AMD ๐ @__tinygrad__
we are looking forward to working closely with @__tinygrad__ to help commoditize the petaflop
- Thats my job ๐
- brick by brick.. We are in it for the long run.AMD's latest GPU software is good, actually. xda-developers.com/way-nvidia-appโฆ
- We are committed to making AMD work well across the board. Lmk if you run into any issues the team is eager to make your experience delightful.Replying to @zephyr_z9๐ Although AMD is now working pretty well for small to medium sized models
- Exactly. Code talks, rest walk. And if you fancy something in low level GPU programming AMD is hiring to power an Open AI ecosystem โก๏ธ[email protected]If youโre a hardcore software engineer and want to build the everything app, please join us by sending your best work to [email protected]. We donโt care where you went to school or even whether you went to school or what โbig nameโ company you worked at. Just show us your code.
- Actually good point. Maybe in the spirit of open source we should make this bounty available to whoever does the best job in Open Source not just to one entity. @__tinygrad__ is welcome to participate in it - and Open Source wins.We've been negotiating a $2M contract to get AMD on MLPerf, but one of the sticking points has been confidentiality. Perhaps posting the deliverables on X will help legal to get in the spirit of open source!
- The beauty of real chiplet design in AMD MI300X. 8 x MI300X in a single node can be partitioned into 64 x MI300X compute partitions with 24GB each. You can unlock advanced pipelining and dataflow. TP=64 in one node anyone ?
- Deepseek-R1 on MI300X outperforming H200 for online serving in the most commonly deployed settings of 8 - 64 concurrent users. (This is MI300X not MI325X which is even better with its FLOPs advantage). Try it out and if you have settings you want us to benchmark leave a note
- AMD is fully committed to Open Source. Stay tuned for more Dev contests from AMD just like the recent kernel developer contest. We continue work with various partners to further the open ecosystem and performance on AMD hardware. Reach out to @roaner / @realSharonZhou or me
- >2X better Deepseek throughput on MI300X vs H200 at the same latency ๐๐๐
- And we are giving away 100 @FrameworkPuter desktops for local AI development - with full Pytorch GPU support.Replying to @shikharontwt @mindcrime and @AnushElangovanThat framework desktop at 1999 is literally the entry point to local AI, with permanent assistant. Need to develop the framework in ROCm to integrate sensors to bring AI to the real world! Hi Anush! (While I writing it you appeared!)
