NVIDIA Ampere A100 Delivers 43% Better Performance Than Turing With RTX Off

[erek]

"The NVIDIA A100 is by far the largest 7nm chip produced to date, featuring a gargantuan 54 Billion transistors packed within a single die. The A100 comes with a vastly cut-down configuration due to early yields but like the Tesla V100, we could see a higher bin version with more cores once yields do improve and that would further increase the performance in this specific benchmark.

The full implementation of the NVIDIA Ampere 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 NVIDIA Ampere 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

One can only imagine what the performance metrics would be once the Ampere cards with RTX enabled are brought to market. If this specific benchmark is anything to go by, then we can see Ampere GeForce RTX 30 series cards easily coming in close to their HPC counterparts"

https://wccftech.com/nvidias-ampere...test-gpu-ever-recorded-43-faster-than-turing/

 

[Zarathustra[H]]

"The NVIDIA A100 is by far the largest 7nm chip produced to date, featuring a gargantuan 54 Billion transistors packed within a single die. The A100 comes with a vastly cut-down configuration due to early yields but like the Tesla V100, we could see a higher bin version with more cores once yields do improve and that would further increase the performance in this specific benchmark.

The full implementation of the NVIDIA Ampere 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 NVIDIA Ampere 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

One can only imagine what the performance metrics would be once the Ampere cards with RTX enabled are brought to market. If this specific benchmark is anything to go by, then we can see Ampere GeForce RTX 30 series cards easily coming in close to their HPC counterparts"

https://wccftech.com/nvidias-ampere...test-gpu-ever-recorded-43-faster-than-turing/

Yeah, but GA100 is - by Nvidias past naming conventions - the full size die.

100 = largest full size, goes into Tesla and other professional cards.

102 - slightly smaller or less featured, 104 even more so, etc. The larger the numbers the less featured the die.

The 100 dies don't make it to consumer parts. The only Titan that ever got one was the Titan V which got a GV100, and it was no longer targeted at consumers at that point, and priced much higher.

In the currebnt gen the Titan RTX and 2080 Ti are both TU102's. The 2070 and 2080 are TU104's, and so on.

In the gen before it, the Pascal Titan Xp, Titan X and 1080ti were all GP102's, 1070 through 1080 were GP104's, 1060 was a GP106.

So all I am saying is that these are the professional Tesla type of cards, that may not even have video output, and are intended for compute.

Who knows how they will compare with what consumers wind up seeing. Different animals.

 

[erek]

Yeah, but GA100 is - by Nvidias past naming conventions - the full size die.

100 = largest full size, goes into Tesla and other professional cards.

102 - slightly smaller or less featured, 104 even more so, etc. The larger the numbers the less featured the die.

The 100 dies don't make it to consumer parts. The only Titan that ever got one was the Titan V which got a GV100, and it was no longer targeted at consumers at that point, and priced much higher.

In the currebnt gen the Titan RTX and 2080 Ti are both TU102's. The 2070 and 2080 are TU104's, and so on.

In the gen before it, the Pascal Titan Xp, Titan X and 1080ti were all GP102's, 1070 through 1080 were GP104's, 1060 was a GP106.

So all I am saying is that these are the professional Tesla type of cards, that may not even have video output, and are intended for compute.

Who knows how they will compare with what consumers wind up seeing. Different animals.

I prefer Big variants of all releases

Big Ampere (100) for Desktop Gaming

 

[Zarathustra[H]]

I prefer Big variants of all releases

Big Ampere (100) for Desktop Gaming

Might have a difficult time ever finding one you can connect to a monitor

They usually look something like this

I mean, with some clever hacking you might be able to transfer the framebuffer to another GPU and display it that way, but I'm not convinced that would work well.

That and the drivers for these things are definitely not optimized for games

 

[erek]

Might have a difficult time ever finding one you can connect to a monitor

They usually look something like this

View attachment 264523

I mean, with some clever hacking you might be able to transfer the framebuffer to another GPU and display it that way, but I'm not convinced that would work well.

That and the drivers for these things are definitely not optimized for games

They used to give us Big on Desktop Gaming

i thought Big Fermi or GF100 they gave us the Biggest design? that was last of it? :'(

 

[GoodBoy]

If the information is correct, it is a good starting point for guesstimating where the performance will land for the consumer cards.

I think it may very well be an Oh shit! moment once we see some benchmarks.

Kyle come out of retirement and bench one.

 

[Red Falcon]

Might have a difficult time ever finding one you can connect to a monitor

They usually look something like this

View attachment 264523

I mean, with some clever hacking you might be able to transfer the framebuffer to another GPU and display it that way, but I'm not convinced that would work well.

That and the drivers for these things are definitely not optimized for games

These are designed for compute, so I'm not sure why anyone would want them connected to a monitor or use them for gaming outside of production rendering.
Even a high fraction of this would be amazing performance for a gaming/consumer optimized variant; AMD has their work cut out for them on the GPU front!

 

[polonyc2]

Ampere is going to be a beast, that's pretty much a certainty at this point...hopefully the pricing won't be as crazy as Turing

 

[MangoSeed]

43% faster than which Turing? A100 has 45% more peak flops than the 2080 Ti. Linear scaling is fine but not that exciting.

 

[Vega]

35% faster than RTX Titan. Pretty lame really.

 

[vegeta535]

Ampere is going to be a beast, that's pretty much a certainty at this point...hopefully the pricing won't be as crazy as Turing

Prices will be even crazier.

 

[erek]

35% faster than RTX Titan. Pretty lame really.

is RNDA2 the card to look out for then? Screw Ampere?

cause if Big Ampere is only 35% faster than RTX Titan, the actual consumer card is going to be less exciting

 

[Stryker7314]

I'll take the full fat 8192 with a display output, name your price nvidia

I would legit pay 5k for that card, but it's probably more like a 10k card

 

[erek]

I'll take the full fat 8192 with a display output, name your price nvidia

I would legit pay 5k for that card, but it's probably more like a 10k card

Stryker7314 what about a full fat 8164 instead?

Chipset

8x 3dfx VSA-100

 

[Stryker7314]

Stryker7314 what about a full fat 8164 instead?

Chipset

8x 3dfx VSA-100

View attachment 264570

Gimme gimme!