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

[JDG1980]

Registry File size. Titan Xp has 3840 CUDA cores, each SM in it has 128 Cores. Each SM is fed by particular Registry File size. The same Registry File size is available to 64 cores in GP100 chip. So they are less starved for resources. This is the reason why Maxwell was such huge jump in efficiency over Kepler. It was not because of Tile Based Rasterization, it does not increase performance, but efficiency(saves power required to move the data). It was because of lowered number of cores in Maxwell architecture that had access to particular pool of resources.

Funniest part: GP102, GP104 have had the same SM/Registry file size layout as Maxwell, that is why there was no difference in performance clock for clock/core for core.

This is not true. The block diagram (which you can see on this page) for GP104 clearly indicates that each SM has 64 CUDA cores, just like on GP100. The only difference with GP100 SMs is that they also have extra dedicated FP64 CUDA cores (which are needed for HPC but serve no purpose during gaming).

Reducing CUDA cores per SM from 128 to 64 had minimal impact on performance; as you note, on a per-TFlop basis, there was little difference between Maxwell and Pascal. (The performance boosts that Nvidia obtained with Pascal were largely gotten through higher clock speeds, and packing in a couple more shaders due to the die shrink.) If going from 128 to 64 CUDA cores per SM didn't do much, then it seems unlikely that the massive gains in perf/TFlop (about 33%) from Kepler to Maxwell were due to going from 192 to 128 cores per SM.

The addition of tiled rendering remains the most likely explanation for Maxwell's substantial boost in utilization (DX11 perf/TFlop). Whether any other such breakthroughs are on the horizon remains to be seen. It's safe to say that Nvidia will try to position GV104 (when it arrives) above GP102, since they have historically tried to have each new card beat the previous generation's card from one tier up. That could be accomplished with 3584 CUDA cores at about 2 GHz, if they were fed with adequate memory bandwidth (probably via GDDR6, and a 256-bit bus as always for 4-series chips). It would not necessarily require improved perf/TFlop and I'm not sure we will see additional meaningful gains on that front. Maybe 5%-10%. I'm assuming a ~400mm^2 chip, similar in size to GM204, with 4/6 as many shaders as the big chip. Presumably the new "12FFN" process offers better clock speeds, since calculating the transistor density increase from GV100 over GP100 shows only about a 4% improvement on that front.

 

[Timmah!]

This is not true. The block diagram (which you can see on this page) for GP104 clearly indicates that each SM has 64 CUDA cores, just like on GP100. The only difference with GP100 SMs is that they also have extra dedicated FP64 CUDA cores (which are needed for HPC but serve no purpose during gaming).

In that article you quoted, right under the diagram you linked, there is this sentence>

The GeForce GTX 1080 is a fully enabled implementation of GP104. This means 2560 CUDA cores split up over 20 SMs operating at a blistering boost clock of 1733MHz.

I suppose that means he was right, since 2560/20 is 128.

 

[JDG1980]

<removed>

 

[swilli89]

Sweepr makes sure to create these "topic threads" so that he can control the OP. He has clearly listed false release window information when even Ryan Smith on the AT article states "A DGX-1V will set you back a cool $149,000. The payoff? It ships in Q3, whereas OEM P100 designs won’t be available until Q4." So the very first products based on the Volta GPGPU card are shipping in late Q3 and actual designs not available until the end of this year. On the OP Sweepr is claiming Q3 release for desktop cards. Its almost laughable except for the fact that its not.

 

[Sweepr]

GTX 1080 was launched and made available to consumers a few months after GP100 was announced, with very few leaks before launch. People like swilli89 tried hard to convince everyone that that wasn't possible for dozens of reasons, including GP100-based PCIe products shipping/release dates, GDDR5X availability, among other excuses, and they were wrong. First GP104 gaming products out in June 2016, while Pascal-based Quadro cards weren't till late 2016. Now I'm well aware it's a rumor, but previous releases indicate consumer Volta in late 2017 could happen. Don't panic.

 

[Arachnotronic]

GTX 1080 was launched and made available to consumers a few months after GP100 was announced, with very few leaks before launch

Yep, gaming is such a huge part of NVIDIA's business that you can bet if they've got GV100 up, running, and running simulations that the other Volta silicon long taped out and is in post-Si validation right now.

