12 Comments
User's avatar
Logan Thorneloe's avatar

Revisiting this because it's incredible. Good work!

Paul's avatar

Great post, thank you very much!

Viswa Kumar's avatar

Great post Abhinav. Learnt a lot. I wonder if you also cover or point me in the direction to understand how a tensor operation would become different at the library level from application point of view . For eg what changes (if any) needs to be done in either pytorch or the likes to better conquer this massive parallelisms offered by TSUs or is this completely taken by the Groq’s compiler behind the scenes . I understand Groqs hasn’t published anything yet but if you came across any nuggets on your research pls do share!

Abhinav Upadhyay's avatar

Not a lot of details on it. But looks like their compiler can take a pytorch Or tensorflow model and compiler for their hardware. But the groq twitter account also hints that sometimes they have to rewrite the code. So it's not quite clear in what situations the compiler works without any manual intervention.

I'm just guessing here.

Michael Spencer's avatar

I'm so impressed with Groq, great to know more about the technical details.

Babbage's avatar

Fantastic post Abhinav! Tremendous research and really topical too.

Abhinav Upadhyay's avatar

Thanks, Babbage :)

Sridaran Thoniyil's avatar

Thanks for the post, Abhinav! This was really insightful and I learned a ton.

Asma's avatar

Great article - loved it!

WIllJ's avatar

Thanks for the post, I wonder whether Groq used any parts of Risc-V, or TSP just works like other ISAs (x86/ARM/Risc-V)