AMD "Vega 20" with 32 GB HBM2 3DMark 11 Score Surfaces

[jabbadap]

all that means is lower memory clock speed, its a fact its 4096bit though

Yep, that is given. That 3dmark 11 score run uses SystemInfo 5.1, which is from June 2017. One interesting fact on SystemInfo it has some suspiciously named dll called gpu-z.dll. So @W1zzard does it use gpu-z sdk? Either way it's obvious that old SystemInfo can't recognize that graphics card.

 

[sergionography]

If this card has 4 HBM2 stacks, this is a pretty big deal. It implies they added more memory controllers and will have double the memory bandwidth of Vega 64. Vega 64 is memory bandwidth starved so Vega20 could see a serious boost in performance.

Or a multi module chip with infinity fabric. Explains the 70% performance uplift at the same clock.

 

[efikkan]

I don't know if you guys caught this one:

Vega 20 meanwhile is a product AMD has teased and mentioned is a 7nm GPU with 32GB of HBM2 memory. The Vega 20 is aimed for machine learning / artificial intelligence workloads albeit not yet launched. So it's not too surprising these patches do confirm new deep learning GPU instructions being present for Vega 20. The deep learning intrinsics added are fdot2, sdot2, udot2, sdot4, udot4, sdot8, and udot8.

Source

 

[FordGT90Concept]

Or a multi module chip with infinity fabric. Explains the 70% performance uplift at the same clock.

That's how starved Vega is for bandwidth. Remember, Vega 64 and Fury X have the same bandwidth. Vega 64's shaders can handle about 80% more throughput per shader than Fury X thanks to NCU and higher clockspeeds but it only gets a fraction of that because Vega 64's shaders are constantly waiting on the memory subsystem to feed it more data. Evidencing this is how miners underclock and undervolt the GPU while overclocking the VRAM. Vega 20 finally gives Vega the bandwidth it has always needed. Granted, if the task at hand doesn't require lots of higher level memory accesses, it won't see much benefit from the extra stacks.



It's definitely not MCM. Vega is a big chip by itself. No body is making interposers big enough to fit two. It could be a dual GPU solution though. That's entirely plausible; however, I doubt they'd call it Vega20 if that were the case.

 

[deu]

But if this b/m is real, the card could well be running at least 70% slower than it is capable of, if you add 70% to the score, then how does it compare to the 1080Ti?

He's either trolling or new to hardware + did not read the comments prior to his own before typing . It is pretty naive to use such a number to base ones GFX' puchase the comming year on Even if you have a 1080Ti (i do too and I am waiting with excitement to see AMD hopefully pushing the performance race into gear! It would be good for everyone if AMD was performance competitive and not only mining and lower segment value for money :0, but could compete with NVIDIA's 1070+ cards on perf/$/watt.

Maybe a twin vega 10 gpu on a single pcb? Vega10 x 2=Vega 20? Makes more sense imho to have 32GB HMB for 2 GPUs and the 7nm will help to have almost double the computing power with same consumption. Which is why they will bring that mainly for the production-computing-server market as for gaming CF doesn't work properly at all times. If they did that it makes much sense as a product. I cannot see the meaning of it just as a single core gpu with that added cost for the 32GB HBM.

As many mention; GPUs are not only used for gaming but for deep learning and other simulations; there you will benifit from 32GB RAM if the software allow (which it would be stupid not to do )

 

[Captain_Tom]

Vega 20 is supposed to have a 4096-bit bus. So faster memory AND double the bus?!

Yeah a 70% IPC boost isn't crazy at all lol. Imagine if 7nm allows 2GHz core clocks too...

 

[Octopuss]

Can anyone name me a few games that actually do use more than 4GB VRAM? No matter what I did, I never noticed any. Admittedly, I don't really play modern games much, but not even GTA5 used much memory, and that's supposed to be resources-heavy I think.

 

[T4C Fantasy]

Can anyone name me a few games that actually do use more than 4GB VRAM? No matter what I did, I never noticed any. Admittedly, I don't really play modern games much, but not even GTA5 used much memory, and that's supposed to be resources-heavy I think.

Ff15

 

[InVasMani]

Well, Vega Frontier has 16GB VRAM on 2 stacks. For 32GB, they'd need 4 stacks, meaning it would actually have double the bandwidth. And HBM2 on it seems to have higher clock too compared to old Vega chips. But still, gaming wise, memory bandwidth wasn't really a limiting factor, it was more GPU's throughput on pixel level.

I think Vega's problem is the same as Ryzen's was latency is a absolute killer in time critical applications. Bandwidth much like cores is great, but is more circumstantial. They have have quite a fair amount of bandwidth or torque, but they lack horsepower effective memory speed aka clock speed on HBM2. It's like a vehicle that will haul great lots of torque, but has no speed horsepower to get there quickly. They need speed with just enough hauling power required for resolutions consumers can use and afford to do so up to 4K.

 

[jabbadap]

Can anyone name me a few games that actually do use more than 4GB VRAM? No matter what I did, I never noticed any. Admittedly, I don't really play modern games much, but not even GTA5 used much memory, and that's supposed to be resources-heavy I think.

Yes many, Like T4C Fantasy said Final Fantasy XV is one of the newest ones that uses tons of available Vram. Few others which comes to my mind are CoD WWII, Dishonored 2, Ghost Recon Wildlands. W1zzard actually takes a look of vram usage on his PC game port reviews, so do check those.

