Unlimited VRAM?! | Vega 64 & HBCC in 2023
HTML-код
- Опубликовано: 24 июл 2024
- Benchmarking the AMD Radeon RX Vega 64 8GB and High Bandwidth Cache Controller (HBCC) in 15 games using an average gaming PC in 2023.
Support Iceberg Tech: / icebergtech
With VRAM becoming a point of contention around graphics cards these days, AMD came up with a solution for those who felt like 8GB wouldn’t be enough in the future. In fact, they came up with this solution about 6 years ago; High Bandwidth Cache Controller, or HBCC, was a feature of the RX Vega graphics cards when they launched in 2017. This enabled the use of system RAM instead of VRAM, meaning that if you had enough spare RAM in your PC, you could dedicate as much as you liked to your GPU.
HBCC was mostly ignored at the time, as 8GB cards like the Vega 64 weren’t exactly being overly stressed by games of the day, but things have changed in 2023. Does HBCC work as advertised? Does it help the ageing GTX 1080 competitor run today’s memory-hungry games? And why wasn’t it more popular?
Hardware Lab's look at HBCC: • AMD 8GB VRAM is better...
Music by Backing Track backingtrack.gg/
Track names: “Future Tomorrow”, “Simon’s Call”, “Vercetti Forever”
00:00 AMD Radeon RX Vega 64 8GB vs. 2023
00:36 Background: History of the AMD RX Vega 64 8GB
01:30 High Bandwidth Cache Controller (HBCC)
03:21 Test System: Ryzen 5 5600X, 32GB DDR4-3600
03:45 Gaming Benchmarks
03:48 Benchmarks: The Last of Us Part 1
05:08 Benchmarks: Resident Evil 4 Remake
06:18 Benchmarks: Jedi: Survivor
07:07 Benchmarks: Ratchet & Clank: Rift Apart
07:46 Benchmarks: Forza Horizon 5
08:41 Benchmarks: Halo Infinite
09:24 Benchmarks: A Plague Tale: Requiem
10:15 Benchmarks: Spider-Man Remastered
11:39 Benchmarks: God of War
11:28 Benchmarks: Uncharted 4
13:18 Benchmarks: Cyberpunk 2077
14:40 Benchmarks: Witcher III Remastered
15:32 Benchmarks: Starfield
17:06 Benchmarks: Fortnite
18:20 Benchmarks: Warzone 2.0
19:24 Conclusion: Attempting to understand the results
A D S
The following links are affiliated. Everything costs the same to you, but I get a commission. Please only click these links if you want to help me out!
G E A R
Audio:
Behringer UM2 amzn.to/3aV17wq
Behringer XM8500 amzn.to/3nCt2WY
Generic microphone arm amzn.to/339yZRC
Video:
Fujifilm X-T3 amzn.to/3nGapRY
Camera & 18-55mm lens kit: • Trading a PS4 for a PC...
Fujifilm 23mm f2 amzn.to/3udXvgn
FeiyuTech G6 Max Gimbal amzn.to/3nHndYs
Interfit LM8 Video LED amzn.to/2SkESJL
Neewer 120cm Softbox amzn.to/2RjlMDm
Raleno Rechargeable Bicolour Video LED Panel amzn.to/3f2l7hV
C O N T A C T
Twitter/Insta @IcebergTop10s
Email icebergtop10s@gmail.com
#AMD #Radeon #scalperpandemic 
Наука
ERRATA:
05:56 Oops, the right hand text block should read 8GB, not 2GB.
😮
Ltt vibes rn
Xd
@@ctsjson498laughing out loud over here BTW. 😊
How do you always manage to make these small mistakes? 😂😂😂 you also wrote 144 0p at 8:18 on the right. I mean that's no issue. I love your videos. I don't even know how I got to your channel, but you are hella underrated. I first thought you were a very big youtuber, but damn, no and you even have some better quality than some bigger channels. Your work is awesome, man! Keep going.
It is important to note, that some games like Halo and Hogwarts Legacy when under VRAM pressure will swap out some textures from VRAM and produce worse image than intended. Enabling HBCC might lead this games to not swapping textures and keeping image quality but at cost of performance.
So when testing such features it is important to check that image quality is really equal.
This was the main test I was looking for.
Good 'ole "texture streaming issues"
that idSoftware's RAGE had, badly.
