Nvidia Ampere Discussion [2020-05-14]

Scott_Arm said:
I didn't realize they boosted so far beyond their listed clock speeds. I have a gtx1660, but I've never really payed attention to the clocking behaviour.
Click to expand...
It varies from card to card a lot, even 100s of MHzs depending on cooling and other factors. And of course, different loads get different boosts
 
a100_pcieiok2p.png


A100 PCIE variant, 250W TDP for %90 performance.
https://www.nvidia.com/en-us/data-center/a100/

However, there are differences in the TDP, which is set at 250 watts, well below the 400 watts of the SXM4 module. This raises the question of how Nvidia can promote the PCIe card with the same performance values as the SXM4 module. Upon request, Nvidia explained that only the peak performance was identical, but that the continuously (sustained) callable performance of the PCIe card was between 10 percent and 50 percent lower than that of the SXM4 module, depending on the application and configuration of the server.
Click to expand...
https://www.computerbase.de/2020-06/nvidia-ampere-a100-pcie-karte-server/


nvidia-a100-pcie-laun8hkyy.jpg

https://www.servethehome.com/new-nvidia-a100-pcie-add-in-card-launched/
 
Last edited:
https://hardwareleaks.com/2020/06/21/exclusive-first-look-at-nvidias-ampere-gaming-performance/

3080 +30% over 2080TI?

PSman1700 said:
NV's TF numbers are very conservative for that reason. Almost 1800mhz on a high-end chip like that, damn.
Click to expand...

Nvidia doesn't give TF numbers for their consumer cards. For the same reason reviewers don't rely on TF numbers to do their comparisons. Or for the same reason a Radeon VII (14TF) is barely able (in some games) of beating the 2080 (10TF).
 
jayco said:
https://hardwareleaks.com/2020/06/21/exclusive-first-look-at-nvidias-ampere-gaming-performance/

3080 +30% over 2080TI?
Click to expand...
Last page called, it wants the link back

Nvidia doesn't give TF numbers for their consumer cards. For the same reason reviewers don't rely on TF numbers to do their comparisons. Or for the same reason a Radeon VII (14TF) is barely able (in some games) of beating the 2080 (10TF).
Click to expand...
Don't be ridiculous, they stopped reporting TFLOPS only when they launched RTX-generation because the new fancy "RTXOps" they invented makes the difference look far bigger than TFLOPS would.
 
It's fair to critique the precision and utility of the specific "RTXOps" metric. However we can't avoid the uncomfortable reality that the utility of raw TFLOPs as a single metric to judge performance is diminishing as chips move towards more specialization in a post-Moore's Law world. Whether it's even possible to come up with a vendor-agnostic and cross-benchmark metric is a tough question. Even "FPS in game XXX" is becoming more challenging due to the difficulty in performing iso-quality comparisons because of divergent feature sets.
 
All told, the PCIe version of the A100 should deliver about 90% of the performance of the SXM4 version on single-GPU workloads, which for such a big drop in TDP, is not a bad trade-off.
Click to expand...

V100 had 4 iterations: PCIe with 250w, SXM2 with 300w, SXM3 with 350w and 450w. The PCIe version delivered lower sustained performance than the SXM versions, but was not documented, this time NVIDIA is documenting this with the A100.

And on this note, I should give NVIDIA credit where credit is due: unlike the PCIe version of the V100 accelerator, NVIDIA is doing a much better job of documenting these performance differences. This time around NVIDIA is explicitly noting the 90% figure in their their specification sheets and related marketing materials. So there should be a lot less confusion about how the PCIe version of the accelerator compares to the SXM version.
Click to expand...

https://www.anandtech.com/show/1586...lerator-250w-ampere-in-a-standard-form-factor

The drop is greater in Multi GPU workloads though:

The PCIe 4 variant of the Ampere A100 GPU has a significantly lower TDP of 250W vs the SXM's 400W TDP. Nvidia says this leads to a 10% performance penalty in single-GPU workloads. However, multi-GPU scaling is limited to just two GPUs via NVLink, where NVSwitch supports up to eight GPUs. For distributed workloads, Nvidia says the performance penalty (looking at eight GPUs) is up to 33%.
Click to expand...

