Big Pascal Is Taped Out.

P

PRIME1

2[H]4U
Joined
Feb 4, 2004
Messages
3,942
http://www.tweaktown.com/news/45753/nvidias-next-gen-gpu-being-tested-16nm-up-32gb-hbm2/index.html

:cool:

NVIDIA's next-gen GPU being tested, 16nm GPU with up to 32GB of HBM2
Click to expand...

3MFAJEi.jpg
 
"How you like me now"
nvidia_huang.top.jpg
 
Last edited:
There is no way NVIDIA would produce such a giant leap forward now that they have monopoly status.

We'll have 16nm, 32 GB GPUs from them in 2025.
 
Xizer said:
There is no way NVIDIA would produce such a giant leap forward now that they have monopoly status.

We'll have 16nm, 32 GB GPUs from them in 2025.
Click to expand...

16nm is next year.. about 32gb possibly for workstation segment, hardly we will see anything above 12gb for the general consumer segment..
 
Xizer said:
There is no way NVIDIA would produce such a giant leap forward now that they have monopoly status.

We'll have 16nm, 32 GB GPUs from them in 2025.
Click to expand...

4 of those would be a 128GB video card in DX12. That means the recommended system ram would be 128GB also right?
 
Xizer said:
There is no way NVIDIA would produce such a giant leap forward now that they have monopoly status.

We'll have 16nm, 32 GB GPUs from them in 2025.
Click to expand...

And by 2084 we will be leaving messages for ourselves saying, "Get your ass to Mars".
 
Those are test samples so they are loaded with the max amount of supported RAM. They have a long way to go before that card is released to the buying public. If we are lucky that is next Fall/Winter in the best case scenario.
 
Is that the actual size of the PCB they're testing? If true, that's actually pretty exciting, if it is anything close to a production version.
 
NVLink on Geforce? Is that to replace SLI?

Is SLI really bandwidth-limited?
 
I think 16GB of VRAM is plenty for desktop graphics. In 2016 of course. There may come a day where we'll need 32GB of VRAM, but I don't think that will be anytime soon.
 
Hornet said:
I think 16GB of VRAM is plenty for desktop graphics. In 2016 of course. There may come a day where we'll need 32GB of VRAM, but I don't think that will be anytime soon.
Click to expand...

32GB sounds like a Quadro card. :)
 
Hornet said:
I think 16GB of VRAM is plenty for desktop graphics. In 2016 of course. There may come a day where we'll need 32GB of VRAM, but I don't think that will be anytime soon.
Click to expand...

*insert Bill Gates' RAM quote*
 
jwcalla said:
NVLink on Geforce? Is that to replace SLI?

Is SLI really bandwidth-limited?
Click to expand...

not bandwidth at all, but latency limited.. they need a more direct and faster communication between cards..
 
I think the link with the CPU is only with POWER processors. I don't think Intel is working w/ nvidia on an NVLink implementation.
 
jwcalla said:
I don't think Intel is working w/ nvidia on an NVLink implementation.
Click to expand...

And I don't think Intel will ever implement NVLink unless NVidia makes it open tech, which would allow both Intel and AMD to come up with an "industry standard" name that any motherboard manuf can slap on their products. AMD certainly won't license it.

If it's not going to become open tech, then getting it out into the wild is going to remain a huge challenge. That would, plausibly speaking, leave NVidia the sole option to go back to making motherboard chipsets. nForce 1000 series incoming? :p
 
What do you all think the absolute performance jump between Maxwell and Pascal will be? With the addition of HBM and finally moving down to a smaller process so the chips can fit an even more retarded amount of transistors than they already do, and the coming of DX12, I'm hoping for well over 100%. That will be the time to upgrade my OG Titans and consider moving to 4K (hopefully they have 120Hz screens then too!) ;)
 
That's Intel CPU ballpark, which GPUs are still advancing in performance way faster than. GTX 580 to Titan was close to ~80% (over 100% after overclocking), and Titan to Titan X is ~50% and will gap itself more with time.
 
Hey at least I didn't say knock off the one. ;)

I think it'll be about the same as the move from Fermi to Kepler.
 
jwcalla said:
Hey at least I didn't say knock off the one. ;)

I think it'll be about the same as the move from Fermi to Kepler.
Click to expand...

This would be my guess as well, although the improvement could be more depending on how much benefits HBM brings exactly.

As far as GP100 go, I strongly suspect the early batches will all become Teslas, and consmers chips won't come until much later. Remember Maxwell has shit DP (1/32, even worse than non-Titan Kepler's 1/24), and the Tesla line is still using Kepler. The most powerful Tesla K80 uses two GK210 chips, and has a theoretical 2.91 teraflops of peak performance assuming perfect scaling and full boost.

Well Intel will be launching Knights Landing in 2H 2015. This is one behemoth built on Intel's most advanced 14nm process, and will offer ~6 teraflops of SP performance and >3 teraflops of DP performance. So almost K80-like performance but with only a single die so no scaling issues, and likely with better thermals. This thing is a compute monster and will present strong competition for nVidia.

Given Maxwell can't DP to save itself (Titan X only has ~13% the DP performance of the original Titan), it's not hard to see why nVidia would want a 16nm Tesla out ASAP to combat Intel's Knights Landing.
 
err doing heavy mathematical computations there won't be any scaling issues :) this is why you can make supercomputers with numerous GPUs. Scaling issues tend to be relegated to rendering.

Of course programming for them you have to make sure its ready for parallel work, but even for one GPU you have to make sure of that.
 
I just took AnandTech's word for it:

Per GPU throughput is lower than on Tesla K40, so given a task that doesn’t scale well over multiple GPUs a Tesla K40 could still be faster.
Click to expand...
 
And the final sentence:

so outside of a few edge cases K80 should be faster, generally quite a bit faster.
Click to expand...

Not sure what those edge cases might be, but would it make you happier if I said "potential" scaling issues instead? lol wasn't even my main point
 
blade52x said:
What do you all think the absolute performance jump between Maxwell and Pascal will be? With the addition of HBM and finally moving down to a smaller process so the chips can fit an even more retarded amount of transistors than they already do, and the coming of DX12, I'm hoping for well over 100%. That will be the time to upgrade my OG Titans and consider moving to 4K (hopefully they have 120Hz screens then too!) ;)
Click to expand...
I'm confident in at least double the performance with the smaller manufacturing process and HBM2. We've been stuck on 28nm for a really long time. But the 10x faster quote NVIDA marketing has been throwing around is really hard to believe at this point.

jwcalla said:
Knock a zero off of that and you're in the ballpark.
Click to expand...
The jump from 780 Ti to 980 Ti was around 50-60%. <= 10% is what we expect in the CPU realm. CPU generations bring motherboard chipsets with new features, which is what people get excited for with a new one. Lower TDP is also nice to have.

DogChainX said:
Skylake-E and Pascal builds then in late 2016?

One can only dream...
Click to expand...
That's what I'm planning for right now.
 
You must log in or register to reply here.
Back
Top