel on RUclips!
seeing the MSRP on the video cards still hurts. i've seen 550s going for $250 in the used market near me in the US. and used vega 64s for $1200
Happy New year!! Prices bring a smile now
This is one of the only channels where I regularly learn something new and/or interesting about PC building and configuration. All the other PC channels just feel soft by comparison. Best PC tech channel is best.
Hey Steve, have you done any overclock testing? Funny thing is I just was setting my frame buffer size about 5 minutes before I started watching your video. 64mb should be enough anyways :) My 2400g is 4ghz stable with a decent cooler at 1.39 v and LLC set to auto. seems like I did ok. Video soon!
Yeah I played around with overclocking in the launch day review.
Excellent video Steve :)
Thanks mate.
You deserve more subscribers, subscribed last year. Good job 🖒
Thanks mate!
That's done it, I've been following you for months but this is the last straw for me.
You spend too much dedication on all your benchs and for this reason I will subscribe to this channel.
I hope you happy now.
Thanks mate, appreciate the support and kind words!
Wow Steve. Nonstop testing man. Awesome stuff. Keep up the great work!
The most appropriate and real review of the amd Apu, I bet there cannot be any better review of amd raven ridge apus in the near future as well.....
You forget about long haired Steve...
First of all sir you are amazing. I keep the UMA Buffer size at 64mb and it handles everything effortlessly
My first try with my new ryzen 2400G ... I set to 2g ... FPS did not improve ... but did cut usable mem for other programs... set to 512mb, frame rate is at between 36 - 60+ and I can play most of the games I like... except those killer games that only run on Windbloz.... Great Video!!!
i swear to good you are the only real benchmarker on RUclips
Thank You! This answers so many questions! Your Raven Ridge content has been super good.
What a great channel for hardware knowledge, also your english is easy to understand. Keep up the good job dude.
LOVED THIS REVIEW AND EXPLANATION! Highly informative and stayed engaged the whole time.
Steve let me tell you, size always matters.
That's what APUs always say, amiright?
I think you are the best reviewer on RUclips right now. Since where every other reviewer is like allocating the 2Gb of Ram for Video Memory. You are the one who came first to address this Rumour.
Thanks mate.
Hardware Unboxed Compliment doesn't cost a penny. 😇✌
HWU notification > like the video > watch it. Don't even have to question whether or not it's gonna be of the very best quality.
Great work again HWU! You da real MVP!
Thanks for doing more tests on whether or not, and how, you can get more performance from Raven Ridge. Can you test 2 single rank, vs 2 dual rank, vs 4 single rank DIMMs next? That seems like the biggest thing that needs tested. Ideally at 2133, 2666, and 2933MHz for each of the 3 configurations, same timings, and a handful of games. I've only seen dual rank vs single rank on Raven Ridge tested in 2 games, and not in much detail. Top end Crucial Ballistix Sport would seem ideal to use as the dual rank RAM to test it.
Nice and clean as always in your assertions. Very good job mate!
Thank you so much for this, pretty much all the other benchmarkers tested exclusively with 16GB RAM despite it being ridiculously expensive while the CPUs are for budget builds.
Can finally do a raven ridge build for my little bro.
I was so misinformed by Jayz.
Thanks for releasing this video.
Great video Steve! Results are around what I would expect, but there could be slightly higher CPU usage with lower VRAM as the operating system doesnt differentiate between dedicated VRAM and this APU-style partitioned VRAM, so it uses the same mechanisms and when you start spilling into system memory, it will still have to copy in and out of VRAM.
Couple small things I cought, Vega's HBM does not run at 1.25 GHz, the default clock is around 950 MHz or around there and because HBM is DDR memory, the effective clock is double of that, so around 1.9 GHz, similar goes with GDDR5/x, which is QDR, so a base clock of 2.5 GHz would mean an effective clock of 10 GHz (AMD and NVidia cards report it differently, AMD would report 2.5 GHz and nvidia would report 5 GHz, but Im not aware of any GPU actually reporting the 10 GHz effective clock)
PCI-E based GPUs arent limited by system RAM speed, 35 GB/s is way more than a x16 PCI-E 3.0 connection, which caps at 16 GB/s + regular draw calls and all that, so in the end you might only have around 10 GB/s for GPUsystem RAM communication.
