At #SC26 in Chicago? My @gimletlabs colleague Tom St. John is hosting the MLPerf BoF tonight, 5:15pm CST in W176abc, with MLCommons, Argonne, AMD, Microsoft and Georgia Tech. Come help shape the future of AI benchmarking 👇
https://t.co/j3E7ia3kDW
@israel_ogbole@anbarth Both recordings are up:
@israel_ogbole on finding Jev’s GPU bottlenecks with Zymtrace:
https://t.co/2wb6n3kFNX
Chaim Rand on taming PyTorch recompiles:
https://t.co/AMaFtijdxw
@typesafeai achieved 46% higher GPU throughput while training Jev using @zymtrace’s cluster-wide, continuous profiling and agentic optimization platform.
See how they did it in the link below 👇
@saltyph would love to have you folks on my and @anbarth's AI Systems Performance Engineering meetup. It's online, super casual, 100k members across the globe. @philipkiely was on earlier this year. https://t.co/mSz4NUt1bJ
The new "AI Systems Performance Engineering, the Huang Edition" is now available! Here are some excerpts featuring Jensen, Lori, and Alice Huang.
One real CUDA & AI concept per reply below.
And now the stock market... Let's compare 2 important Monday's in NVDA history. The stock closed down 17.0% on Jan. 27, 2025 and only 4.3% on Nov. 9, 2026 (each versus the previous trading-day closing price.)
Both caused by Chinese influence.
https://t.co/TfoqJjQPR6
https://t.co/o7ZsI6qjnb
NCCL Broadcast is an important optimization in distributed systems.
Here, Jensen writes 1 apology and NCCL Broadcast copies his bytes from the root rank/GPU to other ranks/GPUs, including Lori, Alice, the NVIDIA board, and us shareholders.
Everyone receives it.
https://t.co/CwFcPbcPXc
https://t.co/oRGw7xbYXZ