volta specs

[crasch48]

http://www.tomshardware.com/news/nvidia-volta-gv100-gpu-ai,35297.html states "The DXG-1 draws up to 3,200W inside a single chassis, so effective cooling is a must." is that possible? 1080 Ti offers just over 11.3 TFLOPS , the volta dgx-1provides a whopping 960 TFLOPs of power from 40,960 CUDA cores. almost 1k increase in processing power?

[Enderman]

This is not a consumer GPU card like the 1080ti.

[Bananasplit_00]

its funny how they draw less power than Vega lol

[DocSwag]

You realize that's like 10 GPUs in one chassis not one gpu right?

[ImNotThere]

1 hour ago, Bananasplit_00 said:

its funny how they draw less power than Vega lol

3200 watts, lets call it 10 cards

320 watts per card

vega 64 specifically maxed out hits 334 so in a way youre right!

but youre stupid to use a vega card and not undervolt

[Bananasplit_00]

24 minutes ago, ImNotThere said:

3200 watts, lets call it 10 cards

320 watts per card

vega 64 specifically maxed out hits 334 so in a way youre right!

but youre stupid to use a vega card and not undervolt

its 300W per card, says so in the post.

[Princess Luna]

25 minutes ago, ImNotThere said:

but youre stupid to use a vega card and not undervolt

Correction, you're stupid to use Vega altogether if not for specific computing loads

[ImNotThere]

51 minutes ago, Princess Cadence said:

Correction, you're stupid to use Vega altogether if not for specific computing loads

tru that fam

although i sold my 1070 and bought vega 56 just because its new lmao

[pyrojoe34]

A few clarifications:

 

- These are not consumer Volta cards, they are the GV100s (Tesla cards)

- The DGX-1 contains 8 Tesla V100 cards for a total of 40,960 CUDA cores (5120 CUDA cores per card compared to the 3584 the 1080Ti has).

- 960 Tflops (120 per V100) is referring to HALF-PRECISION (FP16, actually it's mixed precision so not even FP16, more like FP16+FP32) whereas the 11.3 Tflops of the 1080Ti is talking about single-precision (FP32, which is what most games use) and you cannot directly compare them. A single V100 can do ~14 (maybe 15 depending on the source) Tflops of FP32 so the increase is not as insane as you think.

- The 960 Tflops also counts the new tensor core tech (640 per V100) which is not available in Pascal and as far as I know will probably not be available in consumer cards or will not actually help in gaming unless devs start to code for it. IIRC the tensor cores allow for mixed precision workloads (which may be nice for gaming in the future).

- The 120Tflops of FP16 per V100 is only referring to performance in deep learning applications which the DGX-1 is specifically designed for and in no way will translate to gaming workloads which are very different.

 

The only thing we know here is that the flagship Volta cards will probably have about 5,120 CUDA cores (compared to the 3584 of the 1080Ti) which is about 40% more cores. The FP32 of the flagship Volta will probably be around 14-15 Tflops which is  24-33% faster than the 1080Ti. The V100 has 21 billion transistors which is 40% more than the P100s 15 billion (keep in mind the 1080Ti only has 12 billion). And it may be possible that future developers will be able to use mixed precision in games to accelerate certain tasks which can improve performance independently of traditional metrics like CUDA core count and clock speed.

 

What I would say are safe guesses regarding the flagship consumer Volta cards is that they will be somewhere 20-40% faster than flagship Pascal in games, they will probably have decently larger L1/L2 caches, they will likely draw between 250-350W of power, and they will have a much higher performance/watt ratio than Pascal. Other than those guesses you can't really tell anything from this information (just look at all the massive differences between GP100 and Titan Xp/1080Ti). I wouldn't even be surprised if the flagship consumer Voltas don't use HBM2 (maybe GDDR6 or even GDDR5X instead), I wouldn't be surprised if they don't include tensor cores (for mixed precision), and I wouldn't be surprised if their clocks are lower than the Pascal clocks due to energy density issues.

[pyrojoe34]

5 hours ago, DocSwag said:

You realize that's like 10 GPUs in one chassis not one gpu right?

It's 8 cards, not 10.

 

Total Specs of DGX-1 Volta:

8x GV100 GPUs (each with 84 SMs, 5120 CUDA cores, 640 Tensor Cores, 16GB 900GB/s HBM2, 1455Mhz core boost clock, 15 Tflops FP32)

2x Xeon E5-2698 v4 (20-core) CPUs (40c/80t total)

512GB 2133Mhz DDR4 (which actually seems surprisingly low to me given the rest of the specs, I'm guessing deep learning workloads keep most of the data on vRAM though and the system RAM is only a cache for vRAM, so 1+TB system RAM would not be needed.)

4x 1.92TB RAID0 SSDs (almost certainly PCIe but I can't confirm that)

6x 25GB/s NV-link (instead of PCIe) for a total of 300GB/s

800Gb/s network connection (via Quad EDR IB) + 2x 10GbE

 

Crazy badass system... and only $129,000!

[DocSwag]

3 minutes ago, pyrojoe34 said:

It's 8 cards, not 10.

 

Total Specs of DGX-1 Volta:

8x GV100 GPUs (each with 84 SMs, 5120 CUDA cores, 640 Tensor Cores, 16GB 900GB/s HBM2, 1455Mhz core boost clock, 15 Tflops FP32)

2x Xeon E5-2698 v4 (20-core) CPUs

512GB 2133Mhz DDR4

4x 1.92TB RAID0 SSDs

25GB/s NV-link (instead of PCIe)

800Gb/s network connection (via Quad EDR IB)

 

Crazy badass system... and only $129,000!

Hence why I said "like 10 cards"  

[danrey84]

7 hours ago, crasch48 said:

http://www.tomshardware.com/news/nvidia-volta-gv100-gpu-ai,35297.html states "The DXG-1 draws up to 3,200W inside a single chassis, so effective cooling is a must." is that possible? 1080 Ti offers just over 11.3 TFLOPS , the volta dgx-1provides a whopping 960 TFLOPs of power from 40,960 CUDA cores. almost 1k increase in processing power?

Your comparing Apples to Pear's