As I understand HBCC it's about increasing the available VRAM which one would expect to improve texture quality. I might have expected it to improve 1% lows but I would not expect maximum frame rate to improve.
It's a shame it's no good in Warzone because that's one game that really needs more VRAM.
@@labrat810I remember the foliage looking terrible on the PS3
I would love to see a Blender benchamark, where it needs 8+GB of VRAM
Ah yes amd gou in blender very nice but tbh I'd say it'd do goodish
It does not work
@@yves1926 what does not work in blender amd? It does just way slower
@@Bacongoober HBCC does not work with blender
@@yves1926 that's...
Disappointing
Suspect HBCC mainly helps in cases where the game does a bad job at vram management and allocates vram that it doesn't use much which potentially lets the caching management take up some of the slack better than a normal setup.
I agree yeah
honestly.... vega was designed as an all in one heavy compute card
essentially its a workstation or gaming card --- HBCC most likely helps
with large data sets where they need the memory, some examples
would be comparing data that involves weather patterns i believe
@@wnxdafriz Probably, the card isn't really fast enough to make use of more than 8 GB of VRAM in most games anyhow (not to mention the 16gb radeon VII), and few games at the time were even able to hit 8gb vram usage so it makes more sense for productivity use. Even workstation cards don't come with HBM these days though so I guess it wasn't worth it enough in the end over just putting a bunch of GDDR on them instead and reserving HBM for datacenter stuff.
isn't Resizable BAR one of the aspects of HBCC? The anti-latency seems to be the case.
Starfield
i do wonder if HBCC + fast DDR5 platform would help modern GDDR6 cards with heavy/large/memory intensive stable diffusion workloads...
EDIT: This answer was stupid :)
Thats what i see this being used for. SDXL on my 8gb card needs to swap models with ram 3 times to finish a single image
The VEGA would allow us to test this theory even if it's not the card we'd really want with this technology.
Wondering the same to see how VRAM intensive game/work loads work on my DDR5 7200mhz 13700k system.
I snagged what must have been one of the final V64's sold from UK retailers back in November 2020 for just £219. Had been looking to get a 5600XT that day but saw the Sapphire V64 in stock for considerably less, and just jumped on it. Amazing card still going strong in my daughters PC!
Unfortunately, all she plays is Roblox and Stardew Valley.
I’d that’s the case it’s going to last for the next century
The 5600XT is also a banger of a card. It was practically just a 5700 with 6GB of VRAM and slightly lower clocks, being sold for the price of a 1660 Super. Incredible value at the time, and it's aged so well it makes me glad that I got it instead of the 2060 when I built my PC.
@@ThinkAboutVic its aged the same as the 2060 but the price did make it a better deal if you werent really bothered by DLSS
Dude Thats cute
@@ThinkAboutVic I did get one in the end! A Merc 519 5600XT, with the BIOS update etc. Bought it the following year in fact once Covid prices started to return to normal. Performance was pretty similar to the V64 in most cases. Very nice card. Only recently retired for a 6750XT.
I appreciate that you keep coming up with incredibly unique videos.
Makes the tech dystopia we are suffering through much more bearable!
Would definitely wanna see this with DDR5. Seems interesting.
I doubt it would help much if at all, data is still going trough a PCIe 3.0 x16 interface (16GB/s) which is slow compared to the bandwidth of even dual channel DDR4-1600 at 25.6GB/s, DDR4-1600 is the bottom spec for DDR4, I have never seen it outside of the JEDEC spec, nowdays the most common you will find is DDR4-2666 JEDEC and 3200-3600 XMP
If HBCC really depends more on latency than bandwidth, as said in the video, the results would be worse, as DDR5 has a higher latency than DDR4
@@portalwalker_ DDR5 have more clock cycles in the latency but it clocks at twice the speed. So no, it's not higher latency.
i gave it 20gb of DDR5 5600 mhz Hynix chip memory and i had roughly the same improvements or none at all
I believe HBCC is some actual silicon, not just something to simply be "enabled" on new cards. It's a feature from the fact Vega was never meant to be a gaming card. Super cool. I JUST upgraded my Vega 56 a month ago. The big thing, it let me render Blender scenes normally way outside 8gb of vram
It is a software feature. Its only available for HBM GPUs. This was explained at the beginning.
