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

[caswow]

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

i cant remember any gpu build that big ever.

 

[n0x1ous]

GP100 was the previous largest

 

[Crumpet]

Wow.. that "deal" of 4 Tesla Volta's works out more expensive than the house me and my wife are looking at.

 

[tamz_msc]

It seems FP64 cores has decreased to 8 from 16, but the new L0 cache makes up for it through better utilization.

Similar to Pascal GP100, the GV100 SM incorporates 64 FP32 cores and 32 FP64 cores per SM. However, the GV100 SM uses a new partitioning method to improve SM utilization and overall performance. Recall the GP100 SM is partitioned into two processing blocks, each with 32 FP32 Cores, 16 FP64 Cores, an instruction buffer, one warp scheduler, two dispatch units, and a 128 KB Register File. The GV100 SM is partitioned into four processing blocks, each with 16 FP32 Cores, 8 FP64 Cores, 16 INT32 Cores, two of the new mixed-precision Tensor Cores for deep learning matrix arithmetic, a new L0 instruction cache, one warp scheduler, one dispatch unit, and a 64 KB Register File. Note that the new L0 instruction cache is now used in each partition to provide higher efficiency than the instruction buffers used in prior NVIDIA GPUs. (See the Volta SM in Figure 5).

But they now have separate INT32:

Unlike Pascal GPUs, which could not execute FP32 and INT32 instructions simultaneously, the Volta GV100 SM includes separate FP32 and INT32 cores, allowing simultaneous execution of FP32 and INT32 operations at full throughput, while also increasing instruction issue throughput. Dependent instruction issue latency is also reduced for core FMA math operations, requiring only four clock cycles on Volta, compared to six cycles on Pascal.

 

[swilli89]

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

Negative. Two quarters delayed. GP100 shipped in March of last year and gp102 showed up 6 months later.

Looks like V104 will be Q1 2018 and V102 Q2 2018.

 

[tamz_msc]

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

Maxwell didn't have the FP64 capabilities of Kepler.

 

[swilli89]

Was anything shown in Volta that shows it will be any different than a slightly shrunk Pascal in terms of gaming?

 

[xpea]

For us gamers, the big news is that Volta has a much improved cache and independent thread scheduler.

 

[tamz_msc]

Okay so the biggest change, which was actually desired, is independent scheduling withing the warp. Would like to see how it translates in to things like Async compute.

Spoiler: Thread scheduling

 

[Glo.]

For us gamers, the big news is that Volta has a much improved cache and independent thread scheduler.

Unless Nvidia will decide to reuse GP100 chip architecture without any improvements from Volta .

 

[Head1985]

Full GV100 have 5376SP so full gaming big volta GV102 will have also 5376SP without fp64SP.Same as Gp100 vs GP102 now with pascal.

Big SKU is always 50% more than mainstream part GV104.

GV102-5376SP 600mm2?
GV104-3584SP 400-420mm2?

 

[swilli89]

Okay so the biggest change, which was actually desired, is independent scheduling withing the warp. Would like to see how it translates in to things like Async compute.

Spoiler: Thread scheduling

Do we expect GeForce cards to finally be able to use Async compute with this addition?

 

[tamz_msc]

Full GV100 have 5376SP so full gaming big volta GV102 will have also 5376SP without fp64SP.Same as Gp100 vs GP102 now with pascal.

Big SKU is always 50% more than mainstream part GV104.

GV102-5376SP 600mm2?
GV104-3584SP 400-420mm2?

If that is the case expect 900$ GTX 2080 FE.

 

[tamz_msc]

Do we expect GeForce cards to finally be able to use Async compute with this addition?

I'm hopeful, they have laid the groundwork for it or so it seems.

 

[Head1985]

If that is the case expect 900$ GTX 2080 FE.

why 900usd?
Btw GTX2070 2560sp?I think its reasonable.
GTX2080 will have 40% more sp.

 

[tviceman]

Maxwell didn't have the FP64 capabilities of Kepler.

Neat, but you specifically only talked about die size in respect to cost.

 

[Arachnotronic]

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.

NVIDIA disagrees with you.

• New Streaming Multiprocessor (SM) Architecture Optimized for Deep Learning Volta features a major new redesign of the SM processor architecture that is at the center of the GPU. The new Volta SM is 50% more energy efficient than the previous generation Pascal design, enabling major boosts in FP32 and FP64 performance in the same power envelope. New Tensor Cores designed specifically for deep learning deliver up to 12x higher peak TFLOPs for training. With independent, parallel integer and floating point datapaths, the Volta SM is also much more efficient on workloads with a mix of computation and addressing calculations. Volta’s new independent thread scheduling capability enables finer-grain synchronization and cooperation between parallel threads. Finally, a new combined L1 Data Cache and Shared Memory subsystem significantly improves performance while also simplifying programming.

 

[crisium]

why 900usd?
Btw GTX2070 2560sp?I think its reasonable.
GTX2080 will have 40% more sp.

That'd be a massive cut. Nvidia has only cut to less than 75% in laptop and OEM chips. 75% would be 2688cc.

Assuming 3584cc full chip.

 

[Head1985]

That'd be a massive cut. Nvidia has only cut to less than 75% in laptop and OEM chips. 75% would be 2688cc.

Its already 33% in GTX1070 vs 1080.Why not 40%?It will be like 5% more gap.Now its 25-30% and it will be 30-35%.
They cutdown x70 card more and more since kepler.

 

[tamz_msc]

Neat, but you specifically only talked about die size in respect to cost.

Yes, because 561mm^2 vs 601mm^2 but without FP64 and on a more mature process is obviously going to be cheaper. NVIDIA cited $2bn development costs and asked 1200$ for a 471mm^2 chip. Now they cite $3bn in development costs;one can hazard a guess - Titan X Volta would at least be 1500$ for a 600mm^2 chip.

 

[tamz_msc]

NVIDIA disagrees with you.

Finish reading that article and you'll understand why it isn't much of a redesign, only a few additional improvements compared to Maxwell>Pascal. Except the fancy tensor stuff of course.

 

[crisium]

Its already 33% in GTX1070 vs 1080.Why not 40%?It will be like 5% more gap.Now its 25-30% and it will be 30-35%.
They cutdown x70 card more and more since kepler.

You're right that they have been cutting it down more, but I think there's a reason they have never done more than a 25% cut. They want to compete with themselves by spacing out their cards, but the faster it is the more they can charge. GTX 760 was 25% shader cut, so was 1070. We only see more than that when OEM/laptop chips, often before the smaller chip is ready, where I assume they get even more margin.

But there's a first for everything.

 

[Headfoot]

800+mm2 is ridiculously impressive. Too bad that level of GPU has fully left the price bracket of the mere mortal, it would be sweet to get hands on something like that

 

[Timmah!]

Most likely: no.

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

Regardless, thats the chip i am going to replace my 1080s with in the future. Not with the lowly 1080Ti´s now 5120 CC´s ftw - thats how many cores have both my 1080s together.

 

[caswow]

800+mm2 is ridiculously impressive. Too bad that level of GPU has fully left the price bracket of the mere mortal, it would be sweet to get hands on something like that

like intel with their monolithic dies this is another company that has the money to risk it building it that big.