Nvidia: Blackwell Ultra, Rubin and Vera announced
The Ultra version of Blackwell is the fastest expansion version. Nvidia is also developing its first large server CPU with its own computing cores.
Still visible in the picture: a first-generation Blackwell GPU. Externally, there is little to no difference between it and Blackwell Ultra.
(Image: c't / csp)
Nvidia has announced a faster version of its Blackwell GPU for AI data centers. Blackwell Ultra, alias B300, is said to calculate up to 50 percent faster in compact data formats and at the same time gets a memory upgrade from 192 to 288 GByte HBM3e. Nvidia uses memory chips with twelve instead of eight DRAM layers –, increasing the capacity per chip from 24 to 32 GByte.
It is unclear what exactly is behind B300. At the presentation at Nvidia's in-house exhibition GTC, Nvidia boss Jensen Huang spoke of a 50 percent increase in FP4 computing power: a B300 should be able to achieve 15 instead of 10 petaflops, or up to 30 instead of 20 with Sparsity. This involves removing superfluous zeros in matrices. There were also new “Attention” instructions on a slide.
In the online data sheet for the DGX GB300 server system with the new GPUs, however, Nvidia states an identical computing performance to the previous GB200 variant. It remains to be seen whether Blackwell Ultra uses revised chips, is possibly just clocked higher, or whether the real computing power increases mainly due to the larger memory. Large AI models benefit massively from a lot of fast memory.
Blackwell Ultra systems are expected to appear in the second half of 2025 – from the single GPU to boards with Nvidia's Grace processors to large racks.
(Image:Â Nvidia)
Rubin Ultra as a quad chip
The real successor with a new architecture will follow in 2026: Rubin. The first offshoot looks superficially similar to Blackwell. It remains with two large chips on one carrier, framed by eight memory modules. However, Nvidia is switching to faster HBM4 and the Rubin GPUs are also set to make significant gains under the hood: Huang speaks of up to 3.3 times higher computing power in simple data formats such as FP4.
Huang focused more strongly on the new Rubin Ultra edition for the second half of 2027, which couples two Rubin GPUs, i.e., a total of four chips and 16 HBM4 components. Theoretically, this doubles the computing power, but potentially also the electrical power consumption.
(Image:Â Nvidia)
Vera CPU with custom cores
At the same time, the Grace processor is getting a successor in the form of Vera. Nvidia is switching from ARM's standard Neoverse cores to custom ARM cores developed in-house. Unlike the Neoverse types, these are capable of simultaneous multithreading (SMT) – the total of 88 cores can therefore process 176 threads in parallel.
Videos by heise
The low number of cores is surprising compared to the competition. System integration should be particularly advantageous, as the Vera CPUs can also communicate with GPU accelerators via a fast NV-Link interconnect.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externer Preisvergleich (heise Preisvergleich) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (heise Preisvergleich) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
(mma)
