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

[Arachnotronic]

The Volta launch will look just like Pascal. It's going to lead with the very top end HPC/data center/deep learning chip. Consumer will be several months after. Nvidia is doing this because Radeon Instinct(Vega architecture) is AMD's first legit attempt at a card in that segment. AMD has already laid the tools ground work with the Boltzmann Initiative, MIOpen, and ROCm. They just needed a card to go with it. Radeon Instinct is that card. Nvidia actually has to start competing against someone other than itself. That means if they have an opportunity to launch early they will.

Has nothing to do with Radeon Instinct, and everything to do with the fact that the data center business is NVIDIA's fastest growing segment by far.

 

[Despoiler]

Has nothing to do with Radeon Instinct, and everything to do with the fact that the data center business is NVIDIA's fastest growing segment by far.

That they are trying to protect..... Several of the Vega features directly benefit HPC/data center/deep learning. They don't align directly to gaming. Ignoring that is well....being ignorant.

 

[Headfoot]

nVidia's much higher datacenter pace absolutely has something to do with the fact that they have big, fast moving competitors in that space. Intel with Xeon Phi, AMD with their GPUs, various FPGA and DSP companies chomping at the bit for specialized high throughput computing.

 

[railven]

TIL: NV's behavior from the last 3 GPU generations is going to change because AMD is finally making a strong move in datacenters.

It's not like consumers got GK104/GM104/GP104 before datacenters got GK110/GM100/GP100. Dang AMD and their market movements! It's your fault NV is milking us!

 

[crisium]

Anything Volta relevant at this conference?

 

[tential]

Has nothing to do with Radeon Instinct, and everything to do with the fact that the data center business is NVIDIA's fastest growing segment by far.

Trying to get people to understand that not every business decision has to do with the competition is hard.

 

[Bouowmx]

Right from Anandtech live blog:

 

[n0x1ous]

GV100

21 Billion Transistors
5120 Cuda Cores
800mm2
12nm TSMC

 

[xpea]

Tesla V100
815mm2, TSMC 12nm FFN, 21B transistors
15 TFLOPS FP32, 7.5 TFLOPS FP64
20MB register file, 16MB cache, 16GB of HBM2 at 900GB/sec. NVLink 2 at 300GB/sec

 

[xpea]

Vega Instinct M25 is a toy in comparison

 

[n0x1ous]

Poor Vega - Goodnight AMD

 

[Arachnotronic]

Poor Vega - Goodnight AMD

Yeah, Vega has no chance in the data center against this beast.

 

[Ajay]

815 mm^2!!! What a monster. Hard to imagine yields on this thing are anywhere near normal.

 

[tamz_msc]

Vega Instinct M25 is a toy in comparison

Still MI25 is 36% faster than GV100 in TFLOPS/mm^2.

 

[xpea]

Volta demo live simulating andromeda galaxy with hundreds of million of stars !!!

 

[xpea]

Still MI25 is 36% faster than GV100 in TFLOPS/mm^2.

Wrong. Not in deep learning, the main market of these GPUs

Edit: V100 is 120 TensorFlops for 815mm2 vs 50 Tensor Flops on Mi25 for 500mm2
a Toy I said

 

[Bacon1]

Yeah, Vega has no chance in the data center against this beast.

You do realize that Vega and this are in completely different price segments right? This is 815mm

 

[jpiniero]

Still MI25 is 36% faster than GV100 in TFLOPS/mm^2.

Yeah but the instinct doesn't have DP. Still, for that big (and on the 12 nm node) I would expect more.

Edit: nVidia is still doing the separate FP32 and FP64 cores, and now has the Tensor cores. So I guess that's why it's so big.

 

[tamz_msc]

Wrong. Not in deep learning, the main market of these GPUs

Edit: V100 is 120 TensorFlops for 815mm2 vs 50 Tensor Flops on Mi25 for 500mm2
a Toy a said

I'm talking about FP32 TFLOPS. The yields on this thing assuming a 24mm*34mm die(816mm^2) is GF100 levels at worst and early Kepler levels at best.

 

[xpea]

Yeah but the instinct doesn't have DP. Still, for that big (and on the 12 nm node) I would expect more.

V100 is 120 TFlops. how much did you expect ?
more than 10 times faster than P100 on Deep learning !!!

 

[xpea]

I'm talking about FP32 TFLOPS. The yields on this thing assuming a 24mm*34mm die(816mm^2) is GF100 levels at worst and early Kepler levels at best.

I want to say that NV doesn't care about FP32. Biggest market is deep learning FP8 performance when Volta crushes everything else in the market with the dedicated instruction.
NV invented this market, they know what scientists want. They are not AMD followers. They are again one generation ahead...

 

[Glo.]

V100 is 120 TFlops. how much did you expect ?
more than 10 times faster than P100 on Deep learning !!!

Only for Tensor.

FP16 performance is still 2x FP32.

 

[tamz_msc]

I want to say that NV doesn't care about FP32. Biggest market is deep learning FP8 performance when Volta crushes everything else in the market with the dedicated instruction.
NV invented this market, they know what scientists want. They are not AMD followers. They are again one generation ahead...

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

 

[crisium]

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?

 

[raghu78]

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.