What was cool is that if I remember correctly, there is a Pro version of Fury X that has a build in M.2 SSD slot that it uses as aditional VRAM. Meaning that the card technically had like 1TB of aditional, low speed VRAM
This is basically the cheaper version of the radeon VII
I love my radeon vii. An amazing gpu.
@@bryanwages3518I'm so happy to meet someone who has this card, it's a white whale out there I swear.
The Radeon VII B-roll I recorded for this video might give you a clue about an upcoming video... 😉
I need radeon VII
One of my life's objectives
I feel like the radeon VII is the most underrated amd gpu
Well, not everyone has multiple high-frequency and low-latency memory kits galore. That HBCC feature is interesting. AI seems to be where HBM is going to be used. Vega is aging well though and better than the Nvidia equivalents at 1440p.
1080Ti has entered the chat.
Me and my brother have Identical PCs same CPU 3700x but he's got a rtx 2060 6g I have a vega 64 8g we both play around with stable diffusion and he can beat me any day ATM
At this point in gaming, the Vega 64 GPU itself is going to be the most limiting issue no matter how much RAM it is fed. But it is impressive to see how well it has done after 6 years and so much change in the gaming landscape.
Noticed the B-roll of the Radeon VII
Very cool! Looking forward to that!
Gold content!!!!
I want more!
Definitly subbed and liked!
This channel needs to grow!
I think HBCC did not catch on because it's too expensive. But I think, in the future, this technology or some similar flavour of it, will be come de facto and most manufacturers will include it in their hardware.
This tech have it's own issue. Even in AMD's case HBCC is disabled by default because there are certain conditioned need to be met before you can properly use it. HBCC was created because to put more memory can be very expensive with HBM. if memory are cheap you can just put the amount you want and did not need to go with tricky workaround like HBCC
I still use a vega 64, 6+ years and it still churns out decent frames and makes audio manipulation really fun.
Audio manipulation? What is this feature?
I have a Radeon VII (otherwise known as Vega 2nd Generation) and it still kicks ass.
Love it a lot and don't have any regrets. Even without overclocking or anything, having the 16GB HBM2 memory has ensured that memory bottlenecks aren't really an issue for me (at 1080p). Having this option in HBCC to further expand it into the future makes me really happy. I haven't felt a need for it yet.
So nice to finally find somebody who realized the Radeon VII is technically a "Vega Two" and not a roman numeral "Seven".
@@KomradeMikhail It's a clever name 😂 think it does often go over heads, though.
People are so obsessed with HBM memory it’s not funny. That GPU will not even take advantage of 16gbs of its vram because it’s not powerful enough to.
@@ZackSNetwork lol? I've already seen the utility in having high memory capacity in both games and applications 🙄
There's nothing stopping one from using the memory up to its capacity and I'm struggling to understand why you believe that.
Edit: do you mean it can't take advantage of the speed? The bandwidth itself? I'm still struggling to figure this comment out 😭
thanks for these enjoyable videos. theyre also very informative and interestingly presented
19:11 - IDK, MAN... = Best benchmark results ever! 😁
I think it would be really interesting to see how it runs on DDR5
I know nobody would use it with a PC like that. But it just sounds like something that could be interesting to see
it would be interesting to see how this tech does with AI workloads, since AI usually likes HBM memory... and a lot of it.
Vega Frontier Edition, Vega II and Titan V all have 16 or more GB of HBM2.
Google the dude who got Stable Diffusion running on an AMD Mi25
I'm gonna try loading a full 13B llama model to gpu with HBCC. Sadly I only have 16gigs of RAM so I can't do a 30B
This is actually interesting for other applications, rather than gaming.
If you are trying to run a LLM locally (a model like chat-gpt), then you would run out of VRAM pretty quickly on bigger models.
If you could expand the vram using the regular ram, it would allow you to run bigger models on a cheaper card... Issue is, the support for ROCm (cuda equivalent for AMD) is pretty bad right now
Just tried loading all 41 layers of a llama2 model, instead of the 30 that fit in 8gb with HBCC. Sadly at least with my miniscule 16gigs of ram it slowed down the performance.