Lastly 35 GB/s isnt really accurate for 3200 MHz RAM, real world performance will vary based on a lot of things, such as how heavily youre hitting your CPU, how many cores it has, what programs are running, etc. Theoretical maximum performance of DDR4 3200 MHz is 51.2 GB/s in dualchannel mode and 25.6 GB/s in singlechannel.
This pretty much answers my two biggest questions regarding the Zen APUs: If 8gb dual channel is enough; and if is necessary to sacrifice 2gb of RAM for the iGPU, which would hurt a lot a system with only 8gb of RAM.
This was certainly eye opening. Good work Steve!
Steve your killing it on getting info out to the world. All the big tech tubers are busy deliding and comparing APU's to dedicated cards that cost more than the APU alone.
I've been revisiting these Ryzen 5 2400g videos as possibly buying one soon- best coverage!
Nice work mate, really thorough benches. You could even say the 1-2 FPS delta are within the margin of error, so realistically, the difference is not statistically significant so it's basically the same. Also worth mentioning, is that APU gaming is on low/med settings where it's low quality assets in play in memory so it doesn't need much capacity. You could game perfectly fine with an 8GB budget build, with 2x 4GB sticks of 2933 DDR4 to save on money.
Fantastic review, I was not aware that the game will use the system memory as needed, whether you allocate 64 Mb or 1 Gb. I have always set the reserved memory to 2 Gb on my APU Systems that have 16 Gb of Ram.
Great video. Thank you
This makes perfect sense. As always, do remember to use the highest speed ram you can grab, I'm personally seeing DDR4 3000 for sale around these parts for a decent (so to speak...) price. Great to know how to set this little beastie (I'm going for the R3) up beforehand, very appreciated! Thanks Steve :-)
Momi Welp. good for you that those are the prices you're getting. here it's about 220+ for 16gb 2133 ram and 315+ for 16gb 3000 ram
Momi RAM is so "cheap" in your country, here 8gb 2400mhz (4gb x2) would cost 230 usd. But 3200mhz RAM would cost 247 usd, so no reason to not buy it for an APU.
Actually is not RAM being cheap there but everything being ridiculously overpriced here, I spend 800 usd on a build with a 1050 ti, R3 1200, B350 and 8gb 2400mhz.
Naizuri
Quite good. I pay 21% VAT or 27% if at my dads. That really adds up.
I just opted for 4GB of absurd speed ram.
Michael Livote will 8GB of RAM at 3200Mhz be enough for games like GTA V and DOOM?
Good job m8. I just bought a 2400G and was thinking about this. Thnx for the info.
Awesome video, I was a bit confused with all this memory thing and now it's crystal clear!
Thanks mate for clearing Misconceptions.
(I've asked a *ton* of questions in this post, please do look at it =)
1:44
The problem with a lot of tests of low end GPUs with little VRAM is that the system itself have a lot of RAM.
When all the data needed by the GPU isn't located in the VRAM you end up with that data having to be stored in the system memory...
And low end systems often have little system memory too...
And trust me, when paging happens (especially to a HD and not SSD) the performance suffers...
6:25
Ok, granted I don't have experience with newer AMD APUs...
But the old way that APUs from both Intel and AMD was done was that the GPU could *not* access the system memory directly, instead the GPU had to tell the CPU that it can't find the data it needs in the parts of the memory allocated to it, then the CPU has to move that data from the parts of the system memory it uses to the parts used by the GPU and only *then* can the GPU itself access that data...
In *theory* with HSA the new AMD APUs *should* be able to access that data directly.
But does it in fact do that without any special coding involved?
What kind of performance penalty is there for using the minimum memory allocation when exceeding that allocated VRAM with the game assets?
I'm curious about those results.
11:22
Alright, nice to see the results of your tests. =)
However is it possible to do some kind of latency testing too?
If the GPU has to request data from the CPU and can't get it directly from the system memory outside the alocated parts itself it might have to wait a little longer for the data even if the *amount* of data is just fine leading to higher latency.
And another thing that could be fun testing is the impact on the GPUs bandwidth when the CPU is also loaded.
When both are constrained, what is prioritized for instance.
I'm guessing the CPU, but still would be a interesting test I think.
And what about the frame rate impact in titles where *both* the CPU and GPU is loaded to the max?
If the GPU has to wait for a lot of CPU tasks to finish before the data can be shuffled...
Also, what about a higher priority CPU task of some kind running in the background, like a demanding driver?
If the GPU drivers are high priority a even higher priority piece of software might cause issues...
All in all there's a ton of things I'm curious about.
Oh, and another thing.
