NVIDIA Accelerates Volta to May 2017?

[RejZoR]

The better analogy would be GM's V8 cylinder deactivation. Fury X, in a lot of games, runs like a V4 because the graphics pipeline isn't saturated enough to fill all of the shaders. When enabling TSAA or other async workloads, it puts most of the shaders to work like a full V8.

My understanding is that Pascal doesn't actually do async but they fixed the scheduling problem so that Pascal can rapidly change task instead of waiting for the lengthy pipeline to clear. This change allows it to get a 5% performance boost where AMD sees 10%.

There's two things going on here: async shaders and scheduling. Scheduling involves interrupting the graphics queue to inject a compute task (Pascal does this). Async involves finding idle hardware an utilizing it (Pascal doesn't do this). Both are complex and both are important.


Edit: Don't believe me? Believe Anandtech:

Fast switching is not async. Async is simultaneous processing of two or more workloads at the exact same time. Everything else is pretending to be something it is not. The "idle" shaders is nonsense. Why R9 Fury starts to fly with heavy workload, especially with TSAA and other stuff is because instead of forcing TSAA within the existing rendering thread, making it longer and choking performance, it's actually running in parallel with the usual rendering thread. R9 Fury has the hardware grunt which has been underutilized for all this time. That's the whole point of async. And delivering sufficient compute performance via shaders is exactly the same as today. You simply need to have enough of them to process something. When you run out o them, performance starts to suffer. It has nothing to do with "idle" shaders. It's just utilization as we know it now. Except it's rather inefficient in D3D11 where shaders indeed idle, that's why they had to balance things perfectly depending on game trends and what kind of workloads are expected from game engines. With async, you can basically throw more of everything into a chip and it will perform better exponentially. For as long as the code is written to utilize it.

 

[xkm1948]

I sense Volta based mid-range GPU will make this new TitanX owners cry harder than what 1070 did to last year's Titan.

 

[RejZoR]

You ca be sure that stacked DRAM won't land on lower end models. They'll still use GDDR5 on most, GTX 2070 might even recive GDDR5X, but you can forget about HBM on mid end. If NVIDIA didn't use HBM on GTX 1080, there is no way they'll use it on mid end next year...

 

[PP Mguire]

I sense Volta based mid-range GPU will make this new TitanX owners cry harder than what 1070 did to last year's Titan.

No crying here, there's still only one card higher and I've had this card over a year now. Might even wait this one out too.

 

[xkm1948]

You ca be sure that stacked DRAM won't land on lower end models. They'll still use GDDR5 on most, GTX 2070 might even recive GDDR5X, but you can forget about HBM on mid end. If NVIDIA didn't use HBM on GTX 1080, there is no way they'll use it on mid end next year...

I was talking about the shift to Async Compute. Being as good as Nvidia, once their new architecture does well in Async I will bet Nvidia will be pushing as hard as they can to max out Async usage in games and applications, which will be bad news for non Aysnc optimized architecture GPUs.

 

[rtwjunkie]

I will discuss anything and take any argument for and against any company as long as they are substatiated by facts and knowledge.

We are on the same team, then. As you point out, I have seen too many people lately who seem to join up just to bash the other side, and frankly, it's tiresome.

 

[xkm1948]

We are on the same team, then. As you point out, I have seen too many people lately who seem to join up just to bash the other side, and frankly, it's tiresome.

On a more sinister guess, it could just be some old members who were banned want to let off some steam

 

[FordGT90Concept]

Fast switching is not async. Async is simultaneous processing of two or more workloads at the exact same time. Everything else is pretending to be something it is not. The "idle" shaders is nonsense.

Improving load balancing (fast context switching) also improves asynchronous performance. Maxwell did accept async jobs but because it was statically scheduled, it created execution bubbles (idle time) which caused a loss in performance. Pascal significantly reduces those bubbles so it gains FPS instead of losing it.

 

[xkm1948]

Pascal's performance can also be attributed to the super high clock speed. I wonder how Pascal performs if given the same MHz as Maxwell or Fiji. That would be interesting.

 

[jabbadap]

Pascal's performance can also be attributed to the super high clock speed. I wonder how Pascal performs if given the same MHz as Maxwell or Fiji. That would be interesting.

I would imagine close to same cudacored maxwell, but much lower power consumption.

 

[xkm1948]

Found it.

https://www.chiphell.com/thread-1595720-1-1.html

At exact same MHz 1080 is actually slower than GTX Titan X. And indeed power draw is much lower.

So 28nm--->16nm gives Pascal better power management and higher frequency. However architecture wise there is not much improvement over Maxwell. Very interesting. Only if I have those two cards. It would be interesting to see how they perform at exact same frequency in both DOOM Vulkan API as well as Time Spy.

