Hot Chips 2017: A Closer Look At Nvidia's Volta

[Guest]

Nvidia was on hand to present some of the finer microarchitectural details of its new Volta GV100 SMs. We also had a meeting with Rob Ober, Tesla Chief Platform Architect at Nvidia, for a closer look at the GV100.

Hot Chips 2017: A Closer Look At Nvidia's Volta : Read more

 

[JonDol]

When would it be available ?

 

[10tacle]

There had been rumors all year that Nvidia may do a Pascal refresh (GTX 20xx series) similar to what they did with Kepler (GTX 6xx --> GTX 7xx) prior to Volta. If they were going to do it, they'd have already done it by now.

But 5,120 CUDA cores. Wow. Pascal's Titan X and 1080 Ti have 3,584. I just hope finally we'll have a true single GPU solution for 4K gaming at a solid 60FPS with the heaviest GPU hitting AAA titles (I'm looking at you, Dues Ex and Watchdogs 2).

 

[zippyzion]

I don't think NVidia needs to release Volta right now, and if they did it would be crazy expensive. They announced them quite a while ago and if they still have process issues that are causing low yields, anything they release would be a $1000+ card. Not to mention the cost and short supply of HMB2. My money is on a late Q1 or early Q2 next year launch for desktop graphics. Hopefully that gives them time to work out the kinks.

 

[dg946f_lives]

It's been released and is selling very well. Jensen said as much during the webcast on 8/10.

 

[bit_user]

When would it be available ?

It's already shipping.

https://blogs.nvidia.com/blog/2017/08/07/ai-researchers-tesla-v100-nvail-isml/

But, if you have to ask, then you can't afford it. Like the P100, this surely will not be offered to gamers. You actually can get a GP100-based graphics card, but in the form of a Quadro workstation card. Read it and weep:

https://www.cnet.com/products/nvidia-quadro-gp100-graphics-card-quadro-gp100-16-gb/prices/

They haven't announced any consumer-level Volta GPUs.

 

[TadashiTG]

@ZIPPYZION Volta as an architecture doesn't have "process issues". The V100 GPU core based on the Volta architecture has unavoidable "process issues" due to it's massive size.
A desktop Volta GPU will definitely not be as big as the V100 and will most likely not use HMB2 either. So desktop parts will not cost a lot to manufacture should Nvidia choose to start manufacturing them.

Nvidia has no reason to release any new GPU line up because of the lack of competition at the high end as well as at the perf/watt metric. A refreshed Pascal might make sense in terms of driving a bit more sales I suppose, but I'm not sure about that, I mean they already released 1060 and 1080 with faster VRAM, that could have been saved for a Pascal refresh instead.

 

[bit_user]

But 5,120 CUDA cores. Wow. Pascal's Titan X and 1080 Ti have 3,584. I just hope finally we'll have a true single GPU solution for 4K gaming at a solid 60FPS with the heaviest GPU hitting AAA titles (I'm looking at you, Dues Ex and Watchdogs 2).

You're comparing to the wrong GPU. The V100 should be compared to the P100 - not the GP102.

And the V100 is reportedly much more expensive than the P100 was. Based on that, I wouldn't expect the same ~43% increase in core count for the consumer GPUs. For gamers, the performance increase from Pascal to Volta will actually be smaller than Maxwell -> Pascal. If you already have a 1080 Ti, then you'll probably skip the 1180 Ti. But if you have a 1070, 1080, or a 980 Ti (like me) then 1180 Ti should be pretty interesting.

 

[bit_user]

BTW, 3.2 kW in 3U.

Wow.

I would not like to be around one of the air-cooled units at full bore. I wonder how much power the fans, themselves, actually burn.

 

[Martell1977]

So they went from 16nm process to a 12nm process? That's not much of a change for a GPU. I would have thought they would wait like AMD for the 7nm process and use the 16nm for a couple generations. SHould be interesting to learn exactly what was gained by changing it.

 

[d0x360]

Nvidia would be wise to release a consumer version with at least 12 gigs of HBM2. People are under the impression that amd has no answer for this but they do. The Vega was delayed so much because of issues with HBM2 yields. It's still not sorted as this article obviously states BUT it is the best way to go.

Next year and is going to launch something big that's essentially an optimized Vega that uses less power but delivers far more performance. People also forget amds architecture allows them to do more per clock cycle than nvidia.

That's nvidias mistake and they are making it again here. They keep chasing higher clocks to brute Force rendering. It's managed to beat amd because their architecture wasn't up to the task but that has changed

nVidia needs to be ready and it's likely they are already building consumer versions and running tests then making changes. Unfortunately for them they can't change the way it renders as far as the full path goes. Their architecture is like Intel and amd has once again found a way to be the best. They just need a few more months to refine what is already an excellent design for both the CPU and GPU market.

Brute Force only gets you so far. Finesse is not in their playbook at all but it's the basis for everything and is setting up.

 

[sirnate3000]

Nvidia will wait to release Volta as they are still refining yields and their is no need to rush it, Pascal is still sitting on top. Pascal will be refreshed/updated for 1100 series.