 

[Karnak]

GTX 1080 was launched and made available to consumers a few months after GP100 was announced

Then look at 980 -> 980ti, 980ti -> 1080 and 1080 -> 1080ti.

Yeah, that's right. We're talking about 9-11 Months between each of them. I doubt consumer volta will launch in 2017.

 

[tviceman]

Yep, gaming is such a huge part of NVIDIA's business that you can bet if they've got GV100 up, running, and running simulations that the other Volta silicon long taped out and is in post-Si validation right now.

Yep. I'm guessing GV100's lead time will be a little bit longer than GP100, so I'm going with a November release for GV104.

 

[Sweepr]

Yep, gaming is such a huge part of NVIDIA's business that you can bet if they've got GV100 up, running, and running simulations that the other Volta silicon long taped out and is in post-Si validation right now.

Expecting a fast roll out like Pascal, with GV104 and GV102 first. Worst case scenario early 2018 for the first high-end Volta products, based on Hynix's recent claim:

SK Hynix has been planning to mass produce the product for a client to release high-end graphics card by early 2018 equipped with high performance GDDR6 DRAMs

5120-5376 CUDA cores GV102 Titan with 768 GBs/s bandwidth could enable decent single GPU gaming at 8K if you tone down some settings or much higher frame rates at 4K. Crazy how fast they are improving, and that's without a (radically) new node.

 

[Madpacket]

Hmm. Vega better be damn good or very well priced. Freesync is still a feather in their cap but if Nvidia keeps widening the performance gap it's going to take a miracle to catch up. At least with Ryzen selling well they should be able to allocate a good deal more money to RTG for R&D. Wonder if Navi is getting closer to tape out?

 

[swilli89]

Volta cards will no doubt be powerful seeing as we can expect GP102 to be 600mm^2 or larger or so. The GV100 scaled down 22% would be 635mm^2 for a GV102. Similar to how a GP100 scaled down 22% while keeping the amount of shaders the same resulted in 471mm^2 for the GP102.

There may be a $1299 version of GP102 that sees the light of day Q3 or Q4, but a normal priced $600 variant (true flagship card price.. especially if its using GDDR6 right?) won't drop until Spring 2018 at the earliest.

 

[Arachnotronic]

Volta cards will no doubt be powerful seeing as we can expect GP102 to be 600mm^2 or larger or so. The GV100 scaled down 22% would be 635mm^2 for a GV102. Similar to how a GP100 scaled down 22% while keeping the amount of shaders the same resulted in 471mm^2 for the GP102.

There may be a $1299 version of GP102 that sees the light of day Q3 or Q4, but a normal priced $600 variant (true flagship card price.. especially if its using GDDR6 right?) won't drop until Spring 2018 at the earliest.

I wouldn't get my hopes up for NVIDIA being that late to the game with consumer Volta if I were you. It really looks like Volta is being prepped to do battle with Vega.

 

[Karnak]

I wouldn't get my hopes up for NVIDIA being that late to the game with consumer Volta if I were you. It really looks like Volta is being prepped to do battle with Vega.

Why late? Consumer Volta early 2018 would be pretty in line with everything since GM104.

 

[Arachnotronic]

Why late? Consumer Volta early 2018 would be pretty in line with everything since GM104.

Early 2018 for GV102 based on the SK Hynix PR, but I think GV104 will come out in the fall.

 

[Karnak]

Early 2018 for GV102 based on the SK Hynix PR, but I think GV104 will come out in the fall.

980 -> 980ti / September 2014 - June 2015
980 ti -> 1080 / June 2015 - May 2016
1080 -> 1080ti / May 2016 - March 2017
1080ti -> 2080 / March 2017 - ???

That's why I'm expecting GV104 in Q1 2018, about 9-11 months after the 1080ti.

