📢 Exciting FPTalks coming up!
🗓️ Join us October 2nd at 9:00am PT for Sehyeok Park's talk on RLibm!
Updates / discussion: fpbench.org/subscribe
Nominate contributions: fpbench.org/nominate
📢 FPTalks Community Meeting, Mar 5 at 9am PT: Ichitaro Yamazaki (Sandia) on large scale linear solvers, and the reliability headaches that come with new hardware plus techniques like mixed precision, async comms, and sketching. Updates: fpbench.org/subscribe
📢 Last FPTalks of 2025 coming up!
🗓️ Join us Nov 6th at 9:00am PT for Brett Saiki talk on first-class rounding contexts!
Updates / discussion: fpbench.org/subscribe
Nominate contributions: fpbench.org/nominate
📢 Next FPTalks: Thu, Aug 7 @ 9am PT!
Michael Overton (NYU) will spotlight the new cut of Numerical Computing with IEEE FP Arithmetic, updated for modern ML/HPC.
Baranowski & Gopalakrishnan (Utah) will demo GPU-FPX + nixnan for FP exception diagnosis.
fpbench.org/subscribe
fpbench.org/nominate
Next FPTalks: Thu Aug 7 @ 9AM PT
Detecting and diagnosing FP exceptions in GPUs and CPUs
Baranowski & Gopalakrishnan (Utah)
Come learn how we track down NaNs and INFs in closed-source GPU kernels launched from Python, Julia, etc!
fpbench.org/subscribe
fpbench.org/nominate
FPTalks 2025 is tomorrow!
🗓 Thu 10 Jul, 07:50 – 12:10 PDT
🎙 3 one-hour sessions on new FP formats, compilers, ML stability & more
👥 Speakers from CNRS, Graphcore, ARM, Anthropic, Cornell, Utah…
Full schedule + registration: fpbench.org/talks/fptalk...
📢 Last FPTalks of 2025 coming up!
🗓️ Join us Nov 6th at 9:00am PT for Brett Saiki talk on first-class rounding contexts!
Updates / discussion: fpbench.org/subscribe
Nominate contributions: fpbench.org/nominate
Next FPTalks community meeting: Apr 3, 9am PT! Anastasia Isychev (TU Wien) on "Cost of Soundness in Mixed-Precision Tuning" — challenging the assumption that dynamic optimizers always win. Sound tools + regime inference can match or beat them, with formal guarantees. fpbench.org
📢 FPTalks Seminar, June 4 at 9am PT!
Jeffrey Sarnoff (IEEE): "Fast floats for AI: the IEEE P3109 family"; the new standard for 3-bit to 16-bit FP formats designed for ML training and inference.
fptalks.org/subscribe
github.com/P3109/Public
Also, Aug 6 workshop reg now open:
fptalks.org/talks
📢 FPTalk Seminar May 7 at 9am PT! Mantas Mikaitis (Leeds) on NVIDIA tensor cores: each GPU generation rounds differently, none follow IEEE 754. Mantas reverse-engineered bit-accurate models for V100 through B200.
📄 arxiv.org/abs/2512.07004
🎤 fptalks.org/nominate
fptalks.org/subscribe
fptalks.org
