Skip to main content
Question

Reproducing the FPS/W benchmark claim

  • May 19, 2026
  • 1 reply
  • 8 views

AlbertaBeef
Ensign

Hello,

There seems to be two different published benchmarks for ResNet50.

The first is in the Voyager SDK’s documentation : 1756 FPS (M.2) / 1946 FPS (PCIe) :
https://github.com/axelera-ai-hub/voyager-sdk/blob/latest/docs/reference/model_zoo.md

The second is on the Metis Marketing page : Up to 3200 frames per second (ResNet-50)
https://axelera.ai/ai-accelerators/aipu/metis

 

I was able to achieve 2050 FPS, which is higher than the modelzoo documentation, but MUCH lower than the marketing claim:

https://mariobergeron.com/posts/edge-ai-power-p04-axelera-metis/

 

Where does the 3200 FPS benchmark claim come from ?   How was this benchmark achieved on Metis ?  Which hardware platform ?



Which brings me to the next marketing claim : Industry-Leading FPS / Watt ratios on ResNet50v1 (graph showing 614 FPS/W

https://axelera.ai/ai-accelerators/aipu/metis

 

I was only able to achieve 270 FPS/W :

https://mariobergeron.com/posts/edge-ai-power-p04-axelera-metis/

 

What does the 614 FPS/W benchmark come from ?  Is it for ResNet50 ?  How was it achieved ?

Thank you in advance for any feedback and/or clarifications.

Cheers !

Mario.

1 reply

Spanner
Axelera Team
Forum|alt.badge.img+3
  • Axelera Team
  • May 19, 2026

Hi ​@AlbertaBeef!

Quick correction on the model zoo numbers first: the table at the link you posted actually shows ResNet-50 v1.5 at 1938 PCIe / 1958 M.2 (the columns are PCIe then M.2). Your 2050 FPS is slightly above the published M.2 figure, so you're already reproducing the benchmark.

So, I think the 3200 FPS on the web page is a different measurement: per-chip, inference-only throughput on the AIPU silicon, with no host preprocessing or data transfer in the loop. The 614 FPS/W headline is likely derived on the same chip-only basis. But as you say, the product page doesn't spell this out in much detail so I’ll try to confirm exact methodology for us 👍 

If you want a third-party end-to-end view in the meantime, the HotTech Vision & Analysis report also covers several models and is a really interesting read.