Hi,
I know Metis can be used for ML etc. but wondering whether it’s suitable to do some more low-level operations like big matrix multiplication, FFT of data stored in Metis memory etc. - more like on CUDA (some DSP/Math library)? I assume I could use torch basic math functions?
Dominik
Best answer by Spanner
Great question! And it’d be amazing to see someone create a project like this.
It’s possible to do general-purpose matrix multiplication on Metis, yep, but it needs to be framed as part of a neural network forward path and deployed as an ONNX model using the Gemm operator. It’s not a CUDA-style setup where you can run arbitrary math on memory directly—everything goes through the model deployment pipeline.
There’s a bit more info in the Gemm section of this doc:
👉 https://github.com/axelera-ai-hub/voyager-sdk/blob/release/v1.2.5/docs/reference/onnx-opset14-support.md#gemm
Did you have a specific project or use-case in mind,
+3Great question! And it’d be amazing to see someone create a project like this.
It’s possible to do general-purpose matrix multiplication on Metis, yep, but it needs to be framed as part of a neural network forward path and deployed as an ONNX model using the Gemm operator. It’s not a CUDA-style setup where you can run arbitrary math on memory directly—everything goes through the model deployment pipeline.
There’s a bit more info in the Gemm section of this doc:
👉 https://github.com/axelera-ai-hub/voyager-sdk/blob/release/v1.2.5/docs/reference/onnx-opset14-support.md#gemm
Did you have a specific project or use-case in mind,
My idea is to do some basic image processing on the edge device, where I have multiple images (not a video) taken, where each image might be shifted a bit (camera was moved a bit) and would like to have a one crispy by aligning these images and do some kind of multi-frame image averaging.
Thanks for the link
+3Fascinating! So combining multiple, possibly misaligned or blurred or imperfect images to create one that harnesses the best of all, to produce a high quality image?
Very cool! Keep us posted on that - would love to see it in action one day! How much progress have you made on it so far?
Fascinating! So combining multiple, possibly misaligned or blurred or imperfect images to create one that harnesses the best of all, to produce a high quality image?
Yes, this way we can have (should have ;) ) a much better image. So far, it's in an early development stage. Therefore, looking for options and using your chip popped up as one of the potential options.
+3This is a cool idea, and not one I’ve heard anyone talking about, so you might well be on to something there! Keep us all posted
(By the way, once you’ve found what you need, please hit the “best answer” button to close your question 👍 No rush to do so, if there’s more to come - but when you’re ready 🙂)
Already have an account? Login
Enter your E-mail address. We'll send you an e-mail with instructions to reset your password.