 

[xpea]

GTX 1080 was launched and made available to consumers a few months after GP100 was announced, with very few leaks before launch. People like swilli89 tried hard to convince everyone that that wasn't possible for dozens of reasons, including GP100-based PCIe products shipping/release dates, GDDR5X availability, among other excuses, and they were wrong. First GP104 gaming products out in June 2016, while Pascal-based Quadro cards weren't till late 2016. Now I'm well aware it's a rumor, but previous releases indicate consumer Volta in late 2017 could happen. Don't panic.

^^This
No way, first Volta consumer is a 2018 product

 

[swilli89]

I wouldn't get my hopes up for NVIDIA being that late to the game with consumer Volta if I were you. It really looks like Volta is being prepped to do battle with Vega.

I don't hope its delayed/late. I love when new, faster tech comes out!

Early 2018 for GV102 based on the SK Hynix PR, but I think GV104 will come out in the fall.

But what shape will GV104 take? If its IPC is not a Kepler - Maxwell jump then they are just basically refining and slotting down GP102's die size to the GV104 market for this year. If they have truly redone the architecture and there are large IPC gains then yeah its going to extremely fast.

It just sorta seems like to me that this incarnation of Volta is a tick for 16nm and won't really change their per mm2 performance.

 

[Arachnotronic]

I don't hope its delayed/late. I love when new, faster tech comes out!

Awesome, me too. My Titan Xp is good for 2560x1440 144Hz, but when that 4K 144Hz monitor comes out, this Titan Xp will be really pathetic. I need more power and I'm hoping Volta will get me close to where I need to be when that new monitor comes out.

But what shape will GV104 take? If its IPC is not a Kepler - Maxwell jump then they are just basically refining and slotting down GP102's die size to the GV104 market for this year. If they have truly redone the architecture and there are large IPC gains then yeah its going to extremely fast.

It just sorta seems like to me that this incarnation of Volta is a tick for 16nm and won't really change their per mm2 performance.

We don't know enough about the non-SM bits of the architecture to really make that call, but the SM looks like it has quite a few solid improvements. The memory controller is also apparently much more efficient, hopefully that learning will show up in the consumer based stuff, too.

This is really exciting stuff, it is good to see NVIDIA pushing as hard as it can and not getting complacent.

 

[itsmydamnation]

One thing to consider is mm per transistor hadn't moved very far with 12nm. So unless they have high clocks, bigger dies and more ipc all in one the jump in perf for consumer cards might not be as large if you expect the same price point

 

[Arachnotronic]

One thing to consider is mm per transistor hadn't moved very far with 12nm. So unless they have high clocks, bigger dies and more ipc all in one the jump in perf for consumer cards might not be as large if you expect the same price point

Perf/mm^2 calculation is muddied by inclusion of tensor cores.

 

[IllogicalGlory]

Perf/mm^2 calculation is muddied by inclusion of tensor cores.

It wasn't performance/mm^2, it was transistors/mm^2.

 

[xpea]

It just sorta seems like to me that this incarnation of Volta is a tick for 16nm and won't really change their per mm2 performance.

Correct but at same time, Volta is much much more efficient at same TDP (around 40%)

 

[xpea]

One thing to consider is mm per transistor hadn't moved very far with 12nm. So unless they have high clocks, bigger dies and more ipc all in one the jump in perf for consumer cards might not be as large if you expect the same price point

I don't think so.
Compared to GP100, GV100 has 33% bigger die for 42% more FP32 FLOPS at same 300W TDP. And I don't even count the addition of the new scheduler and the Tensor units.
Volta offers MASSIVE gain of efficiency !
It means that for same TDP, it offers much more performance. If GV100 is any indication of consumer Volta, then we can expect around 40% more theoretical performance for same die size...

 

[tamz_msc]

I don't think so.
Compared to GP100, GV100 has 33% bigger die, 42% more FP32 FLOPS for same 300W TDP. And I don't even count the addition of the new scheduler and the Tensor units.
Volta offers MASSIVE gain of efficiency !
It means that for same TDP, it offers much more performance. If GV100 is any indication of consumer Volta, then we can expect more than 35% performance for same die size...

He is clearly talking about transistor density which hasn't improved much with 12nm FFN.

 

[xpea]

He is clearly talking about transistor density which hasn't improved much with 12nm FFN.

I have edited my post.
But transistor density is not an important metric for us end-users. Performance is.