Does the APU have HSA enabled?
AMD has advertised HSA capability in earlier APUs but not with this one, curiously...
Is there any HSA compatible software you can test on the APUs?
Wow! Thank you, Mr. Your Host Steve! I did not know this. From what I've loosely picked up from claims on different fora, along with a lack of knowledge that made it seem "like common sense", I though more was best to a far bigger degree than your benchmarks show here. Of course, I try not to blindly follow one singe source of benchmarks :P but this was very interesting to say the least. Excellent information :D Thanks!
Thanks mate appreciate that and glad you enjoyed it.
Thank you for clearing out the confusion about dedicated memory!
You man, you are making really scientific videos! Thanks!!! Intro-results-discussion is on point 👌🏼
myth buster for sure lol great content sir!
this industry needs more of people like you.
Nice video! U deserve more subs for that hard work!
Steve, the King of Benchmarks delivers again!
Oh, and congrats on hitting 200K subs ;)
back when I used integrated video, in the early 2000s, I could set 8MB or 256MB and it was the same too...
ah.. the good ol' days with my ATI Radeon Xpress igpu..
You are just about the best Tech Tuber out there.... And I know I give you some shit from time to time, BUT you are.. No one does the amount of testing that you do.. No other Tech tuber knew what you knew about the memory allocation not mattering... Everyone else said it had to be set at 2 GIG's for max performance.. But now we know it doesn't matter and it really doesn't matter to me because, I already have a 1700/3.9 in a Taichi motherboard with an Rx 480, ( because GPU's are way to fucking expensive right now) but anyway, alot of PPL are wanting to know if we should flash to the latest BIOS even tho we have Gen 1 Ryzen chips, Like the 1600 and 1700.. So can we update our BIOS to the latest Bios without hurting performance and stability with our Gen 1 chips in order to be more prepared for the Gen 2 chips in April... I know with Asrock there's a bridge BIO's and another Bios after that in order to use the Vega APU's, But seriously fuck those APU's, Most PPL only care about, weather or not they are going to be able to keep their X370 motherboards and get the same performance out of overclocking as the x470's that are coming out???
1:00AM HWunboxed video? Count me in!
Very good info Steve.
Nice job on the video and excellent explanation. I agree with you wholeheartedly about the reason that AMD would want to tout 4GB. In this era of 8GB video cards, more must be better. A paltry 512MB or 64MB sounds like we're back to the Diamond Viper days of the 1990s.
Again, well done!
Love your videos always very helpful, thank you for your time and effort
Great video. Never heard of the channel I was already sold 3 minutes in. Channels that briefly provide comparisons and spend the most time explaining and discussing their testing are fantastic but so rare. I'm interested in the system on the left, has he made a build video of it?
Amazing video Steve!.
Most useful ryzen APU video for me good work!!
Very informative video Steve, thank you :)
Steve I'll take you back to 2010 when I got my first 'gaming' PC, it was based of an ECS (Your favorite brand) AM3 board. It had an integrated HD 3300 ATI chipset, and I tweaked anything possible to make gaming experiences better, overclocking, setting dedicated Ram etc, and I remember increasing the memory allocation gave a big boost.
Perhaps all is not what it seems with this RR allocation, with early firmware and all it could be that the settings did in fact not take place/work as intended.
Were you using Windows 10? ;)
Interesting point, no I was on Win7, is there a big difference in how memory is managed?
Honestly I'd have to install Windows 7 and check, haven't used it properly in a very long time. I don't think there is a Win 7 Raven Ridge APU driver though.
the statement that you made about games that require a certain amount of vram to even launch very interesting point and that could be a reason for the higher vram settings so they do not lock some people out of there games.
Thank you so much for this. Answered all my remaining questions about Raven Ridge.
And I've gotta say that Vega 8 GPU is the real star in these APUs. Overclocked it looks like it maxes out what an APU can do, and it would appear that the system itself can't really push past 8GB of RAM. It's the perfect APU for a really small budget.
Hi Steve. It might be worth noting that while you don't see any real performance change using lowest possible dedicated VRAM setting in most games, you might see significant reduced performance in workload test with the highest possible VRAM setting as effectively reducing total memory for the CPU from 8GB to 6GB can have quite a large negative effect.
Of course, if you have 16GB RAM it wouldn't hardly matter but I figure that most people buying a budget system will try and get the fastest 8GB RAM kit they can find.
Another test of possible negative impact of highest possible VRAM settings on a 8GB system, perhaps?