I'm upgrading the system soon and will get 32 gigs, maybe I'll give it another go
Had the v64 nitro loved it. Over clock, under volt got 15% more perf it was nuts. Also a space heater. Ended up selling during covid 3 or 4 years later for more than price. Really an upside down experience, loved tweaking with it . Got Witcher 4k about 40 to 50 fps in prior version coming from 1080p res and was blown away at the time
There was a macbook pro 15" 2018 with an HBM version of the RX 5700 xt iirc. Could be a fun one to look at
Fascinating! Yeah would be interesting in the future to see how faster DDR5 works :)
Interesting video. How does HBCC differ from resizable bar?
I like this vid so much. I never heard of this hb... Thing. Your commentary is good. When you explaind that it is, I new faster ram would be a good benefit. Waiting for the ddr5 vid of this one. Just for curiousity.
I feel like this is another great case of Fine wine. Once DDR5 advances into higher bandwidths, these cards with HBCC could suddenly be much more playable and last longer than anyone ever expected
Im curious how this will handle v-ram intensive production applications, stuff like Stable Diffusion for example. You can run SD on AMD now through DirectML, and I have thought of trying that on my Vega 56. Might be a good motivator.
Never tried running Stable Diffusion locally tbh. What kind of workloads would cause it to use a large amount of VRAM? Just, like, regular images? Or would it have to be particularly high resolution ones?
@@IcebergTech Stable Diffusion in general is a very V-ram intensive application, but generating larger images and generating more images per batch (multiple at a time) require more vram. I have found that 8GB cards like the 1070 and the 1080 can generate images from anywhere between 512 x 512 and 1024 x 1024 (though usually closer to the lower end) but V-ram usage goes up exponentially as resolution increases.
I don't know if Stable Diffusion on AMD is any more or less V-ram efficient through the DirectML backend. I feel like Vega can handle it, but I did try running SD on an RX 580 8GB which resulted in massive temperature spikes on the core and forced a system shutdown. It could be a 580 specific issue, or it could be a "GCN runs hot and hungry" thing. I unfortunately don't have any Navi stuff to test that.
If you're interested in getting SD running locally there are two ways to do it:
1: Using the DirectML backend. This is the only way you can run it on Windows right now.
2: Using the ROCm backend. I haven't tried this, but it requires that you use Linux or WSL. I have heard that Arch based distros (like Arch, Manjaro, and Garuda) are the easiest to set this up on.
I recall that in Far Cry 5, HBCC really helped with smoothness with the HD texture pack. It was a stuttery mess without it, but totally smooth when i turned it on.
And the game really used all the allocated VRAM i gave it, went all the way to 16GB used after like 15 minutes of playing.
HBCC might be a feature slightly more suitable for systems with budget intel 12th gen cpus or newer with ddr5
worth mentioning is also the fine-vine technology that applied here quite niecely. If you would retry the tests, the vega56 is now pretty close to 1080 in most of the titles.
Would love a retest with a Radeon VII and tuned DDR5(the fastest you can get) with a 7800X3D, kind of interested to see the limits of HBCC
But a Radeon VII has 16GB VRAM so almost no game will be able to use that much VRAM hence HBCC will basically be useless
@@AliShaikh_744 Ah shit good point. Fuck why do they put this feature on the cards that dont need it? Like imagine this on the Nvidia cards of last gen that would have atleasted helped.
@@0SynergyRadeon VII is more a heavy compute card. You might get good results out of HBCC running non gaming workloads on the card.
For simple gaming the GPU is kind of bad value for money.
Maybe but it would still be limited by the theoretical max 16GB/s speed of PCIE 3.0 while one channel of DDR4 3200 has a bandwidth of 25.6 GB/s so the PCIE bus may already already be a bottleneck in the current dual channel DDR4 3600 system and higher bandwidth ram might not help a lot more but idk. (Idk what the actual bandwidth is in practice though)
The 7800x3d only needs 6000mhz DDR5 and just like the other Zen 4 CPU’s it doesn’t benefit when you go higher.
I have a bit of Frankenstein of a system. It's an old Alienware Aurora R5 I've owned since new. I ordered it during the first big GPU shortage during a cyber Monday special. I spec'd it with a 6700k 2 Rx480's 8GB and a few other good options like the cd-drive. I've pretty much replaced everything but the 6700k and it's factory AIO liquid cooler.
