NVIDIA has announced the next-generation Rubin AI accelerators, which will replace Blackwell Ultra in the second half of 2026. Rubin Ultra is scheduled for release in the second half of 2027. They will be accompanied by Arm Vera processors. The series is named after astronomer Vera Florence Cooper Rubin, known for her research into dark matter.
NVIDIA noted that there was a “mistake” in the naming of previous accelerators. In Blackwell, each chip consists of two GPUs, but, for example, the GB200/GB300 NVL72 only mentions 72 GPUs, although it actually refers to 144 GPUs. So, starting with Rubin, the company will use a new naming scheme that no longer takes into account the number of chips, but refers exclusively to the number of GPUs. Thus, the next generation of super accelerators, packaged in the same Oberon rack used for Grace Blackwell, is called Vera Rubin NVL144.
Rubin largely follows the Blackwell design, as the R200 still includes two GPU dies (as part of the SXM7) capable of delivering up to 50 Pflops in FP4 calculations (without sparsity), and 288 GB of memory in eight 12-Hi stacks, but this time it is HBM4 with a total bandwidth of 13 TB/s (2048-bit bus). The GPU dies will be manufactured using the TSMC N3P process technology, and they will be accompanied by two IO chiplets responsible for all external communications, writes SemiAnalysis. Everything together will be packaged using CoWoS-L. The TDP of the new products is not specified.
Image source: NVIDIA
The chips will switch to the NVLink 6 interconnect with a speed of 1.8 TB/s in each direction (3.6 TB/s in duplex), which is twice as fast as the current generation NVLink 5. The switching capacity of the NVSwitch and NVLink C2C will grow in a similar way. However, if the previous scheme is maintained, when one CPU services two GPU modules, each of the latter will apparently get half the bus bandwidth. The Vera processor itself will receive 88 custom (not Neoverse CSS in the case of Grace) 3-nm Arm cores, and with SMT, which will give 176 threads. Each CPU will receive about 1 TB of LPDRR memory and will be twice as fast as Grace with a thermal package of about 50 W.
According to NVIDIA, the VR200 NVL144 will be 3.3 times faster: 3.6 EFLOPS in FP4 calculations for inference and 1.2 EFLOPS in FP8 for training. The total volume of HBM memory will be more than 20.7 TB, system memory – 75 TB. The external network will be represented by ConnectX-9 SuperNIC adapters with a speed of 1.6 Tbps per port, which is twice as much as the ConnectX-8 serving GB300.
In the second half of 2027, the Rubin Ultra (R300) accelerator will appear with FP4 performance of over 100 Pflops (without sparseness), combining four GPUs, two IO chiplets and 16 HBM4e memory stacks 16-Hi with a total capacity of 1 TB (32 TB/s) in an SXM8 package. Moreover, the accelerators will apparently also receive LPDDR memory. The Vera processor will migrate to the new platform unchanged, one CPU will be for four GPUs. The internal bus will be NVLink 7, which will retain the speed of NVLink 6, but will receive a four times more productive NVSwitch switch. But the external connection will still be served by ConnectX-9 adapters.
The new Kyber rack will completely change the layout. The nodes now resemble vertical blade servers used in supercomputers. Each node (VR300) will include one Vera processor and one Rubin Ultra accelerator. There will be 144 such nodes in total, which gives a total of 144 CPUs, 576 GPUs and 144 TB of HBM4e. The Rubin Ultra NVL576 super accelerator will consume 600 kW and provide a performance of 15 EFlops for inference (FP4) and 5 EFlops for training (FP8). At the same time, it is mentioned that the volume of fast memory will be 365 TB, but how much of it will go to the CPU is not specified.
NVIDIA’s future plans include the release of the first accelerator based on the new Feynman architecture, named after theoretical physicist Richard Phillips Feynman, in the second half of 2028. Feynman will reportedly rely on “next-generation” HBM memory and likely Vera CPUs. This generation will also receive NVSwitch 8 (NVL-Next) switches, Spectrum7 network switches, and ConnectX-10 adapters.