Provide peak Tensor throughput performance numbers for GPUs

[elevated_quark]

The title is pretty much the ask.

I'm a Machine Learning engineer, and evaluating Intel's GPU offerings for Deep Learning work (small-medium scale experiments).

Please provide peak Tensor throughput performance numbers for the following:

1. FP32 FLOPS

2. FP16 FLOPS

3. BF16 FLOPS (also whether or not it is supported)

None of this information is available from a credible source as of writing.

[Andres_Intel]

Hello elevated_quark,

Thank you for posting on the Intel®️ communities. I know how important is or you to know about the FLOPS, I will be happy to help you.  

To understand your request as clear as possible, answer the following question:

• Are you looking for this information for a specific product? If so, let me know the model.

Regards,  

Andres P. 

Intel Customer Support Technician 

[elevated_quark]

Arc A770 and A750

[Andres_Intel]

Hello elevated_quark,

Thank you for your response, and for the information provided.

I will start with an investigation to provide you with the information you need, as soon I have it I will let you know.

Regards,  

Andres P. 

Intel Customer Support Technician 

[Andres_Intel]

Hello elevated_quark,

Thank you for your time.

We have been working on the investigation of your request, now I have a couple of questions for clarification, please answer them:

• Are you referring to Tensorflow as the tool you are using? Or are you using a different tool specifically?

For more information about Tensorflow:

Intel® Extension for TensorFlow.

Running TensorFlow* Stable Diffusion on Intel® Arc™ GPUs 

Regards,  

Andres P. 

Intel Customer Support Technician

[elevated_quark]

Hi Andres,

The information I'm looking for, is independent of what tool is used.

I'm asking for the "peak Tensor FLOPS performance" numbers for Arc A770 and A750 at an architectural level and utilization.

If you need an example, here's one: https://www.nvidia.com/content/PDF/nvidia-ampere-ga-102-gpu-architecture-whitepaper-v2.pdf

Again, I'm not looking for a full-blown whitepaper, only the TFLOPS numbers.

-K

[Andres_Intel]

Hello elevated_quark,

Thank you for your explanation and clarification, that helps a lot.

Now, I will continue with the investigation to provide you with the information that you need as soon as possible.

Regards,  

Andres P. 

Intel Customer Support Technician

[Andres_Intel]

Hello elevated_quark,

Thank you for your wait time and patience.

We have been working on the investigation, for Arc GPUs, the online manuals. Check 'Volume 4 - Configurations' and download the PDF. This volume provides device attributes, including FLOPS/Clk for Half Precision and Single Precision on page 7. From that the TFLOPS can be calculated.

Another source of public information is the , section “oneAPI GPU Optimization Guide”. On that page you will find a table of “Xe Configurations”, with FLOPs/clk for single-precision and half-precision, including for the A770 and Data Center GPU Flex 170. Again, from that TFLOPS can be calculated.

Wikipedia also has a lists TFLOPS in an easier-to-read table although the sources can't be verified but can serve as a baseline for subsequent TFLOPS calculations.

Let me know if you have further questions

Regards,  

Andres P. 

Intel Customer Support Technician 

[Andres_Intel]

Hello elevated_quark,

Were you able to check the previous post? Remember to check the online manuals, oneAPI GPU Optimization Guide, and Wikipedia for FLOPS information.

Let us know if you still need assistance.    

Best regards,   

Andres P.   

Intel Customer Support Technician 

[Andres_Intel]

Hello elevated_quark,

We have not heard back from you, so we will close this thread. If you need any additional information, please submit a new question as this thread will no longer be monitored.  

Best regards, 

Andres P. 

Intel Customer Support Technician