NVIDIA's Ian Buck at SC26 disclosed the Rubin HPC GPU (expected 2028), Nemotron for Science toolkit, and production co-packaged photonics deployment at Lambda.
NVIDIA Reveals Rubin HPC GPU, DGX Spark Kit, and Science AI Tools at SC26
The Rubin HPC GPU and DGX Spark represent NVIDIA's next-generation HPC hardware roadmap, with the former not shipping until 2028 and the latter already available. MLPerf results from CoreWeave give the first public multi-rack Vera Rubin performance benchmarks against an earlier generation.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
At SC26, NVIDIA announced the Rubin HPC GPU, the DGX Spark HPC Developer Kit, the Nemotron for Science toolkit, and production deployment of co-packaged photonics switches. The Rubin HPC GPU targets science, engineering, analytics, and AI workloads, carrying over 50,000 CUDA cores (roughly three times the count of the Hopper architecture), supporting a precision range from FP64 down to FP4 with sparse-data acceleration, and connecting via seventh-generation NVLink; it is expected to ship in 2028. The DGX Spark HPC Developer Kit is a four-node desktop cluster linked in a ring via NVIDIA ConnectX-7 networking, requiring no external switch, with 512GB combined memory and 20TB NVMe storage capable of running models up to 700B parameters including DeepSeek V4.1; it is available through the NVIDIA Marketplace and OEM partners. Nemotron for Science pairs an open vision-language model with chemistry datasets and post-training recipes for life sciences, designed to run on a DGX Spark, workstation, or researcher-owned supercomputer. NVIDIA's co-packaged photonics switches are deployed at cloud provider Lambda in what NVIDIA described as one of their first real-world installations, with a claimed 15x improvement in bit-error and link-flap reliability. CoreWeave also published MLPerf Training v6.1 results, claiming to be the first cloud provider to submit multi-rack Vera Rubin benchmark results, training DeepSeek-V3 671B to target quality in 17.65 minutes on 256 Rubin GPUs, 2.7x faster than their GB200 NVL72 result.
How it developed
NVIDIA announced the DGX Spark HPC Developer Kit at SC26, a four-node desktop cluster available through NVIDIA Marketplace and OEM partners.
Sources
Related
- Related storyNvidia Vera CPU deployment contracts
- Google Cloud Vera Rubin GPU instances
- NVIDIA Vera Rubin and NVLink Fusion platform
- CoreWeave's Vera Rubin and Forge at FullyConnected26
- AWS Adds 2M More NVIDIA GPUs to Roadmap for 2027-2028
BackgroundCompute: the chips, clouds and data centers behind AI
Get the daily email
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.