Satoshi Matsuoka

This account is verified. Learn more

@ProfMatsuoka

理研計算科学研究センター長 Director RIKEN R-CCS, 東科大特定教授 Prof. Inst. Sci.. ACM/ISC/JSSST/IPSJ Fellows, IEEE Fernbach(2014)&Cray(2022) Awards, 令4紫綬褒章 Purple Ribbon Medal 2022

神戸市 Kobe, Japan r-ccs.riken.jp Joined July 2009

898 Following 27K Followers 45K Posts

Latest posts 5

Pinned post

Satoshi Matsuoka @ProfMatsuoka Aug 24

(Apologies for JP-only post) 思い立って一般ビジネスパーソン向けのAI利用法の新書「創造的ビジネスパーソン向けのフロンティアAI」をAIを駆使して書きました。正に研究開発で自ら行ってて本に述べられてるAI活用法を改良しながら執筆したので、一種のRecursive Self Improvement の実践でもあります。CC0で配布するので自由にコピーや利用が可能です。drive.google.com/file/d/1oiL-JF… 今後もLive book としてアップデートする予定です。

Satoshi Matsuoka @ProfMatsuoka Oct 18

We are getting to the point where, once memory prices drop & PF quantized perf for PCs becomes general, laptops will run models ‘good nuf’ & a single enterprise GPU serving 10k users (10 % util) meaning we need less than 800k GPUs/NPUs globally, even less :-)

Susan Zhang Verified account @suchenzang Oct 17

it took deepseek one month to move the cost/performance pareto on this benchmark

(and 5 months to move from v4 to v4.1)

zero chance of pacing, what a travesty !

View quoted post on X

Satoshi Matsuoka @ProfMatsuoka Oct 8

Very honored to participate in this important event as one of the panelists today in DC. Hopefully we could demonstrate the new era of US-Japan Scientific progress thru the Genesis-JP AI4S collaboration and other endeavors.

Director Michael Kratsios Verified account @mkratsios47 Oct 7

View quoted post on X

Satoshi Matsuoka @ProfMatsuoka Sep 30

Submit your papers now! It will be a great conference following on the success of 2026 in Osaka which gathered nearly 3000 participants.

Torsten Hoefler 🇨🇭 @thoefler Sep 30

When AI, HPC, and Quantum meet at SC Asia 2027 in SIngapore.

📢 Due to popular demand, the paper submission deadline has been extended.

We invite the community's very best papers.

🔗 sc-asia.org/sca2027-call-f…

#SCA27 #Supercomputing #AI #HPC #QuantumComputing #Singapore

View quoted post on X

Satoshi Matsuoka @ProfMatsuoka Sep 18

Great coverage by Supercomputing News on the progress on my FP8/ Ozaki work. Numerous corrections/updates/additions made. Bad news: FFT difficult due to conversion costs. Good news: this can be overcome with on-the-fly HW conversion. Great news: part of the HW if added to Rubin could improve the DGEMM perf to ~1/2 Petaflops. Bad news: 200 pages / 3 papers now, dense math and algorithms galore…

Supercomputing News @supercompnews Sep 17

@ProfMatsuoka's revised FP8 papers concede the 3-D FFT in software. Four of five kernels reach the memory roof on FP8 tensor cores; the FFT needs a hardware ladder. One rung could roughly double emulated DGEMM on Rubin, per the model. supercomputing.news/hpc/fp64-emula…
View quoted post on X