Falling in love with my new drug discovery machine. Let's see how proteins fold in a Blackhole.
Moritz Thüning
434 posts
Moritz Thüning
@moritzthuening
AI Drug Design @ Tenstorrent. My views.
- TT-Boltz is more than 1 minute faster again. The blue protein structure in the image was computed in 3 minutes and 20 seconds on a single Tenstorrent Wormhole processor. At Dev Day, it took 15 minutes, and the goal of 2 minutes seemed so far away. I was hoping for a speedup of
- We've integrated FlashAttention into TT-Boltz, reducing the runtime for a protein with 686 amino acids to 4 minutes and 45 seconds. At Tenstorrent Dev Day just over two months ago, it was still 15 minutes. We're getting closer to the 2 minutes of Boltz-1 on an Nvidia RTX 4090
- TT-Boltz now runs on Tenstorrent Blackhole too. Just had to change a single line of code. The runtime dropped from 3m 20s to 2m 30s using only 64 of 140 Tensix cores for the heaviest operations. Very soon we'll use all of those cores. For 9 months my goal was 2 minutes - now
- I'm getting more confident that Tenstorrent boards can become open & better alternatives to GPUs in biotech. Now that I parallelized triangle self-attention of AlphaFold 3 across both Tenstorrent Wormhole processors, we can predict the structure of proteins with up to ~750 amino
- Is this the first time one can buy something other than GPUs to predict protein structures fast? Computed on Tenstorrent Wormhole in 1 minute and 25 seconds.
- I implemented the last module with high computational complexity on Tenstorrent Wormhole and the runtime of TT-Boltz (AlphaFold 3) dropped from 15 minutes to 10 minutes for a protein with 686 amino acids. We only have to save another 8 minutes and we are faster than an Nvidia
- Release of TT-Boltz during the keynote by @DavidBennett__ at Dev Day in SF. What a day! github.com/moritztng/tt-b…
- When I started porting Boltz to Tenstorrent Wormhole just a few days after the release of Boltz, not choosing the original AlphaFold 3 seemed contrarian. Now, Boltz is the industry standard and Boltz-2 can predict binding affinity. We should be so grateful to @GabriCorso,3x in a couple of months... hardware acceleration is so cool!
- 38 seconds nowIs this the first time one can buy something other than GPUs to predict protein structures fast? Computed on Tenstorrent Wormhole in 1 minute and 25 seconds.
- I'll present "AlphaFold 3 on Tenstorrent Wormhole" at Tenstorrent DevDay in San Francisco on April 3rd. Thanks @Saboo_Shubham_ and others at TT! I’ll be around April 1st to 5th. Ping me, if you're interested in making Tenstorrent boards a better alternative to GPUs in biotech :)
- You can now watch my full talk about TT-Boltz at the AI Plumbers Conference in Berlin: open.substack.com/pub/aifoundryo… Hands down one of the coolest conferences I've ever been to. Met so many awesome AI plumbers who build cool stuff there. Loved the deep technical discussions we had,
- Special event. One of the very first Tenstorrent events in Germany. Daisytuner has a super cool product. I'll be there. Please join us and make Wormhole go brrr!!👋 Darmstadt, join us for the #Daisytuner x @tenstorrent hackathon this weekend and learn how to compile legacy code on Wormhole™. Register here --> bit.ly/tt-daisytuner
- Was busy with uni stuff. Tenstorrent is the light at the end of the tunnel. I'm back now. Tenstorrent all day every day. Awesome TT-Boltz update dropping very soon.
