Hi everyone, thank you for today’s call (2025.04.17). Specials thanks to @icarus for leading the call and discuss, and to @MalithHatananchchige for sharing insights. This is the generated summary (short version, please find the long version here ): Today’s DeAI Working Group call focused on TensTorrent’s new Black Hole ASIC cards, which offer significant performance upgrades from their previous generation by moving from 12nm to 6nm lithography, using cost-effective GDDR6 memory instead of NVIDIA’s HBM. The group discussed practical implications of model sizes and agreed on the need for standardized open-source models suitable for agentic workflows on the Internet Computer. Malith shared insights on performance bottlenecks in multi-GPU setups and highlighted advantages of NVIDIA’s advanced memory interfaces like NVLink. Lastly, the group acknowledged the necessity of evaluating emerging inference software such as Modular’s inference server as potential improvements over existing solutions.
Links shared during the call:
- Tenstorrent hardware update (Blackhole cards): https://openfuture.tenstorrent.com/
- Blackhole™
- tt-forge/README.md at main · tenstorrent/tt-forge · GitHub
- tt-inference-server/README.md at main · tenstorrent/tt-inference-server · GitHub
- Tenstorrent Bounty Program · GitHub
- https://www.youtube.com/watch?v=S3n_9gX-9lw&t=2400s
- NVIDIA DGX Spark
- NVIDIA H100 Tensor Core GPU
- NVIDIA H200 Tensor Core GPU
- Modular: MAX 25.2: Unleash the power of your H200's–without CUDA!
- New node specification my bucket list. Option 2