NVIDIA Tesla A100 with GA100 Ampere GPU based on 7nm, 54b transistors and 6912 CUDA Cores

it matches the leaks too

The full implementation of the GA100 GPU includes the following units:

• 8 GPCs, 8 TPCs/GPC, 2 SMs/TPC, 16 SMs/GPC, 128 SMs per full GPU
• 64 FP32 CUDA Cores/SM, 8192 FP32 CUDA Cores per full GPU
• 4 third-generation Tensor Cores/SM, 512 third-generation Tensor Cores per full GPU
• 6 HBM2 stacks, 12 512-bit memory controllers

The A100 Tensor Core GPU implementation of the GA100 GPU includes the following units:

• 7 GPCs, 7 or 8 TPCs/GPC, 2 SMs/TPC, up to 16 SMs/GPC, 108 SMs
• 64 FP32 CUDA Cores/SM, 6912 FP32 CUDA Cores per GPU
• 4 third-generation Tensor Cores/SM, 432 third-generation Tensor Cores per GPU
• 5 HBM2 stacks, 10 512-bit memory controllers

Leaks

• TSMC 7EUV
  
• GA100 8GPC * 8TPC * 2SM 6144bit

 

What a procreating beast this is for servers. I understand business can pay but it's sad they do this BS for consumers.
The bigger problem is die size: at 826 mm2 it's even bigger than the last generation. Adding the extra cost of the expensive 7nm node, this will without a doubt be more expensive.
We need AMD to, again, come up with a way to disrupt the market. We need multi silicon tech like Ryzen to come for GPUs. Having 3x300mm would be more affordable and beat this monolithic approach in performance.

 

wow, the maximum reticule limit is 858mm2 and Nvidia is pushing 826mm2 on a brand new process node, that's some serious balls they have. Let hope Titan Ampere and 3080 Ti won't cost twice as much as their Turing counterpart...

 

The size includes HBM2, same as GV100 @ 815 mm2

Actual Dies are under 800mm2

 

I can't wait to see a new GPU line crushing all the existing GPUs.

My upgrade path is:

Can't remember > Dual GTX 6600GT > ATI 4870 (300 USD) > GTX 560 ti (MSRP: 250 USD) > GTX 970 (MSRP: 330 USD) > GTX 1080 (MSRP: 600 USD)

The GTX 1080 is the only one I bought "used" (I think for 500€ when the GTX 1080 ti was released), because it started to be in a price range I don't like to play.

 

So if the Titan is 6912, cores hopefully the 3080ti will be around 6500~ cores, Just hoping it's under the £1000 mark this time around. Gotta say based on the table, this is clocked fairly low, maybe due to the insane core number, wonder if it has a lot of OC room.

 

The titan will not be 6912 cores, GA102 is projected to be 5376 and thats only the fully enabled package, the 3080ti will be 5120 at best.

 

You're right, Telsa is normally more than Titan cards, that's my bad, Though I cannot see it being 1800 core difference, having a look at Turning it's 4608 vs 5120 cores (which is 512 core difference or 768 cores vs the 2080ti). Just seems like far too much of a gap right? unless they want a 3080ti super year down the line? and a titan super?

Unsure where you got 5120 from, everyone I am looking shows the 3080ti at 5376 cores, which still baffles me about the huge gap in between unless the Titan is going to have a far bigger difference in core count this time around maybe to allow for that 80TI super this time around?

 

Ti's are almost never fully enabled on a new process
Thats why i'm going with the Ti having 4sm's fused off and the Titan/Quadro 102 having the full 84.

GP102 - 3840
1080ti - 3584
Titan P - 3840

TU102 - 4608
2080ti - 4352
RTX Titan - 4608

And just to point out, even nvidia are fusing parts of the A100, so defect levels must significant. 6912 out of 8120.

 

LOL you blame "user error" then list a bunch if things that aren't user related. Typical cognitive dissonance seen from AMD fan boys.

 

Porsche at least has the good sense to name it an entirely different model if it's skipping 2 or 3 price brackets in a single generation.

 

If their Tesla-class enterprise chip is so severely cut down ( 108 out of 128, or barely 84%, I can't imagine how much lower a theoretical 3080 Ti could be with this defect rate...)

Yields on 7nm must be terrible that they have to fuse so many SMs, and completely drop a stack of HBM2 ... (5 out of 6, one in the picture being just mechanical support)

---
It's still an impressive chip, but damn, not a good start for yields !

Perhaps by the time they release the gaming ones they can produce them with less defects, so the professional line is upgraded to ... let's say, 116/128, and these 108 SM models become the actual 3080 Ti chips.

In any case, a non-Ti 3080 could be 4/6 of the full-fat Ampere, so 84 or 86 SMs as the full chip, or 80 when slightly cut down, that's still a massive 4800 cuda core monster, way better than current Turing 2080 which has only 3072 cuda cores.

If everything being equal (which most certainly is not), 3080 can still be 56% faster than 2080 !! That puts it way above 2080 Ti...
I truly expect that 3070 will be faster than 2080 Ti too, even if just barely...

What a monster !!

 

It´s not a question of balls. Nvidia knows that some customers are desperate for more performance and they don´t care about how much the hardware can cost, so Nvidia is catering for them with this kind of products. It´s just a smart business decision, nothing else.

 

The TSMC 7nm node has been getting a lot work from Apple and AMD in particular, especially with the upcoming consoles. It's likely matured enough to a decent yield rate.

 

"Crysis Remastered is my biyatch." - Lord Ampere

 

The predecessor DGX 2 box that is on the market today is $399,000. This chip is intended to go in the DGX 2's replacement DGX A100 and 8 of these chips will make up the DGXA100.

 

Agree. Im not surprised knowing Nvidia will charge about $400,000 for the DGX A100 which will have 8 of these chips in it. Nvidia has plenty of margin for this beast.

 

That was never expected, this is a HPC conference.

 

Based on info elsewhere it seems like Nvidia decided to go with two 64-core EPYCs. Not bad, and certainly the right choice.