Nvidia Post-Volta (Ampere?) Rumor and Speculation Thread

[trinibwoy]

The kind of which you've never seen before.

Titillating.

 

[Kyyla]

Holy shit this is taking so long. I have my cash in hand here. Get something launched already.

 

[SlmDnk]

@Kyyla

Taking so long?

I also can't wait for RTX v2, but... RTX launched just 17 months ago.

 

[DegustatoR]

Holy shit this is taking so long. I have my cash in hand here. Get something launched already.

I think it's almost guaranteed that we'll see the "Big Ampere" in a couple of weeks via the "virtual GTC". As for the gaming models, well, there's no rush and these will likely come in May/June or Aug/Sep.

 

[troyan]

I dont necessarily agree with Wolfram overall but hes correct in terms of transistor counts. 5700xt is 10.3 billion. The 2070 super is 13.6 billion.

5700XT uses 28% more power for ~7% (in console ports) more performance over the RTX2070. Without the 20% clock advantage of the 7nm process Navi10 wouldnt even faster than TU106.
And Turing has the better graphics, compute, Raytracing and DL/AI architecture.

The only fair comparision is GTX1660TI and 5500XT. Both uses chips with the same amount of transistors and have the "same" feature set. The GTX1660TI delivers 30% more performance at the same power consumption. And TU116 is still on 12nm.

 

[DegustatoR]

Both uses chips with the same amount of transistors and have the "same" feature set.

Not really though. "Small Turing" doesn't have BVH cores and has a version of tensor array cut down to FP16x2 support only but the rest of Turing features are there: VRS, mesh shaders, sampler feedback.

 

[Osamar]

Not really though. "Small Turing" doesn't have BVH cores and has a version of tensor array cut down to FP16x2 support only but the rest of Turing features are there: VRS, mesh shaders, sampler feedback.

I understand troyan "same" as 5500XT and GTX1660Ti have more "similar" features than RTX ones.

 

[DavidGraham]

Not really though. "Small Turing" doesn't have BVH cores and has a version of tensor array cut down to FP16x2 support only but the rest of Turing features are there: VRS, mesh shaders, sampler feedback.

And the concurrent INT32/FP32 execution.

 

[DegustatoR]

And the concurrent INT32/FP32 execution.

Not really a feature you can use.

 

[CarstenS]

Yeah, that's more of a front-endish type of thing, enabling you to make better use of available ressources. Transistors and die-space not translating into raw TFLOPS but into achieved percentage of theoretical throughput.

 

[Deleted member 2197]

Rumor: Nvidia Ampere GA100-GPU would get 8192 cores and boost speed up to 2200 MHz

A Chinese forum called Stage1 apparently has shared some reliable info in the past, and this round he talks about a chip called 'ga100', And GA obviously would be GeForce Ampere.

Here's what he posted and calls GA100:
128SM, 8192cuda, 24 / 48GB HBM2e, boost frequency up to 2.2Ghz, double the tensor core, 300W TDP

GA100 would get far more compute performance then expected, the 7nm fabbed GA100 would indeed get 8192 cores and thus 128 CUs, meaning NVIDIA if pretty much going monolithic and doubles up on their transistor budget, which going from 12nm and 14nm to 7nm is very possible. Very interesting is the mention of a boost clock up-to 2200 MHz, which seems high especially with a 300W TDP.

All that in combination with 24, 32 or even 48 GB of HBM2e graphics memory would/could deliver 32 teraflops of performance. Obviously we're talking data center and supercomputer centric products here, but Ampere paves the way architecture wise for the consumer products as well. Apparently the number of tensor cores would double up as well.

https://www.guru3d.com/news_story/n...192_cores_and_boost_speed_up_to_2200_mhz.html

 

[DavidGraham]

Not really a feature you can use.

Yes, but I meant that it accelerates RT workloads compared to Pascal.

 

[w0lfram]

Look at what Degustator said: 2070S is carrying even more dead transistors wrt to general gaming perf than just RT cores and Tensors: The inactive 512 ALUs.

Apart from pricing and product positioning, one should compare RX 5700 XT and RTX 2080 Super, if one would want to compare how much gaming performance each vendor crams into each mm².

edit: abstracting from product level and going by the highest ALU count with the highest clock frequency offered for that count across gaming, professional and mobile, TU106 fares comparatively worse than TU104 and 102 at 17,4 TFLOPS/mm², whereas 104 and 102 are at 20,5 and 21,6 respectively.

Navi 10 OTOH is comfortably ahead at 40,4, but trailing Vega 20 with 43,2.

Best-in-class for older generations where Polaris 30 with 30,7 and GP104 with 28,9.

NO..!
You are simply trying ostracize the facts, not illustrate them. Why not compare Navi-10 DIRECTLY with Turing..? There is a reason why NVidia came out with SUPER re-spins. It is because the full TU-106 die, could not compete with rdna1

Again, the rtx2070 (full TU-106 die) is 10.6 billion transistors, 2304 Cores, 64 ROPS & 144 TMUs. Nearly equal to the navi in every way.... yet the full TU-106 losses to Navi-10 (5700xt) by 25% in many DX12/Vulkan games. With Navi-10 surpassing even the rtx2080 in some games.


And that is rdna1, with rdna2 being 50% better performance/per watt.

 

[Cyan]

It is rumoured that the new nVidia Ampere is going to be 3 more powerful than Xbox Series X.

https://www.tweaktown.com/news/7107...d-specs-3x-faster-than-xbox-series/index.html

edit: fixed typo

 

[BRiT]

Xbox One X != Xbox Series X

 

[w0lfram]

Xbox One X != Xbox Series X

Yeah, they are doing a little "mis marketing"...

Using old data, instead of new. Of just looking at generalities, the new Xbox Series X (XSX), should equal the 2080 in gaming performance.

 

[techuse]

It is rumoured that the new nVidia Ampere is going to be 3 more powerful than Xbox One X series.

https://www.tweaktown.com/news/7107...d-specs-3x-faster-than-xbox-series/index.html

This isnt even in the realm of possibility WRT any consumer GPU.

 

[CarstenS]

NO..!
You are simply trying ostracize the facts, not illustrate them.

Talking about perf/mm².

Why not compare Navi-10 DIRECTLY with Turing..?

Because you end up comparing apples with oranges. You cannot compare those two directly and objectively without choosing some metrics while ignoring others.

 

[w0lfram]

Talking about perf/mm².


Because you end up comparing apples with oranges. You cannot compare those two directly and objectively without choosing some metrics while ignoring others.

What...?
No, you can run both dies up against games... then compare. The only metric that counts is frames per second, which has already been done and navi comes out ahead by 20% or greater in some instances. If you have to, run them both at the same freq, to see how each compares. But in the end, rdna is more powerful at gaming, than the 2070... because we know 5700xt also beats the 2070 Super in some games, while the 2070 is not capable of that feat. Navi10 also beats the 2080 in some games, and (again) the 2070 will never do that.

So therefore... we already know that rdna1 has greater potential in gaming than Turing, therefore we know that rdna2 is superior to Turing at gaming.

 

[DegustatoR]

I'm very interested in seeing the "potential" of RDNA1 in games which will be made for future consoles and how it will actually compare there with Turing. Fun times ahead.