Nvidia scores its first DOE win since 2022 with Doudna supercomputer The US Department of Energy's next supercomputer will be built by Dell Technologies and powered by Nvidia's next-gen Vera-Rubin accelerators - a notable switch from the usual Cray-AMD tag teams that build such machines. It's the first DOE win for Nvidia since the Venado system in 2022. Named for Nobel laureate Jennifer Doudna …
Yeah, a nice upgrade to Perlmutter (#19 on Top500) ... but the relatively low FP64 (tensor) performance is unfortunate, imho (will have to wait and see where Doudna ends up in HPL, HPCG, and Green500 I guess). But Nvidia's decision to focus their designs on performance at TF32 and lower precisions is clearly their own.
The divergence of HPC (proper) and AI computational requirements has been highlighted for a while, and it looks like we might have reached that fork in the road, with (for example) AMD developing separate MI430X UL4 and MI450X for each of those.
On the AI side though, trends suggest even FP32 is overkill, with -1,0,1 quantization, or 1.58 bits being all the rage, potentially down to 0-bit neuromorphs, and alternative architectures to SIMD GPUs seem to have a performance advantage, so ...
FWIW it's probably best not to keep all of one's dinner eggs in just one computational paradigm basket here, or at least lay them on a diversity of sides, to hedge investments against cracks, bursts, crashes, cold turkey sandwiches, and the likes. And when playing this roulette, always bet on HPC, imho (it's still the new black!)!
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
