NVIDIA GeForce 20 Series (Volta) to be released later this year - GV100 announced

[tamz_msc]

Volta is confirmed to be built at TSMC 12FFC. 5120 cuda cores. 15 TFLOPS FP32, 7.5 TFLOPS FP64. 815 sq mm die. 21 billion transistors. This is a monster GPU. Heck it takes GPU die size to a new record - 33% larger than GP100. Incredible. Poor AMD. That stupid "Poor Volta" image is going to haunt AMD.

FFN not FFC.

 

[xpea]

Of course a GPU will be good at matrix multiplications, it doesn't take an Einstein to figure that out.

Easy to say NOW... but when NV invested this market 5 years ago, what AMD, Intel and the others were doing ?

 

[tamz_msc]

So let's take a guess that GV102:GV100 will be the same as GP102:GP100. What can we expect? And if we assume that the x104 follows the same ratio, what can we expect for GTX 2080 (hopefully later this year)? What would the cuda cores and clockrates look like?

That would depend on whether there is dedicated hardware just for doing those new Tensor ops.

 

[Bouowmx]

Gonna predict V104 (GeForce GTX 2080) with 3584 shaders, and V106 (GeForce GTX 2060) with 1792 shaders, respectively 2/3 and 1/3 of V100, which has 5376 shaders (5120 enabled in Tesla V100). As for frequency, looks like it is about the same as Pascal's.

More reading (including a block diagram): https://videocardz.com/69378/nvidia-announces-tesla-v100-with-5120-cuda-cores

 

[Arachnotronic]

https://devblogs.nvidia.com/parallelforall/inside-volta/

Totally new & redesigned SM. Amazing work, NVIDIA.

 

[Arachnotronic]

That would depend on whether there is dedicated hardware just for doing those new Tensor ops.

It is dedicated hardware.

 

[tamz_msc]

Easy to say NOW... but when NV invested this market 5 years ago, what AMD, Intel and the others were doing ?

The FFT algorithm was invented in 1965. People have known at least since that time that a processor needs to have parallelism to maximize throughput.

 

[lobz]

Easy to say NOW... but when NV invested this market 5 years ago, what AMD, Intel and the others were doing ?

You're kidding, right???

 

[Glo.]

https://devblogs.nvidia.com/parallelforall/inside-volta/

Totally new & redesigned SM. Amazing work, NVIDIA.

It isn't redesigned. The only change that is affecting performance of the GPU is massively increased Registry File size to avoid starvation of the cores. Each SM still has 64 cores, just like GP100 architecture. Increased also L2 cache size also increases performance.

Interesting to see.

 

[raghu78]

Gonna predict V104 (GeForce GTX 2080) with 3584 shaders, and V106 (GeForce GTX 2060) with 1792 shaders, respectively 2/3 and 1/3 of V100, which has 5376 shaders (5120 enabled in Tesla V100). As for frequency, looks like it is about the same as Pascal's.

More reading (including a block diagram): https://videocardz.com/69378/nvidia-announces-tesla-v100-with-5120-cuda-cores

yeah GV104 should easily beat GP102.

 

[xpea]

orders start now. Replace DGX-1
same price
delivery Q3
same schedule as last year Pascal version

 

[caswow]

an 815mm² gpu beating everything on the market? oh gee thats some news

Trolling and Threadcrapping are not allowed
Markfw
Anandtech Moderator

 

[William Gaatjes]

If i would take the tesla p100 as example and compare the tesla v100, and use the fp32 numbers only and compare the amount of cores... This thing runs at a bit slower clock than tesla p100 does.
It is even a bit slower in throughput.
Compare the numbers for 3840 cores and 10,6TFLOPS FP32 for p100 and 5120 cores for 15TFLOPS FP32 v100.
If this is an indication for consumer volta, than it seems pascal was already highly tuned and there will be only much to gain for volta from the new 12nm process when it comes to clockspeed or in amount of cuda cores.

 

[tamz_msc]

https://devblogs.nvidia.com/parallelforall/inside-volta/

Totally new & redesigned SM. Amazing work, NVIDIA.

So the GV102 will be 5376 CUDA cores. SM configuration is unchanged compared to GP100 - still 64; has more TPCs, 40 up from 28. I think that GV102 will still be of similar size compared to GP102.

 

[Bacon1]

$149,000 for server with 8

$69,000 for server with 4

 

[Glo.]

So the GV102 will be 5376 CUDA cores. SM configuration is unchanged compared to GP100 - still 64; has more TPCs, 40 up from 28. I think that GV102 will still be of similar size compared to GP102.

Most likely: no.

GV102 chip will be around 600mm2, with 5120 CUDA cores, and without FP64 cores.

 

[tamz_msc]

Most likely: no.

GV102 chip will be around 600mm2, with 5120 CUDA cores, and without FP64 cores.

That is what I thought initially, but you have to account for a factor of ~1.43 because of the extra TPCs. With this correction, the size turns out to around 450mm^2.

NVIDIA already sold the fist version of the 471mm^2 chip at 1200$, could you imagine what it would be priced at if it were 600mm^2?

 

[tential]

Really irritated with my employer for not paying out bonuses in time for me to buy call options last week for Nvidia. Just missed all of the q1 trades I wanted. The usual. Short amd buy Nvidia.


I honestly don't know what AMD has to do with this.

Also, disable your tapatalk signature.

AT Moderator ElFenix

 

[Glo.]

That is what I thought initially, but you have to account for a factor of ~1.43 because of the extra TPCs. With this correction, the size turns out to around 450mm^2.

NVIDIA already sold the fist version of the 471mm^2 chip at 1200$, could you imagine what it would be priced at if it were 600mm^2?

GP100 - 600 mm2
GP102 - 476 mm2.

There are only 4 HBM2 memory controllers in the GPU. GP102 will have 12 GDDR6 memory controllers.

 

[tamz_msc]

36% bigger than GP100, 43% more cores, 50% more performance in part due to the new instructions. Apart from the new Tensor op, not as big a change as some of us were expecting, except in size of course.

Maybe the GV100 is a throwback to G80, given that it has been 10 years.

 

[tamz_msc]

GP100 - 600 mm2
GP102 - 476 mm2.

There are only 4 HBM2 memory controllers in the GPU. GP102 will have 12 GDDR6 memory controllers.

There are 28 TPCs in GP100 compared to 40 in GV100. I'd be surprised if they're going to sell a 600mm^2 consumer chip.

 

[tviceman]

NVIDIA already sold the fist version of the 471mm^2 chip at 1200$, could you imagine what it would be priced at if it were 600mm^2?

Kepler Titan sold for $1,000, was on the same node, smaller, and half the vram than Maxwell 980 TI which sold for $650.

Anyways, Nvidia is pricing themselves at what the market will bare. Obviously it's working well, as their smaller chips are still priced competitively and outselling the competition 2.5:1.

 

[tviceman]

an 815mm² gpu beating everything on the market? oh gee thats some news

Go crap somewhere else.

 

[caswow]

Go crap somewhere else.

how about you tell that to people crapping on amd in a volta thread. doesnt matter to you right? there will be infactions either way i dont give a crap.

Trolling and Threadcrapping are not allowed
Markfw
Anandtech Moderator

 

[badb0y]

This is amazing... Has their ever been a 815mm^2 GPU from nVidia before?