Why no Nvidia NVIDIA RTX A6000 and RTX3xxx results so far?

Janosch Simon

New member
Hey there :) i love the idea of an Ai-Benchmark and thinking about perhabs upgrading my RTX2080 ti but it seems to be a great card at the moment would love to now if the 3xxx are better :) are there no results of the RTX3090 as they seem to be not available anywhere?

thx janosch
 

Andrey Ignatov

Administrator
Staff member
@Janosch Simon, thank you for your feedback!

there no results of the RTX3090 as they seem to be not available anywhere?

The reason is that AI Benchmark v1.0.0 with new tests and inference modes is about to be released, thus we decided to update the ranking with new results once they are available. When using the current benchmark version, the scores of the RTX 30XX GPUs are as follows:
  • GeForce RTX 3060: up to 23K
  • GeForce RTX 3070: up to 45K
  • GeForce RTX 3080: up to 60K
The results of the RTX 3900 and RTX A6000 video cards don't demonstrate almost any inference / training speed improvements compared to the RTX 3080 GPU, thus one should basically choose them only if lots of RAM for large batch sizes is needed.
 

Janosch Simon

New member
thats very interessting :) thx for the answer :) so if available the 3080 should be a nice option as its nearly 2x of my 2080 ti

exciting times but im still not getting which faktor is the most important for ai :)
 

gbolcer

New member
Just got an offer to upgrade from 3080 to 3090, so did it. (Yeah, I needed the memory, so while the perf difference isn't significant,it's welcome).

2021-03-28 10:06:46.432643: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1406] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 21821 MB memory) -> physical GPU (device: 0, name: GeForce RTX 3090, pci bus id: 0000:0a:00.0, compute capability: 8.6)
2021-03-28 10:06:46.432699: I tensorflow/compiler/jit/xla_gpu_device.cc:99] Not creating XLA devices, tf_xla_enable_xla_devices not set
19.1 - inference | batch=1, size=1x20: 135.2 ± 0.8 ms

Device Inference Score: 20390
Device Training Score: 20393
Device AI Score: 40783

For more information and results, please visit http://ai-benchmark.com/alpha

PS W:\projects>
 

kheerthiharan

New member
@Janosch Simon, thank you for your feedback!



The reason is that AI Benchmark v1.0.0 with new tests and inference modes is about to be released, thus we decided to update the ranking with new results once they are available. When using the current benchmark version, the scores of the RTX 30XX GPUs are as follows:
  • GeForce RTX 3060: up to 23K
  • GeForce RTX 3070: up to 45K
  • GeForce RTX 3080: up to 60K
The results of the RTX 3900 and RTX A6000 video cards don't demonstrate almost any inference / training speed improvements compared to the RTX 3080 GPU, thus one should basically choose them only if lots of RAM for large batch sizes is needed.
Hi Andrey,
I am using 2 x 3070 but my scores seem quite low:

Device Inference Score: 8959
Device Training Score: 9279
Device AI Score: 18238

Do you have any suggestions?
 

Hiroshi

New member
3080 Ti's result here does not even close to suggested 60K. It looks the estimation was a little optimistic?

* TF Version: 2.6.0
* Platform: Linux-5.11.0-27-generic-x86_64-with-glibc2.29
* CPU: N/A
* CPU RAM: 31 GB
* GPU/0: NVIDIA GeForce RTX 3080 Ti
* GPU RAM: 9.7 GB
* CUDA Version: 11.4
* CUDA Build: V11.4.100

Device Inference Score: 20696
Device Training Score: 21018
Device AI Score: 41714
 

gbolcer

New member
3080 Ti's result here does not even close to suggested 60K. It looks the estimation was a little optimistic?

* TF Version: 2.6.0
* Platform: Linux-5.11.0-27-generic-x86_64-with-glibc2.29
* CPU: N/A
* CPU RAM: 31 GB
* GPU/0: NVIDIA GeForce RTX 3080 Ti
* GPU RAM: 9.7 GB
* CUDA Version: 11.4
* CUDA Build: V11.4.100

Device Inference Score: 20696
Device Training Score: 21018
Device AI Score: 41714

My experience has been that the TF 1.x libraries don't support the GPU compute level of the new 3 series, but perform faster than the properly configured and compiled compute level on the 2.x TF platform. Tensorflow 1.x is definitely faster than Tensorflow 2.x at least for now.
 

rpsantosa

New member
* TF Version: 2.7.0
* Platform: Windows-10-10.0.19041-SP0
* CPU: N/A
* CPU RAM: 64 GB
* GPU/0: NVIDIA GeForce RTX 3070
* GPU RAM: 5.3 GB
* GPU/1:
* GPU RAM: 16.0 GB
* CUDA Version: 11.3
* CUDA Build: V11.3.58


Device Inference Score: 12295
Device Training Score: 12203
Device AI Score: 24498
 
Top