-i7 6700k @ 4.7GHz @1.4V
-Vega 64 Limited Edition Liquid cooled 8GB HBM2 (which I snagged off marketplace for a deal.) with the memory over clocked to bump up the bandwidth from 483.3 GB/s to 555.5 GB/s. (helps HBCC I think) cleaned it up and replaced it's thermal paste and shoehorned it into the chassis after some modifications to the case and removing the lower drive bays and trimming some plastic.
-swapped out some of the old fans for some Artic Slim P12's , 2 on the CPU AIO as a push pull and one on the GPU AIO as a push, all acting as exhaust creating a nice passive intake using the factory side vents.
-4x 16gb sticks of HyperX fury GDDR4 RAM running at 2400MHz (MOBO limited, 34/3600 MHz capable iirc.) for 64GB total physical memory.
-For storage I got rid of my old 256gb M.2 drive and 1.8TB HDD combination and replaced it with an Intel DC P4608 series PCi-E SSD.
It's an older enterprise grade data center SSD that has faster read/writes on average than any M.2 I could have gotten to work ( non NVME.)
and it has insane random read/writes paired with high endurance. It's a 6.4 TB drive but it's essentially 2 3.2TB discs on one drive.
So with that in mind I decided divvy an even page file cache from each of them to tune of 32GB each. So with my 64GB of GDDR4 plus the allocated combined 32gb+32gb I now have a useable 128GB system memory.
Once I got that sorted I moved onto the HBCC. I then allocated 32 GB of system memory to be used as VRAM cache.
When I open task manager it shows dedicated GPU memory is 32GB and shared GPU memory is 32 GB but total GPU memory is 64GB. I'm not sure where it pulls the memory from, either it uses the RAM physical memory then the RAM is supplemented by the file cache or vice versa.
Truly an alien piece of hardware now.
Oh, I'm definitely curious to see how this works with 32 or 64 GB of DDR5 RAM at 7200 MT/s with low latency (and 2 sticks only, DDR5 on 4 sticks is horrible at the moment).
Also, I think that HBCC can help properly only if the game also knows to load things sufficiently in advance (so you don't need it instantly, you're not so succeptible to latency). Or simply, like someone else noted, maybe having a lazy approach of leaving things in memory if there's space, and having this more means that you might benefit from that when those assets are reused. In any case, it does look like the game's programming (or game + driver) does matter.
I guess the ultimate test will be the (very?) few games that will have high enough textures that even 1080p might need more than 8 GB of VRAM. Or that have 1440p running well enough to be used by Vega 64, while also having very high textures. It's kind of a bit un unfortunate that the market itself, the developers, are actually taking care to only push over 8 GB of VRAM when it's actually needed, since soooo many people and cards have 8 GB, not caring for that would destroy the performance on too high of a marketshare, so to speak.
most people talking about hbcc dont test Space Engineers, it eats up gobs of ram i think i saw it at 26gb ussage at one point. im using a frontier edition though at 16 gbs vram. awsome video!!!!!
Probably just a case of lacking software support. I could see the technology doing well if implemented well. Could also depend on if there is a cpu vs gpu bottleneck.
The tech have it's issues. AMD did not solve that and just release the tech as it is. That's why HBCC being turned off by default in AMD drivers.
The recent Intel Xeon max series can use its integrated HBM2e as cache and system ram or use only hbm as system memory
Interesting, I didn’t know that. Thanks!
This would be interesting with quad channel ddr4 as well. I got quad channel ddr4 @ 3600 in my x299 pc. Sometimes I really dont know if quad channel matters at all for gaming numbers, but maybe the extra bandwidth could help here?
Nice. Curious. Could you run a LLM larger than 24 gig that the 4090 has?
modding the cooler and bios and figuring out voltages and clockspeeds with my vega 56 was the most fun I ever had with a graphics card.
Didn't find the part in the video withAfterburner setup. Have the same card and am interested how to oc the card more than I have being stable. My card does not like more than 1080Mhz on the HBM2
You can enable sam on amd drivers with a registry change. Ancient gameplay has a link to the site to get it.
great video effort , I've always went with the best latency Ram for price and mostly ATi AMD cards for various reasons. (i have 2 3060's and a 2080ti in my main computer and my other has 2 6800XT's) and my 3rd has a Vega 64) So I feel where your coming from. I like 3D rendering obviously and the gaming isn't so different from what ever box I use.