This will further strength my prediction that Volta will be a completely new design. And it sort of explains why Nvidia is pushing foward with Volta launch. DX12/Vulkan is coming faster than they expected, boosting frequency won't give Nvidia much edge in Async Heavy situations.

 

[FordGT90Concept]

Titan X is a bigger chip (different layout too). On top of that, dynamic scheduling does have a penalty versus static.

 

[PerfectWave]

http://www.pcper.com/reviews/Graphics-Cards/Whats-Asynchronous-Compute-3DMark-Time-Spy-Controversy

what is written here is only trash. wondering why ppl still believe maxwell can do async!

 

[PP Mguire]

Found it.

https://www.chiphell.com/thread-1595720-1-1.html

At exact same MHz 1080 is actually slower than GTX Titan X. And indeed power draw is much lower.

So 28nm--->16nm gives Pascal better power management and higher frequency. However architecture wise there is not much improvement over Maxwell. Very interesting. Only if I have those two cards. It would be interesting to see how they perform at exact same frequency in both DOOM Vulkan API as well as Time Spy.

This will further strength my prediction that Volta will be a completely new design. And it sort of explains why Nvidia is pushing foward with Volta launch. DX12/Vulkan is coming faster than they expected, boosting frequency won't give Nvidia much edge in Async Heavy situations.

Wasn't a forward, it's been planned on contract. We still might not see consumer Volta until late 2017 or early 2018. Mentioned why in one of my previous posts.

 

[FordGT90Concept]

http://www.pcper.com/reviews/Graphics-Cards/Whats-Asynchronous-Compute-3DMark-Time-Spy-Controversy

what is written here is only trash. wondering why ppl still believe maxwell can do async!

The link that links to is awesome:
http://www.futuremark.com/pressreleases/a-closer-look-at-asynchronous-compute-in-3dmark-time-spy

Note the blue column at the top of each picture. GTX 970 has no tasks on separate GPU queues where Fury and GTX 1080 do. Those tasks that are not on "3D" are being executed asynchronously.

GTX 970 (DIRECT only):

Fury (COMPUTE + DIRECT):

GTX 1080 (COMPUTE + DIRECT):

In conclusion: GTX 970 accepts multiple queues but executes them synchronously. Fury and GTX 1080 accepts multiple queues and executes them asynchronously. Pascal does have async.

 

[matar]

Great News NVidia may 2017 is not far if they do deliver on there promise.

 

[ssdpro]

NVIDIA is behaving quite aggressively if they hold a May 2017 launch. The 1070 and 1080 both handle any games we have today at 1440p and lower, many games at 4k just fine even with higher settings. I imagine the next line if on the same node would be very profitable as the cost and yields improve.

While I tend to stay squarely planted on team green, we could really use a spark in Q3/Q4 from AMD before my xx70 cards are $500. AMD stock is pretty hot right now but will need an ace with Zen to survive.
Here is to AMD Making Something Great Again soon.

 

[the54thvoid]

http://www.pcper.com/reviews/Graphics-Cards/Whats-Asynchronous-Compute-3DMark-Time-Spy-Controversy

what is written here is only trash. wondering why ppl still believe maxwell can do async!

I read that article last night. It also helpfully talks about bias (being none) in Time Spy. It also alludes to how games may be developed for DX12, with devs balancing workloads between vendors.

 

[thematrix606]

I read that article last night. It also helpfully talks about bias (being none) in Time Spy. It also alludes to how games may be developed for DX12, with devs balancing workloads between vendors.

Can you please stop? You have the debating skills of a high school teenager who knows all the wrong facts and cannot put 2 and 2 together.

I call you "paper warrior". Just look at the FPS, frame latency, and power draw. What REALLY matters.

Who cares if one side supports paper feature X? Only you paper warriors do. You never link to any graphs because you lose. You only promote hearsay. It's sad and you should stop before you grow old and realize the dumb stuff you did when you were younger. Or WORSE, you don't realize it.

 

[Frick]

Can you please stop? You have the debating skills of a high school teenager who knows all the wrong facts and cannot put 2 and 2 together.

I call you "paper warrior". Just look at the FPS, frame latency, and power draw. What REALLY matters.

Who cares if one side supports paper feature X? Only you paper warriors do. You never link to any graphs because you lose. You only promote hearsay. It's sad and you should stop before you grow old and realize the dumb stuff you did when you were younger. Or WORSE, you don't realize it.

Paper features are also feauters, and these will be important in the future, mostly because they can be used to increase performanceon consoles. Obviously it's not as sure as the rising of the sun or the workings of the weak force, but it's pretty sure. I mean it would be just daft not using those features.

