The RTX 3080 TI isn't going to be- EDIT - No We Were All Wrong!

[g67575]

Ok, I thought dual GPU graphics cards like the GTX 690 showed up in the Windows device manager as 1 display adapter, but it seems not. I suppose this means it basically just SLI'd the two GPUs together and had similar poor support in games / programs

So, as you say - Hopper would need a complete redesign for MCM (and might not be ready by 2021). Why not simply have 2 powerful GPUs with an 'infinity fabric' type link between them?

Just saw a Nvidia research paper about MCM from 2017 thats kind of interesting. In the conclusion, the authors say "The greatest challenge towards building more powerful GPUs comes from reaching the end of transistor density scaling, combined with the inability to further grow the area of a single monolithic GPU die"

Also, the performance of an MCM GPU could be "within 10% of that of a hypothetical monolithic GPU" (that is too large to build in practice). So, you would probably still lose a bit of performance...

The paper suggests a possible configuration of 4 GPU modules + an interlink (SYS + I/0).

For a single GPU, they estimate that "~800mm2 is expected to be the maximum possible die size that can be manufactured". Interestingly, the A100 7nm Ampere GPU exceeded this, with a die size of 826 mm². I wonder if they already reached the limit for 7nm?

Link: https://research.nvidia.com/sites/default/files/publications/ISCA_2017_MCMGPU.pdf

 

[Grim5]

Ok, I thought dual GPU graphics cards like the GTX 690 showed up in the Windows device manager as 1 display adapter, but it seems not. I suppose this means it basically just SLI'd the two GPUs together and had similar poor support in games / programs

Nah they didn't lol. Same goes for AMD's attempt at dual chip GPUs at that time. That's why those designs have fallen by the way side, because developer support for multi GPU's has dropped off and the scaling was never great to begin with.

So windows always recognized these cards as two seperate GPU's and you had to enable SLI/CF to make both work, and if the game didn't support SLI/CF - it would just default to one of the GPU dies running and the other would sit idle.

The guess to make it a simple comparison, with a GTX690 (assuming the game supports SLI) it may have run in a way that the first die generates even frames and the second die generates un-even frames or even the first die generates the first pixel and the second die generates the second pixel and so forth. But with MCM, the whole card is still working to produce 1 pixel at a time as a single unit - the benefit of MCM is that you more easily scale up the number of cores etc without having such a big impact on heat and voltages and also the yields are better, so manufacturing several small processing units is much cheaper than manufacturing one very large monolithic die - e.g look how hard it is for Intel to make CPU's with high core counts, while it's trivial for AMD with Ryzen.

 

[g67575]

Getting back to the RTX 3080 TI - To make a reasonably accurate estimate of performance, I think we would need to know the exact die size and TDP. Supposedly, the die size of the GA102 is 627 mm², according to Techpowerup. That is 75.9% the size of the already released Ampere GPU. but who knows if it's true...

 

[g67575]

I think the current die size estimate (on techpowerup) may be a bit low for the RTX 3080 TI, I think a die size of 627 mm² only applies to the RTX 2080 and lower.

The die size of the RTX 2080 TI is approx. 27% larger than the RTX 2080 Super.

If you apply a similar increase to the RTX 3080 TI Die size:

627mm² + 27.5% = 799.425 mm², which is just under the 800 mm2 maximum die size for a single GPU, estimated by Nvidia in 2017.

The only doubt I have about that figure is that this is very similar to the Ampere A100 7nm die size.

The estimate of 40,000 million transistors just looks like a rough figure / placeholder.

Techpowerup RTX 3080 TI spec estimate:
https://www.techpowerup.com/gpu-specs/geforce-rtx-3080-ti.c3581

 

[Grim5]

Techpower Up's estimated 3080ti size is based on the GA100 known die size and the assumption of a linear pattern based on the leaks that GA102 supports 5300 cores.

I.e. GA100 has 6900 cores (about 8.3 cores per mm2) - multiply that by the leaks of GA102's 5300 cores = 627mm2 (they assume the 3080ti will keep the 500 tensor cores that GA100 has, so its 1 to 1 ratio)

The problem with this number is that GA100 doesn't have any RT cores, those take up extra space so indeed if the 3080ti has 5300 Cuda Cores, 500 Tensor cores and then some RT cores, it may be bigger than 627mm2 - not much bigger though probably no more than 700mm2.

 

[g67575]

The problem is, the RTX 2080 TI die is already 754 mm², with 12nm fab. process, and the RTX 3080 TI supposedly more than doubles the amount of transistors of the last gen.

RTX 2080 TIs have 18,600 million transistors, and the previous gen Telsa GPUs (v100) have 21,100 million, surprisingly close.

So, the RTX 2080 TI has 88.15% the amount of transistors of the v100.

Assuming the RTX 3080 TI has 88.15% the amount of transistors of the Ampere (Tesla) GA100, it would have 47.7 billion transistors.

If the die size is approx. proportional to transistor count (probably isn't quite), you would expect a die size of 728.1mm².

Also, 768 seems to be the magic number, when it comes to the number of extra shader cores per Nvidia generation. So I think a total of 5120 shaders is more likely.