Setup any nerd with what I have and they will come off the gaming box Meth Habits programed in their orbs over their shoulders pretty fast. I won't be brainwashed any time soon again. You get a Thumbs up for this for sure, I hope you get an optimized system to explore with for sure, deeper!
A little late to the party but One other card with 16gb of hbm2 and hbbc was the forgotten vega frontier edition, also has both pro edition and adrenaline driver support. Currently running one with the x399 platform and has been a very solid card.
Not a bad show for the vega 64 im impressed with the performance specialy in starfield.
Ratchet and clank is strange, I went from 35fps to 45fps when upgrading from 16gb to 32gb ram. My settings were 1440p native with full rt on rx 6700xt and 5600x. That is a massive almost 30% increase by just upgrading ram. Gpu power draw also went up, indicating that 12gb weren't enough for those settings. That is probably why the gains were much smaller with RT disabled. More like 10-15% at max settings 4K with fsr performance. I observed the same thing in spider man, 4K fsr balanced with RT, I went from 45 to 70fps, by just upgrading RAM, crazy.
20:55 how is going to be ddr4 (low latency) bandwidth be aproblem by itself when you have the much slower pcie 3.0 or 4.0 to bottleneck latency and BW?
This feature would probably be nice for neural net training or something like that, although there’s no cuda on amd so…
If you want a good showcase for HBCC in a game check out Greedfall in the later cities. It allowed me to keep textures on Ultra and have a smooth gameplay, it was maybe the first WOW moment for HBCC back when I still had a V64. Without it the game would run out of VRAM as you kept playing and would murder the frametimes.
The Starfield clip shows pretty well how a game can run on an optimized platform. It’s just insane.
Some important note: there is a big performance bug in Cyberpunk 2077 with Vega cards that will result it idling for certain milliseconds that ISN'T affected in Linux VKD3D, some graphics feature is hard failing and causing terrible FPS.
You can see this when you run the game at 240p and the framerate stays the same as 1080p+
This. I was shocked at how good the performance was on Proton, 1080ti levels if not better when OCed. While on Windows I initially swapped my 570 for a Vega and saw virtually no difference on all versions after 1.52. I actually gathered quite a bit of info on it and sent it to CDPR, they replied that they’re investigating this. Hope they can fix it, Radeon VII running worse than a 6600 is a joke.
@@zextakamania2086 very late on this i realize, but CDPR did eventually go back and fix vega performance.
@@thanatoast4 Yeah, they replied to me and it was fixed in the next patch. Impressive that they actually handled it instead of telling to buy newer hardware.
It would be interesting to see HBCC with fast DDR5 but I understand that you don’t have the ability now. In theory it should help. Maybe with quad channel or more you’ll get a proper improvement as well.
I discovered this channel a few days ago and can’t stop watching
Great video thank you. When I had a Vega 56 flashed to Vega 64 bios I used to undervolt and be able to push the core and vram to really sweet clocks.
My card allowed settings of 1800Mhz core and 1150Mhz vram max. I ran it at 1785 core and 1100 vram. In the last of us part 1 core would run 1778Mhz vram 1100Mhz and fps it would get 68-82 FPS. I was extremely happy with that result. Settings 1080p mostly max except shadows. Power usage though 258-270 watts which is pretty high. If I locked at 60 fps vsync power usage dropped quote a bit down to 126-155 watts weird lol.
In Starfield I just wanted to say on AMD ryzen 5000 cpus turning off AMD SMT gains you quite a bit of fps at least it does on my 5900x. As for fps for Vega in Starfield I'm not sure as I now have a 7900xt card. My Vega is in its box waiting to be used by someone again hopefully soon.
do you have screenshot of voltage settings? If I under-volt i have to lover the max clocks- So I menaged to get 1550mhz core with 1v and 1100 hbm on my v56@64 rom
Thanks for your tests with Vega 64. 👍I have a water cooled Radeon VII.
Can you combine this tests with DXVK+VKD3D to use Vulkan as graphpics output? 😊
I like to patch to use "smart access memory" and the "Vulkan wrapper" on Vega graphics card.