It appears much of the additional power TDP on the SXM variant goes to inter-GPU communications, based on the 10% performance difference for a single GPU. 400W is also the maximum power limit, so the GPU may not actually use that much in some workloads.
Click to expand...

https://www.tomshardware.com/news/nvidia-drops-its-first-pcie-gen4-gpu-but-its-not-for-gamers
 
Last edited:
Kaotik said:
Pascal launch, TFLOPS number to be seen
View attachment 4143
Click to expand...

Yep, it was so important that then they decided to remove (well, not even bother to add) that number for the official spec list. NVIDIA does marketing and sells a lot of bullshit they days they announce new hardware, but they know it's not an absolute measure of performance, at least not for gaming.

I was checking and there is not a TF number for even the pascal GPUs. I wonder if NVIDIA decided to stop giving meaningless numbers after the RSX announcement.
 
Malo said:
Yeah, because the RTX launch definitely didn't see any meaningless numbers on spec sheets... :nope:
Click to expand...
Ampere hasn't fared any better so far, giving big fat numbers which are really big fat lies (like claiming 20x FP32 performance, not telling you that actually they're comparing FP32 of one card to TF32 of Ampere, and others in the similar fashion)
 
troyan said:
"big fat lies"? nVidia has published every number for every precision, for example in the whitepaper page 23: https://www.nvidia.com/content/dam/...ter/nvidia-ampere-architecture-whitepaper.pdf

And they have put these litte marks under every comparision, for example here: https://developer.nvidia.com/blog/improving-computer-vision-with-nvidia-a100-gpus/

More open isnt even possible...
Click to expand...
Should have been more specific, "in their slides", like this:
https://www.computerbase.de/2020-06/nvidia-ampere-a100-pcie-karte-server/
1-1080.fea5ef29.jpg


That slide doesn't even have fineprint to remind you to check details somewhere else
 
Last edited:
Kaotik said:
Ampere hasn't fared any better so far, giving big fat numbers which are really big fat lies (like claiming 20x FP32 performance, not telling you that actually they're comparing FP32 of one card to TF32 of Ampere, and others in the similar fashion)
Click to expand...
Exactly where do they claim a 20x FP32 performance increase when comparing between Ampere and Volta?
New TensorFloat-32 (TF32) Tensor Core operations in A100 provide an easy path to accelerate FP32 input/output data in DL frameworks and HPC, running 10x faster than V100 FP32 FMA operations or 20x faster with sparsity.

For FP16/FP32 mixed-precision DL, the A100 Tensor Core delivers 2.5x the performance of V100, increasing to 5x with sparsity.

Tensor Core acceleration of INT8, INT4, and binary round out support for DL inferencing, with A100 sparse INT8 running 20x faster than V100 INT8.
Click to expand...
https://developer.nvidia.com/blog/nvidia-ampere-architecture-in-depth/

Other reviews had no problems interpreting the graphs along with what Nvidia stated.
Benchmarking conducted by Nvidia realized speedups for HPC workloads ranging between 1.5x and 2.1x over Volta (see the “Accelerating HPC” chart further down the page). Peak single precision performance gets a theoretical 10-20X boost with the addition of TensorFloat-32 (TF32) tensor cores.
Click to expand...
https://www.hpcwire.com/2020/05/14/nvidias-ampere-a100-gpu-2-5x-the-hpc-20x-the-training/
 
That's a slide from the press deck. And i'm reading FP32 training. TF32 is nVidia's default precision for Training using A100.
 
You must log in or register to reply here.

Similar threads

J
Nintendo Revolution Potential Specs? & Bits On GBA2?
2 3
Replies
49
Views
7K
jvd
J
A
Full transcript of John Carmack's QuakeCon 2005 Keynote
3 4 5
Replies
98
Views
34K
nemisis
N
U
  • Locked
Overclocking Guide
Replies
1
Views
7K
Reverend
R
Back
Top