The word “solution” is sweeping through the Aussie tech channel world!
I have just ordered parts to build an AMD apu system, thanks for the great video as this will help setting up the PC for best performance for the price
Did you take what Adore tv takes? That was a really informative video. Gj!
haha thanks mate :D I like to think we do test kinds of videos every now and then.
Good Morning from Germany ✌🏻
Even if windows managed the system/VRAM dynamically I feel like the purpose of manually setting a reservation is sort of lost on people. A few things to think about.
Near Capacity - If your video is more important and you don't want stability issues of the system fighting back and forth to share that memory when your near memory capacity... Set it manually to slightly higher than the minimum you need to ensure your video performs smoothly.
Latency - When your system memory is used for an IGP dynamically, the location of those assets are constantly being shifted around. This adds latency because of the slow system memory that has to refresh much more often. Your memory's latency (timings) are going to be very important. When set manually, that 2GB portion is reserved and registered to the IGP which then knows EXACTLY where to store it's assets, no waiting for a refresh(s) on the memory.
Stability - In rare cases, certain applications would use memory space that overlapped with one another and can cause a program crash. While this is much more uncommon these days thanks to modern OS memory management, it can still happen with a badly designed piece of software.
Overall it is best for the system to have a dedicated portion of memory to store and work out of for video resources. I'm curious to see your results if you test again and pay attention to the 99th percentile frame rate latency and load times on things like maps or load areas in games. Also check to see if you notice a difference between the high and low frames, not the average.
Granted, these things may not be very noticable to most users and be of minimal benefit for the average user. The obvious best option if you are worried about performance and using more than 8gb for both system and video is to buy 16GB. You would also gain a marginal boost in performance from having more dense DIMMs assuming the same latency timings. Price is of course a limiting factor thanks to RAM prices these days, but what is new.
Amazing work, and quite eye opening!
Long time ago, before most of ya'll were born. Setting the frame buffer size determined which resolutions were available to you in Windows, Linux, OS/2.
very interesting work Steve. i found one limitation of the low capacity framebuffer already: i tried to start ETH mining on my new 2200g for the heck of it. and it seems like the max 2GB VRAM does prevent the miner from placing the DAG file into the framebuffer. i suppose you would need to allocate 3GB to archive that. a very theoretical problem though ;)
I would like to see the difference between 3200mhz ram and say 2400mhz or 2666mhz as high speed ram is so expensive right now. Also dual channel vs single channel.
That's all in the day one review.
Douglas m there is significant difference between 2400 and 3200
Allocating 2G memory for older APUs (A-Series) actually helps a lot in boosting the iGPU performance. If I leave that setting to "Auto" which is 1G, my PC would have a hard time running a lot of games. Setting it to 2G gives me decent frame rates but of course some of those memory go to waste when I don't need the GPU's power. So seeing these results is actually good.
Today in Hardware unboxed: "Bigger is better - Debunked"
It is not the size of the allocated framebuffer, but how to use that memory when gaming.
Wow that is some knowledge.Thank you for the effort!!!!
Thanks updated bios to UMA FIXED 64m for my 5600g. Its mostly used for hypervisor use so will appreciate the extra ram.
As it has 32gig ram it auto sized to a bit 2 gig.
[RAM SPEED Comparison NEEDED] What about difference in performance oF 2400g with different ram speeds
jbc029 yeah but for the APU this time
Already done as jbc029 said, please watch our day one coverage.
This was very informative. when the first load of benchmarks came out and everyone was testing with 16 gb kits. i did start to wonder if 8 gb kits would comprise performance with the shared ram. turns out that isnt true.
However i think a 1200 + gtx 1050 + 8gb ram offers more bang for your buck gaming wise than a 2200g + 16gb ram
D-GPU processes 1GB data then has to move it to the CPU or CPU processes 1GB data then moves it to D-GPU in each case over the available bandwidth. APU has UMA, I-GPU processes 1GB data leaves it where it is (in system memory) and passes a pointer giving the location of the data to the CPU or vice versa. So in an APU if operating system and drivers are working properly a lot of data transfers should never need to occur. Are AMD delivering on the promise of UMA or are they forcing the APU to mimic the more traditional CPU/D-GPU setup?
You really are PCMR National treasure!.
Excellent works mate!.
you should mention the PCI-E bandwidth bottleneck and way higher latency for a card like the 550 having to use the CPU ram.
This video has nothing to do with that. It was however covered in the day one review.
