R. M. Caplan @sumseq

AI adoption shouldn’t require rebuilding your data center. AMD Instinct MI350P PCIe GPUs are designed to deliver production AI inference performance in standard air-cooled servers, helping enterprises scale AI with existing infrastructure. Learn more: bit.ly/48Hxfl4

22 31 306 22K 16

View Details

R. M. Caplan @sumseq

a month ago

@AMD Will there be a blower-fan version for our workstations?

0 0 0 1K 0

View Details

Predictive Science Inc. @predsci

a month ago

Our POT3D code (buff.ly/QtRzBJc), which is part of SWiG, has been accepted and released as part of the new SPEC CPU 2026 Benchmark suite! More details here: buff.ly/6qmq6lS

0 1 8 96 1

View Details

SPEC @spec_perf

a month ago

The SPEC CPU 2026 benchmark suites are here! As the industry standard for vendor-neutral performance, this suite uses 52 real-world apps to help server buyers evaluate how the latest CPUs, memory & compilers handle intense workloads. 🔗 spec.org/pressreleases/… #SPEC #Server #CPU

0 6 13 2K 5

View Details

Verónica Vergara (she/her/ella) @verolero86

a month ago

Early morning ☀️ ☕ big announcement at #CUG2026, @CrayUserGroup 2027 will be in Houston, TX 🇺🇸 🚀! @exxonmobil will be next year's host

1 4 3 157 0

View Details

R. M. Caplan @sumseq

a month ago

@AnushElangovan Please implement the "reduce" clause on "do concurrent" in Fortran for GPU offloading in either the AOMP or AFAR compiler. Code to test: github.com/predsci/hipft

0 0 2 252 0

View Details

R. M. Caplan @sumseq

a month ago

Writing High-Performance Fortran for GPUs – A Practical Guide (Part I) youtu.be/Rgay5I2HFQ4?si… via @YouTube

0 0 0 49 1

View Details

Predictive Science Inc. @predsci

a month ago

🚨 New paper alert! 🚨 PSI's Roberto Lionello led a team (including U of Reading's Matt Owens) to investigate the magnetic connectivity of the corona/heliosphere system under busy (solar maximum) and quiet (solar minimum) conditions. Details below! buff.ly/lBbygiz

1 1 4 121 0

View Details

FortranTip @fortrantip

a month ago

fortran-lang.discourse.group/t/anecdotal-fo… "The Intel Fortran team was curious if the [compiler] was an enabling technology for the Artemis moon mission completed a few weeks ago. NASA and confirmed that Intel Fortran was used to build their workhorse fluid dynamics solvers OVERFLOW and FUN3D"

0 4 13 413 0

View Details

R. M. Caplan @sumseq

a month ago

@darwesh_singh See here how the performance of HPCG is fully determined (more or less) by the peak memory bandwidth:

0 0 0 29 0

View Details

R. M. Caplan @sumseq

a month ago

This can be misleading. It is important to separate FP64 in a FLOP-bound code versus FP64 in a memory-bandwidth-bound code. In the case of FLOP-bound codes, older GPUs (especially in teh consumer cards) can be better (check out the Titan Black!) In the case of memory-bound codes (like the HPCG benchmark), they typically run at only 1%-3% of peak FLOPs on the best supercomputers. So even if the FP64 is very low on paper, for such codes it can be fine. For example, on my 3090Ti, my codes get 30%-40% of peak FLOPs (since the FP64 is cut by 1/64). I get the same performance as a V100S since the memory bandwidth (1TB/s) is the same (even though the V100S has way higher FP64 peak FLOPs). The reason the Radeon VII is so "good" here is because of its massive 1TB/s bandwidth.

1 0 0 66 0

View Details

R. M. Caplan @sumseq

a month ago

@hopefullyidont1 @darwesh_singh See my coment

0 0 0 10 0

View Details

R. M. Caplan @sumseq

2 months ago

Webinar - Fortran on GPU youtu.be/bymFlvFoMXc?si… via @YouTube

0 0 0 33 0

View Details

R. M. Caplan @sumseq

2 months ago

@__tinygrad__ @AMD @AnushElangovan @LisaSu I want blowers! How else can you stack 4 GPUS in a workstation without special water blocks?

0 0 3 498 0

View Details

R. M. Caplan @sumseq

2 months ago

@NVIDIAHPCDev Actually - this would be kinda cool...

0 0 1 207 0

View Details

R. M. Caplan @sumseq

2 months ago

@TeksEdge We had a A100 PCIe (similar issues) and ended up sending it to Bizon (bizon-tech.com) to integrate into a water-cooled desktop (single slot water block). They used a standard new MB with a Ryzen CPU. It works great! (Linux Ubuntu 24.04)