Keep four labels separate
Architecture family
A shared design generation, such as Turing or Hopper.
Chip implementation
A particular silicon design implementing that architecture.
Product or system
A sellable GPU, module, board, server, or cloud offering.
Compute capability
A CUDA programming-model designation for hardware features.
A selected lineage, not a complete database
Each family contains variation. A family label alone cannot establish memory capacity, form factor, software support, delivery date, or application performance.
[6][10][11][12][14][18][23][114]A safer comparison sequence
- Name the workload. Graphics, training, inference, HPC, and media do not weight features equally.
- Name the integration level. Compare GPU with GPU or system with system.
- Check the exact product. Architecture does not fix memory, power, interconnect, or form factor.
- Check the software envelope. Driver, toolkit, libraries, and framework requirements may be decisive.
- Use measured evidence. Do not turn a compute-capability number into a performance score.
Rubin: hardware availability and software support are separate
NVIDIA’s May 31 announcement describes Vera Rubin as ramping into production and says production shipments are planned for fall 2026. This is a vendor announcement, not confirmation of stock at every partner.
Software states are separate: CUDA 13.4 highlights describe Rubin preview support, while TensorRT 11.3.0 Enterprise lists general-release Rubin support. Check the component and platform matrix you will actually use.
[114][113][115]