Then test the game Starfield. 😁😁😁
I have this card and the only game i found use massively the hbbc memory extension is 7 days to die. It is happy to eat all your vram extended over the 8 Gb. My max was 14.3gb. and it helped a lot in mean of FSP.
and the radeon pro SSG i have the second generation that was veeeery short lived . it has 24gb ram and also has an raid 0 ssd 2tb setup built in :) its a freaking fast card
How I miss my Vega 64. Overclocked + undervolted with a tweaked fan curve and it shredded everything i needed it too.
you can enable SAM with a registry entry and the official drivers on older cards.
crazy that HBM is really needed for today games than it was back in the old days (-2017'esque)
14:18 I hate they can modify the game after purchase. For me the problem wasn't performance, but once I had played the game through after launch, I had obtained items that I can no longer use because the level requirement has been modified since. It's so annoying.
I got a gigabyte vega 64 about a month ago for £100 , so getting your sapphire for 117 was a great deal , thanks for the vid
@IcebergTech Awesome vid. Thanks for this.
I really want to see HBCC used in things that eat up vram like stable diffusion. That would be huge for old gpus.
I'm curious how HBCC changes performance in something like Blender. Some scenes can be complex enough that VRAM is insufficient, and the scene has to be rendered on CPU+System memory. I wonder if HBCC could allow a scene with theoretically infinite complexity to be rendered on GPU, as long as it had access to enough system memory. Stutters wouldn't be an issue either, since Blender Cycles renders in the timeframe of seconds per frame rather than Frames per Second
I'm still using my Vega 64 OC from gigabyte. I have problems in the past and fixed last year. Graphics card reset my entire rig and investigating further, i found that my actual conection to the Power suplly was bad, my fault.
If anyone have the same problem, your Vega64 shutdown your computer, try to change 1 cable 8 pins for 2 cables independent 8 pins. Obviously in a modular power suplly.
thanks for the video!!!!!!!
Can you try this on an AI task such as Stable diffusion XL whit a very large batch or Deep Floyd IF or LLama 34 B/ 70B ...etc
Ideal for watching after school
I wonder if HBCC would work better on a DDR5 platform. Do you have a system you could test that with?
It shouldn't... PCI-E bandwidth is far lower than typical DDR4 bandwidth. Latency of DDR5 is slightly worse or at best same (typically).
I recently got a Vega 56. My main game is COD Warzone and I was getting huge lag spikes. The used system I put it in was also new to me so I wasn't sure the issue. I was able to fix it by dropping the in-game VRAM usage to something low like 50% or 60%. Now it plays smooth. I thought maybe this video would show that the HBCC would help, but I have my fix anyways. Hope this helps someone out there.
Yeah I have a Vega 64, pretty good card and it really benefited from undervolting, I tried the HBCC in Forza, while it worked, I didn't get the notice of running out of video memory, the game would end up loosing so much performance over time it would end up in the teens or even single digits if left idle in freeroam. Call of Duty Modern Warfare 2 does play well on low to med sitting on the card, but that game if you got the Vram it will use that vram, enabling HBCC will also seem to make things slow over time, I didn't do much testing with it myself, I just kept it off as I didn't feel like restarting my game every 30min to an hour.
Wonder how HBCC would fare with this card paired to a modern AM5 system /w DDR5 RAM.
If it could work, Stable Diffusion people would really love this feature
I had hoped there would be some exploration of AI improvement. Seeing performance difference for 13b or SDXL but perhaps one of the AI focused folks will see this and experiment.
Ah, HBM2... I miss it. Those were intriguing days
So in theory hbcc works like infinity fabric on amds current RDNA architecture especially RDNA 3 and the high cache on the 7900xtx
What about 'reverse operation'?
Backwards HBCC, HBM as System Memory over PCIe?
I can't find any 'off-the-shelf' options but, the more I dig, the more I see the possibility.
System RAM and VRAM (seem to) already get enumerated in a shared space.
Whatever allows HBCC to get direct access to sysRAM, seems conceptually 'recyclable' for giving more (high-latency, high-bandwidth) working RAM to old-arse PCIe machines. Why? Why not?
Which BTW,
Vega(s) seem to work in PCIe Gen1 systems, just fine.
(I am #73 for Radeon WX 9100 on 3Dmark FireStrike Ultra, with dual S940 880s on NF4pro)
Alternative to the titan V?