I understand that this is not the main topic of the video, it's just a small comment on what you say around 2:25, because in reality the hit should be even harder for the 550 than 30GB/s once it's out of onboard ram, because of PCIE limitations, it should be more like the read/copy on Aida speed I guess, anyway, I understand that the relevance of this is very limited for this video, since the topic is the IGP, and in that case yes, the memory access over the "dedicated" amount should be just as fast as the dedicated portion, as your video proves, anyway, not trying to be annoying, good job with the video showing the fixed amount of ram on Vega IGP is not really relevant for graphics performance, which lots of people seem to believe it is.
Sorry you are completely right, I misunderstood your comment. Yes, I overlooked the obvious here, you are indeed limited by the PCIe bus before the system memory bandwidth. Thank you for pointing that out, I'll try not to make that mistake again.
Yes, that is the point I tried to make earlier in a conversation w/ you Steve.
The important thing your quoted test numbers revealed, & you seemed to have overlooked, that the alleged 8 lane gpu link on the Fabric bus apu, in fact performs like a ~24 lane link to system ram, as I would expect from a miniaturised new gen bus.
Ie., on occasions when either discrete or igp DO have to link to system RAM, as BOTH do, the apu igp is in fact at a great ADVANTAGE over discrete.
In other situations, the apu igp is not at as big disadvantage as commonly perceived.
Needless to say, this fasterFabric apu bus also applies to cpugpu links, - a further apu edge.
it's a good point, I noticed on the Digital Foundry 2200G video, he was pushing high textures in games that normally kills the performance for 2GB discrete cards just fine, basically the game was using over 2GB even if he had the IGP on the bios set to allocate that amount, and since there is no PCIE bottleneck it handles it OK with the dynamic allocation, it's a fairly interesting point, now I look at the GT 1030 and it uses a x4 link only, it's kind of something that could perhaps be interesting to explore.
I think it's basically the same problem as with CVTs on cars; a true optimized CVT will hold the engine at the best performance rpm for any given situation (economy, torque or power) so that you're always getting the most out of your engine, but almost nobody would buy a car with such a CVT because according to customers they sound boring/broken.
The actual performance is better than with any other transmission, but it's much easier to add virtual gears than convince customers that the car isn't broken and explain why it has the best acceleration at a constant engine speed.
Thank you very much, great video. If I had access to your array of hardware it would be interesting to see if the rx550 suffered worse than the vega 8 in superposition if it were forced to use say 4-5 gigs of vram on a 2 gig card. Basically determine if the vega 8 can access system memory faster than the rx550 can?
super professional as always :-)
I love reading all the comments you loved, they really left me nothing to say! xD
Hey Steve, did you try setting everything to auto and seeing what the performance was? See how good the AMD drivers are at allocating the memory for game use?
That's not what the auto setting does mate. On auto the motherboard just detects how much RAM you have and sets accordingly. So it you have 8GB's most boards will set it to somewhere between 512MB and 1GB.
You might want to talk to JayzTwoCents and GamersNexus and help them get their Ryzen APU setups working, they've been having so many issues they can't even benchmark anything.
I learned that i could use AIDA for this thanks steve great video
I think the difference between the theoretical peak of DC DDR4-3200 and the measured 30GB/s comes from the limitations of the PCI-E interface.
PCI-E 3.0 can deliver up to 16GB/s per direction, which may explain the 30GB cap.
Correct me if I'm wrong.
Its so obvious yet for some reason I never thought about it like that. Very interesting vid indeed
Thanks mate.
Very informative, thanks!!
5:36 - What would be good if you could limit the ram to 500 meg for OS/2D mode and 2 gig for 3D/game mode
The most best channel ...love you bro keep it up👍
I searched in vain for a chart, but I have seen the results for openCL on these apuS, and they are outstanding, as you would expect from such tightly integrated cpu/gpu.
Instant like, because you know it's a good video.
Too kind!
I have a guess. If someone's CPU is completely saturated and have spare RAM they might get a noticeable boost with a higher buffer setting. Maybe it's DMA so there is no CPU involvement because the GPU is handling it itself. Even on high end systems, if there are stutters do to overwhelming CPU usage, there's a very slight chance that making that buffer bigger might help.
but there are lots of problems with adobe collection.. premiere cannot enable openCL by itself, illustrator cc greyed out GPU acceleration, and photoshop always show us a notification about less of vram at launch.
This is a very helpful video, thank you