NVIDIA Verifies: “Vera Rubin” in 2026, “Blackwell Ultra” this year
Jensen Huang, the CEO of NVIDIA, made some predictions on upcoming products during the company’s most recent FY2024 earnings call. The second part of 2025 is when the next Blackwell B300 series, codenamed “Blackwell Ultra,” is expected to be released. It will have notable improvements in performance compared to the B200 series. Eight stacks of 12-Hi HBM3E memory, totaling up to 288 GB of onboard memory, will be included inside these GPUs together with a 512-port Mellanox Spectrum Ultra X800 Ethernet switch. According to earlier rumors, this chip has a 1,400 W TBP, which indicates that NVIDIA is cramming a lot of computation onto it. Performance gains of up to 50% are possible when compared to products of the present generation. Although NVIDIA has not formally verified these numbers, they can be achieved using educated guesses about the number of cores and the increase in memory bandwidth.
NVIDIA is getting ready to introduce its next-generation “Rubin” architecture, which looks beyond Blackwell and is expected to provide what Huang called a “big, big, huge step up” in AI computation capabilities. In order to create a complete environment for advanced AI workloads, the Rubin platform, which is aimed for 2026, will incorporate eight stacks of HBM4(E) memory, “Vera” CPUs, NVLink 6 switches that deliver 3600 GB/s bandwidth, CX9 network cards that support 1600 Gb/s, and X1600 switches. Huang revealed something even more unexpected: NVIDIA will talk about post-Rubin technologies at the next GPU Technology Conference in March. Details on Rubin Ultra, which is anticipated for 2027 and might include 12 stacks of HBM4E using 100 mm × 100 mm TSMC substrates and 5.5-reticle-size CoWoS interposers, could be included. This would be another major architectural advancement in the company’s rapidly advancing AI infrastructure plan. These may sound far off, but because of the enormous demand for its products, NVIDIA is fighting supply chain issues to get these GPUs to its clients.