 

[thelazy_1]

Nvidia will sell its best Volta cards to Big Business millions of them for artificial intelligence autonomous driving deep learning and crypto Mining and give consumers leftover scraps that are half as Volta pascal Hybrids that barely beat today's generation of CPUs

 

[bit_user]

So they went from 16nm process to a 12nm process? That's not much of a change for a GPU. I would have thought they would wait like AMD for the 7nm process and use the 16nm for a couple generations. SHould be interesting to learn exactly what was gained by changing it.

They're not only competing with AMD. This GPU seems aimed squarely at machine learning, meaning they also have to worry about things like Google's TPU gen2 and Intel's Knights Mill. Not to mention half a dozen startups trying to build purpose-built machine learning ASICs.

So, Nvidia saw an opportunity to strengthen their product offerings and went for it. We should expect nothing less. Plus, the substantial performance improvements mean they can sell this for a lot more money than P100.

 

[bit_user]

People also forget amds architecture allows them to do more per clock cycle than nvidia.

That's nvidias mistake and they are making it again here. They keep chasing higher clocks to brute Force rendering. It's managed to beat amd because their architecture wasn't up to the task but that has changed

... Their architecture is like Intel and amd has once again found a way to be the best. They just need a few more months to refine what is already an excellent design for both the CPU and GPU market.

Brute Force only gets you so far. Finesse is not in their playbook at all but it's the basis for everything and is setting up.

I can't effectively disagree with you, because I simply lack the raw data. But this makes me all the more curious exactly what you're basing this on, and where you got such information.

BTW, I don't know if you saw that they doubled the number of warp schedulers, improved the granularity of thread scheduling, improved L1 cache bandwidth, and claim >= 50% more efficient SM's than GP100. And yet, I saw no mention of clock speeds (Wikipedia claims it's clocked at 1.37 GHz, which is just 5% faster than the GP100). This doesn't exactly line up with what you're saying about brute forcing clock speeds...

 

[bit_user]

Pascal will be refreshed/updated for 1100 series.

Someone please explain to me why Nvidia wouldn't just use their new Volta cores. Is it because Volta is too closely tied to HBM2 or the 12nm FFN node? Given that no one is forcing their hand, in the consumer space, why wouldn't they take their time and put together an offering that provides a meaningful improvement over the 10-series?

 

[zippyzion]

I'm not sure how Vega fares against Volta, but I agree with d0x360 that Vega is clock for clock doing more work than Pascal. Considering the level of performance for clock speed this seems to be the case. Performance of RX Vega 64 is on par with the 1080 despite a 360 MHz base clock and 187 MHz boost clock disadvantage. It is much the same story on the Vega 56 side where it is at a 350 MHz base clock and 212 MHz boost clock disadvantage. (All are reference clocks of course) I haven't even mentioned the lower memory clock speeds. All in all, Vega seems to do more work per cycle than Pascal.

 

[mtl96]

Nvidia has no need to release consumer Volta atm. Any custom cooled factory oc pascal beats Vega quite easily with less power. Nvidia owns the pro market and the big dollar corporate customers are buying every Volta they can make.

 

[JakeWearingKhakis]

I remember reading somewhere that gaming Volta will use GDDR6 or an improved generation on GDDR5X.

Will be a monster for sure no matter what. Hopefully AMD is ready for a response.

 

[kabat818]

Volta looks exciting.

why are people talking about vega in the comments?
-uses too much power
-can't even beat 1080
-unstable pricing
-its garbage.

 

[Martell1977]

Volta looks exciting.

why are people talking about vega in the comments?
-uses too much power
-can't even beat 1080
-unstable pricing
-its garbage.

-If that is a concern for you, for many, it's not
-beats the 1080 pretty handily, reference to reference. We don't know what the factory OC editions will do
-retailers are trying to gouge the mining crowd right from the start
-in your opinion

 

[Rob1C]

The Article says:
>"The package pulls an average of 300W at a little below 1V, so over 300A flows into the die.".

Is there any chance you meant 300W, 48V, 6.25A ?

 

[bit_user]

The Article says:
>"The package pulls an average of 300W at a little below 1V, so over 300A flows into the die.".

Is there any chance you meant 300W, 48V, 6.25A ?

No, they mean the chip's package - not the 3U chassis. The 8-GPU chassis is drawing up to 3.2 kW (need to power the Xeons, system RAM, peripherals, etc.).

 

[bit_user]

Okay, so if we're just talking about gaming performance of Pascal vs. Vega (neither of which this article is about)...

I agree with d0x360 that Vega is clock for clock doing more work than Pascal. Considering the level of performance for clock speed this seems to be the case. Performance of RX Vega 64 is on par with the 1080 despite a 360 MHz base clock and 187 MHz boost clock disadvantage. It is much the same story on the Vega 56 side where it is at a 350 MHz base clock and 212 MHz boost clock disadvantage.

Vega 64 has 4096 shader pipelines vs. GP104's 2560, giving Vega a 60% more units. Even accounting for the clock speed discrepancy, Vega 64 is doing less useful work with more raw horsepower (not to mention more "wall" power).

I haven't even mentioned the lower memory clock speeds.

HBM2 runs at lower clockspeeds but has a much wider interface. The memory bandwidth of Vega 64 is 484 GB/sec vs. the GTX 1080's 320 or 352, depending on whether it's using 10 or 11 GHz GDDR5x. Either way, Vega is doing less with more.

 

[pasaf]

How its in Crysis?