Well pascal have you covered
Tesla P100 with 16gb of hbm2, it might not have video displays, nor "officially sourced" drivers (you might be able to get them from nvidia if you own a company, for the rest of plebs, drivers have to be "found" online)
Also 1080ti core
I wonder how the old x99 platform will handle this. I ask because tonight i plan to build my vega 56/xeon e5 2680 v4 system and enable this feature and maybe..just maybe..quad channel ram will give this a boost. Fingers crossed
I doubt it, even dual channel DDR4-1600 (yes that speed exists in the JEDEC spec) has significantly more throughput than PCIe 3 x16, but it might depending on the software and HW implementation: does the GPU use DMA or does the data need to be elaborated on the CPU? because this 2 scenarios i would expect to behave differently at varying throughput and latency
@@rubenfasola5402 there was nothing scientific about my experiment sadly, and I probably should have turned it off and on to compare, but for some games I can definitely tell the 0.1% lows must have been better. It produced a sickly smooth fps in combination with a 1080p monitor with black frame insertion.
I have 32 GB of RAM so the card will be used on my physical RAMs, right?
Just got done playing Halo infinite with Frame Gen and FSR on! Frame times 2.9ms with 180-195FPS 1080p High/ultra.
Rx6600 non xt undervolt to 1125.
Went from 75FPS to nearly 200 FPS with noticeable gains seen on my MSI 27” curved monitor 170hz. Just nasty frame rates with a great looking game too!
I wonder if the GDDR6X found in the 10GB 3080 is fast enough for a similar technology to work
you can get a Mi25 which is one of these with 16GB of VRAM for $60, you gotta find a way to cool it though lol.. i found the VBIOS easy to flash but that could cause trouble also
I have a rtx 3070 on a MSI B550-A PRO ATX AM4 Motherboard that uses this. All I had to do was turn it on in the pcie bios settings. Then I no longer had vram issues in Detroit become human on high resolutions.
Seems to me there was a reason HBCC was abandoned. I wonder how my CL16 3800Mhz RAM paired with my 3950X of 1:1 Infinity Fabric enabled would do. If faster RAM is what the HBCC requires then maybe it'd actually see some kind of performance increase.
You needed to try the benchmarks ona X platform which had quadchannel memory. Since the bandwith should be higher even tho the memory speed is lower
Always wanted to play with a Radeon 7 or Vega 64, bought a Vega 56 for my cousins build, but never got to play with it
This kind of feature would absolutely dominate at heavy AI workloads, especially for image generation. Unfortunately, I can't imagine any manufacturer hurting their high-end professional card sales like that.
These sounded so cool in theory
Vega is cool af. Still rocking one
pcie bandwidth would make a difference, too, I imagine. not just system ram speed.
I found after installing the Vega 64 a few restarts in MW2 irons itself out and behaves itself. Slide show after installation however. Have the Vega in a cheap living room gaming rig for console like sofa experience. Happily does 1080/60 so good IMO
Can you do benchmarks with FSR enabled?
I'd suggest to use high res texture mods to some games that have them to create VRAM deficit without impacting anything else
just got an upgrade in the mail and uninstalled my vega56 and searched for vega content and saw your video uploaded just a couple hours earlier; what a coincidence. I'm impressed by the relevancy of titles benchmarked. Usually I'm left scratching my head why people would choose a set of two dozen random single player titles of the last 5 years that together demonstrably have a couple thousand unique players per month. Your mix of games is perfectly relevant or interesting to me. Like how fortnite is usually a good indication of unreal engine performance.
I feel like this would show its worth much more in productivity tasks where vram is eaten up and the gpu horsepower isnt as required and nor would consistent fps either. I wonder what vrchat and skyrim with mods would be like since those usually consume ram like the gpu version of google chrome.
My 2020 PC had a 3600XT, 16 gigs of ram and a Vega 64 - incremental hardware upgrades (64 gigs of ram first) led me to use the HBCC functionality as the Vega was the last part to be upgraded. The 2 games I played most were HEAVILY modded Cities skylines (which would fill up 48 gigs of ram, 16 gigs of HBCC and 8 gigs of VRAM and still only just run at 30 FPS) and DCS which could fill up the VRAM with textures easily.
I now have a 5800X3D, 64 gigs of ram and a 7900XTX - and the only downfall of the system is PCIE 3 - the motherboard is still the same one from 2017, socket AM4 forever :) - its lasted me longer than even my Socket 7 system from back in the day.
Looks like my i5 3570K isn’t the only bottleneck anymore. Next year seems like a complete upgrade for me.
Thanks for Vega 64 benchmark in 2023😅