How the 2070 TFLOPs of Jetson AGX Thor(T5000) is calculated?

Robotics & Edge Computing Jetson Systems Jetson Thor
1

The T5000 datasheet states a maximum of 2070 FP4 TFLOPs, I think this value appears to represent the total performance of each core within the T5000 SoC. What are the maximum computing capabilities of each core, such as the GPU cores, Tensor cores and other? Could you please explain how the 2070 TFLOPS figure was calculated?
Also, what is the maximum performance when converted to TOPS?

image
image702×257 9.63 KB

1 Like
3

Hi,

You can find the info in our Thor spec below:

https://developer.download.nvidia.com/assets/embedded/secure/jetson/thor/docs/Jetson_Thor_Product_Brief_PB-12379-001_v0.1.pdf

Thanks.

4

Hi,

Thank you for your reply.

I am unable to access the document (PDF) linked below.
Is there a special method required to view this document?
https://developer.download.nvidia.com/assets/embedded/secure/jetson/thor/docs/Jetson_Thor_Product_Brief_PB-12379-001_v0.1.pdf

Best Regards,

5

Hi,

The document is shared from the announcement below:

We are checking with our internal team and will update more info to you later.

Thanks.

6

Does Thor support NVFP4 data type? Is it has the same TOPS as FP4, 2070TFLOPS?

7

Hi,

Please find the new version of the document below:

Please check our TensorRT-supported matrix below:

Jetson AGX Thor can natively support FP8 and FP4.
But for NVFP4, we need to double-confirm with our internal team.

Will share more information with you later.
Thanks.

1 Like
8

Thanks, any update?

9

Hi,

We are still checking the details.
Will provide more info to you later.

Thanks.

10

The new version of the specs document you linked says the AGX Thor has Gen 4 Tensor cores but the marketing material says Gen 5. Which is correct? There are other inconsistencies as well.

Having said that, I notice that the specs doc sometimes says Gen 4 and other times Gen 5 so I think it’s just that there are typos and mistakes in the document. Thor is Gen 5.

11

Hi,

Thanks a lot for this information.

Thor has Gen 5 Tensor Core.
We will feedback to our internal team and update the typo.

Thanks.

12

Hi,

As developer community adopts blackwell, we will see more NVFP4 supported models in the open source.

NVIDIA will also release NVFP4 checkpoints for many of the models.
We will be bringing NVFP4 support for the models in VLLM and SGLang in coming months.

Stay tuned!

13

I’m really looking forward to it. Got my Jetson Thor from FedEx today and I’m typing this reply on it.

14

Hi,

You can find our introduction for Thor below:

There are several use cases you can try with.
Thanks.

15

How come you get to use VLLM on the Thor and I don’t? ; ‘)

16

Hi,

You can find some related containers in the link below:

Thanks.

18

The manual says 96 tensor cores. Is that incorrect?

It seems to have 20 SMs, 2560 CUDA cores, and 80 tensor cores.

FP32 non-tensor: 2560 * 1575 * 2 = 8.064 T (MAX)

FP16 tensor: 80 * 1575 * 2048 = 258T (MAX)

FP16 tensor: 80 * 1386 * 2048 = 227T (120Watts)

FP8 tensor: 80 * 1575 * 4096 = 516T

2 Likes
19

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.