It will be interesting to see if having ~88% the transistor count of the Ampere GA100 translates into an equivalent reduction in performance. The RTX 2080 TI is upto 13% slower than the highest spec Volta Tesla GPUs (which have more Shaders, ROPS etc).

 

[g67575]

What do you guys think? For the RTX 3080 TI, would ~88% the performance of the GA100 Ampere GPU (in games) be enough of an upgrade for you?

Or, are you more interested in the performance of the RTX 3080 or 3070?

 

[K1LaW8tt5]

more interested in the performance at the £500 mark because realistically I won't spend much more.

 

[Phixsator]

What do you guys think? For the RTX 3080 TI, would ~88% the performance of the GA100 Ampere GPU (in games) be enough of an upgrade for you?

Or, are you more interested in the performance of the RTX 3080 or 3070?

Don't care about who has the best halo product. I'm only interested in the most performance for around 400-450 quid mark.

 

[Grim5]

What do you guys think? For the RTX 3080 TI, would ~88% the performance of the GA100 Ampere GPU (in games) be enough of an upgrade for you?

Or, are you more interested in the performance of the RTX 3080 or 3070?

The 3080ti will be faster than A100 in games. Compute cards are not made for gaming.

 

[g67575]

Grim5 - Pretty sure you are wrong there. the Tesla V100S PCIe 32 GB has higher theoretical Performance than both the Titan RTX and RTX 2080 TI.

V100S Specs:
Pixel Rate - 204.4 GPixel/s
Texture Rate - 511.0 GTexel/s
16.35 TFLOPS

Link: https://www.techpowerup.com/gpu-specs/tesla-v100s-pcie-32-gb.c3467

In the link above in the Relative Performance section, they estimate that the V100S is 13% faster than the RTX 2080 TI.

The performance advantage isn't that surprising, since the RTX 2080 TI has 68.75% the amount of ROPs and 88.5% the amount Shading Units + TMUs of the V100S.

 

[Grim5]

I don't think I've ever seen a Quadro card beat a GeForce card in games

 

[g67575]

The Turing Quadro has a lower spec core config than the top Turing Tesla GPU (e.g Shading Units
TMUs, ROPs)


https://www.techpowerup.com/gpu-specs/quadro-rtx-8000.c3306
https://www.techpowerup.com/gpu-specs/tesla-v100s-pcie-32-gb.c3467

However, both cards have superior specs and theoretical performance compared to the RTX 2080 TI.

 

[Guest2]

However, both cards have superior specs and theoretical performance compared to the RTX 2080 TI.

Key word here 'theoretical' Why not just benchmark the V100S using somethink like the Quake 2 RTX demo, firestrike or anything

 

[g67575]

Well, there's no reason for the 3080Ti to have more transistors than the Ampere A100. The A100 is already maxing out the die size limit, which is approx. 800mm squared, with 54.2 billion transistors, already more than double last gen Turing GPUs. Also in previous generations, Tesla GPUs always get the most transistors.

This GPU is the fastest single GPU, according to recent benchmarks, with 446 points vs the Titan V's 401.
Link: https://www.tomshardware.com/uk/new...s-fastest-gpu-crown-in-first-benchmark-result

With 40-47.7 billion transistors, and therefore fewer shader units, I can't see how it could possibly be faster than the A100. Perhaps some of the factory overclocked 3080Tis with +20% GPU clocks might come close.

I've realised that it could have upto 5888 shaders (5120 + 768), but still be slower than the A100.

Why the 3080Ti has been estimated to be faster than the Tesla A100 is beyond me, apart from blatant Nvidia shilling I guess. Although to be fair, many ppl in the rumour mill have guessed at the Ti's performance without comparing it to the Ampere Tesla GPU.

 

[craptakular]

Will it run crysis though?

 

[justwondering]

Will it run crysis though?

More importantly, will it run Crysis remastered?

 

[nomadd]

More importantly, will it run Crysis remastered?

Yes - via Yuzu.

 

[g67575]

A 47.7 billion transistor GPU would be kind of impressive though.

AMDs flagship Radeon VII GPU (7nm) had just 13.2 billion transistors. That's how far behind AMD is in technology.

Even if they double or triple the transistor count, they will likely still be a bit behind Nvidia in performance. Also, 3x the transistor count seems pretty unlikely in a single generation.

The Xbox X series RDNA 2 GPU has 14.75 billion transistors, wonder if they could double this amount for RDNA 2 desktop GPUs?

 

[BigBANGtheory]

What do you guys think? For the RTX 3080 TI, would ~88% the performance of the GA100 Ampere GPU (in games) be enough of an upgrade for you?

Or, are you more interested in the performance of the RTX 3080 or 3070?

I think its a silly comparison tbh, because assuming a 3080Ti is coming anytime soon (it could well be a year or more away) the performance measure for it would be graphics and for GA100 it would be compute, deep learning and analytics.

Going by the rumours the 3080 seems the most interesting because people say/assume it has a 102 die. That means Nvidia are giving you more potent silicon than they would have ideally wanted to probably because they see a threat from Big Navi not from the kindness of their corporate hearts. Which means you should take AMDs offering seriously