 

[bug]

Can you please stop? You have the debating skills of a high school teenager who knows all the wrong facts and cannot put 2 and 2 together.

I call you "paper warrior". Just look at the FPS, frame latency, and power draw. What REALLY matters.

Who cares if one side supports paper feature X? Only you paper warriors do. You never link to any graphs because you lose. You only promote hearsay. It's sad and you should stop before you grow old and realize the dumb stuff you did when you were younger. Or WORSE, you don't realize it.

Show some proper debating skills and link to something to support your assertions.

 

[the54thvoid]

Can you please stop? You have the debating skills of a high school teenager who knows all the wrong facts and cannot put 2 and 2 together.

I call you "paper warrior". Just look at the FPS, frame latency, and power draw. What REALLY matters.

Who cares if one side supports paper feature X? Only you paper warriors do. You never link to any graphs because you lose. You only promote hearsay. It's sad and you should stop before you grow old and realize the dumb stuff you did when you were younger. Or WORSE, you don't realize it.

Eh?

Did you quote the wrong post before you strained your logic muscles insulting me?

I'll reply back if you confirm my post is what irked you.

 

[las]

I sense Volta based mid-range GPU will make this new TitanX owners cry harder than what 1070 did to last year's Titan.

Titan X released 16 months ago. 1070 is a 256 bit 8GB card with same performance when both are overclocked, or worse if people need the VRAM. You think Titan X owners care? They have had this kind of performance for a long time and is probably pretty bored by now. They are probably looking to replace it with the new Titan X coming in a few weeks. Hell, even 1080 is 25% faster than 1070. And the new Titan X should be at least 50% more powerful than 1080. Meaning it will be like 75% faster than 1070.

The 1070 is actually pretty meh if you ask me, 980 Ti easily matches it when both are overclocked, or even beats it. Here is a video of a custom 980 Ti vs a custom 1070, and 980 Ti wins..

But I think many Titan X owners will probably go Vega or wait for Volta, to get hardware async compute support. I would feel pretty stupid buying a card right now, with no async support, like 1000 series. And people that bought these cards knew Pascal lacked proper support (or they will found out real soon). Pascal is nothing but a cashgrab in my eyes. It was not even on Nvidia's roadmap before very late, probably because Volta was delayed, which explains the faster release.

 

[MxPhenom 216]

Your endless tears are endlessly amusing.

Seriously, hes been bitching about async compute on Nvidia cards for months now. Rather entertaining really.

 

[GreiverBlade]

I sense Volta based mid-range GPU will make this new TitanX owners cry harder than what 1070 did to last year's Titan.

errr actually a 1070 can make a Titan X owner cring ... (i would if i did buy a 1900ish chf card and then a 531chf card beat it ) altho ... funny enough a 8gb card hold a 12gb card in respect in all resolutions

Titan X released 16 months ago. 1070 is a 256 bit 8GB card with same performance when both are overclocked, or worse if people need the VRAM. You think Titan X owners care? They have had this kind of performance for a long time and is probably pretty bored by now. They are probably looking to replace it with the new Titan X coming in a few weeks. Hell, even 1080 is 25% faster than 1070. And the new Titan X should be at least 50% more powerful than 1080. Meaning it will be like 75% faster than 1070.

The 1070 is actually pretty meh if you ask me, 980 Ti easily matches it when both are overclocked, or even beats it. Here is a video of a custom 980 Ti vs a custom 1070, and 980 Ti wins..

But I think many Titan X owners will probably go Vega or wait for Volta, to get hardware async compute support. I would feel pretty stupid buying a card right now, with no async support, like 1000 series. And people that bought these cards knew Pascal lacked proper support (or they will found out real soon). Pascal is nothing but a cashgrab in my eyes. It was not even on Nvidia's roadmap before very late, probably because Volta was delayed, which explains the faster release.

well yep ... nonetheless the 1070 is not bad at all ... (considering the pricing of the 980Ti ... which is almost simple to double from a 1070 were i live ...) also the 980Ti match it or beat it if overclocked, yep indeed more core more ROP more TMU more bandwidth more transistor, although the 1070 equal it at stock clocks and those clocks are not far from a stock 980Ti (+300mhz core/-68mhz mem), ok once OC the 980Ti beat it, with more powerdraw tho? right? so then the 1070 has another advantage, albeit having less transistor/core/TMU/ROP etc etc etc and not needing a OC to pass above or equal, while being cheaper, a overOC 980Ti

altho that video i wouldn't call it a 980Ti win ... the difference is too small between the 2 card ( and honestly if they are on par, the one with the lesser specs should be considered as the best of the 2 )