But does them really need more than that is another question. Modern game engines can use more flexibly available vram, so game might take all available video memory but does not really need it to run well. And then there's many games with high-res texture packs for additional eye candy(i.e. Middle Earth Shadow of War, where game with high-res texture pack uses whopping 8665MB of vram at 1600x900)...

 

[Xzibit]

Remember the slides from 2016.

DP is increased from 1/16 to 1/2

 

[T4C Fantasy]

Remember the slides from 2016.

DP is increased from 1/16 to 1/2

the next 7970, which btw has more dp than Vega 64 xD

 

[FordGT90Concept]

ECC HMB2 too. These are definitely not gaming chips. Think Volta.

 

[T4C Fantasy]

ECC HMB2 too. These are definitely not gaming chips. Think Volta.

remember i said there would be a 7nm vega before all the news? xD

 

[efikkan]

Vega 20 will not be a gaming chip to "save" AMD in the gaming market, it's a chip to compete with Tesla, Quadro and maybe Titan. Don't expect gaming versions anytime soon…

 

[T4C Fantasy]

Vega 20 will not be a gaming chip to "save" AMD in the gaming market, it's a chip to compete with Tesla, Quadro and maybe Titan. Don't expect gaming versions anytime soon…

They have lots of extra vega 20 IDs left also, they probably plan to have gaming chips

 

[FordGT90Concept]

Akin to Frontier Edition, sure. I can't help but shake the feeling that Navi will be the next mainstream gaming architecture for AMD. AMD might be leapfrogging architectures like NVIDIA has been doing.

 

[jabbadap]

the next 7970, which btw has more dp than Vega 64 xD

In which way? Tahiti's full fp64 speed was 1/4 of it's fp32 compute power. 1/2 is the same as Hawaii has, so the next R9 290x.

 

[T4C Fantasy]

In which way? Tahiti's full fp64 speed was 1/4 of it's fp32 compute power. 1/2 is the same as Hawaii has, so the next R9 290x.

HD 7950 0.874 TFLOPs DP @975 (Tahiti) 1792 SP
HD 7970 1.075 TFLOPs DP @1050MHz (Tahiti) 2048 SP
R9 280X 1.024 TFLOPs DP @1000MHz (Tahiti) 2048 SP
R9 290X 0.704 TFLOPs DP @1000MHz 2816 SP
Vega 64 0.859 TFLOPs DP @1677MHz 4096 SP

Tahiti is the best FP64 chip AMD has ever made

 

[FordGT90Concept]

They pulled back because there's not much that uses FP64. Still really isn't outside of specific compute workloads.

 

[T4C Fantasy]

They pulled back because there's not much that uses FP64. Still really isn't outside of specific compute workloads.

some mining takes advantage of it too.

 

[evernessince]

I really don't get the point of a 32GB card. AMD took a chunk of the die just for the HBCC, which gives them access to up to 1TB of memory. What's the point of hampering the card's gaming performance if they aren't even going to use the feature they traded it for. I swear, what the hell was Raja thinking with that.

 

[sergionography]

That's how starved Vega is for bandwidth. Remember, Vega 64 and Fury X have the same bandwidth. Vega 64's shaders can handle about 80% more throughput per shader than Fury X thanks to NCU and higher clockspeeds but it only gets a fraction of that because Vega 64's shaders are constantly waiting on the memory subsystem to feed it more data. Evidencing this is how miners underclock and undervolt the GPU while overclocking the VRAM. Vega 20 finally gives Vega the bandwidth it has always needed. Granted, if the task at hand doesn't require lots of higher level memory accesses, it won't see much benefit from the extra stacks.



It's definitely not MCM. Vega is a big chip by itself. No body is making interposers big enough to fit two. It could be a dual GPU solution though. That's entirely plausible; however, I doubt they'd call it Vega20 if that were the case.

Im not sure if your assessment is correct, because if anything; vega does well with less memory size and bandwidth, as demonstrated by the apu vegas that hold up against discrete entry level cards with much higher bandwidth.

As for the MCM aspect, raja once tweeted to a questionnaire regarding vega vs polaris, and the main thing he highlighted is vega having infinity fabric. So we know infinity fabric is an essential part of vega, so seeing it on an mcm shouldn't be too out of the question. Remember here we are talking about 7nm vega 20, not the current 14nm vega 10. Current vega 10 is a 484mm2 chip. The same thing on 7nm should easily be less that 300mm2 or perhaps even close to polaris in size

 

[Octopuss]

What exactly does "starved for bandwidth" mean? Isn't HBM's bandwidth numerous times higher than GDDR5?

 

[jabbadap]

HD 7950 0.874 TFLOPs DP @975 (Tahiti) 1792 SP
HD 7970 1.075 TFLOPs DP @1050MHz (Tahiti) 2048 SP
R9 280X 1.024 TFLOPs DP @1000MHz (Tahiti) 2048 SP
R9 290X 0.704 TFLOPs DP @1000MHz 2816 SP
Vega 64 0.859 TFLOPs DP @1677MHz 4096 SP

Tahiti is the best FP64 chip AMD has ever made

Come on T4C, you should know better than that. Amd crippled FP64 performance on R9 290x. Have a look on workstation and server Hawaii cards:
https://www.amd.com/en-us/products/graphics/workstation/firepro-3d/9100
https://www.amd.com/en-us/products/graphics/server/s9100
https://www.amd.com/en-us/products/graphics/server/s9170