Nvidia previews Rubin HPC, a 50,000-core GPU for scientific computing

Started by Pele84, Today at 07:36 PM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Topic: Nvidia previews Rubin HPC, a 50,000-core GPU for scientific computing   Views(Read 58 times)
Active members in this topic:
Pele84(1)

Pele84

Nvidia is building a chip aimed squarely at traditional science workloads. SiliconANGLE reports that Rubin HPC will have over 50,000 CUDA cores, a new generation of NVLink and sparse data acceleration, and is due to ship in 2028

For comparison, the flagship Rubin GPU has over 28,000 CUDA cores. Rubin HPC is aimed at double precision work, which climate, physics and engineering simulations depend on

Nvidia also announced Nemotron for Science, built on Nemotron 3.5 Super VL with 124 billion parameters. It comes with nearly 50,000 chemistry questions for post training

A new cuPhoton.xBeam library cuts some Event Horizon Telescope tasks from about a month to eight minutes, Nvidia says. No double precision figures or memory specs have been published yet

Scientists have worried that AI focused chips were neglecting the high precision maths they need, so a dedicated science GPU is welcome. Cutting telescope processing from a month to minutes would change how researchers work. 2028 is a long wait, though

Should Nvidia make more chips for science rather than AI? And which research field would benefit most?