NVIDIA's next-gen Rubin GPUs enter production, gets HBM4 samples from all major DRAM makers

NVIDIA's next-generation Rubin GPUs have reportedly entered production, with the company also having all major DRAM manufacturers provide it with HBM4 memory samples.

4

NVIDIA CEO Jensen Huang has been in Taiwan over the last few days, visiting TSMC and having extensive discussions with the semiconductor giant, with NVIDIA seeking more 3nm production capacity for its Blackwell and Rubin AI GPUs, seeing TSMC increase 3nm production by 50%. In even more recent news, UDN reports that NVIDIA's next-gen Rubin AI GPUs have now entered production, and that the company has received HBM4 memory samples, too.

NVIDIA only just received its first Rubin GPU at its labs a few days ago, and now we're at the production stage in just a few days, which is pretty crazy. UDN reports that regarding product and supply chain timing, Jensen said that demand for Blackwell is strong, and that it's not just for GPUs. Jensen explained: "NVIDIA is also manufacturing CPUs, network chips, switches, and many other chips related to Blackwell".

In regards to Rubin, Jensen said that the next-gen Rubin chips are already entering the production line, adding: "we have already seen Rubin on the production line", with TSMC working very hard to support the related AI chip demand.

Each of NVIDIA's next-gen Rubin GPUs are surrounded with a bunch of power circuitry, with each of the Rubin GPU chips sporting 8 HBM4 sites, and two Reticle-sized GPU dies. The Vera CPU features 88 cores and 176 threads of custom Arm-based CPU cores.

4

NVIDIA's next-gen Vera Rubin NVL144 AI server platform will debut in 2026, while its monster Rubin Ultra NVL576 AI server drops in 2027 with upgraded Rubin Ultra GPUs, scaling from 144 GPUs to a huge 576 GPUs. Each of the Rubin Ultra GPUs are beefed up to four Reticle-sized chips (double the chips on Rubin) with gigantic 1TB of even faster HBM4E memory on 16 HBM sites.

There's a total of 13TB of HBM4 memory with 75TB of fast memory, which is a 60% increase over GB300 NVL72, with 2 x the NVLINK and CX9 capabilities, rated at up to 260TB/sec and 28.8TB/sec, respectively. NVIDIA's new Vera Rubin NVL144 launches in the second half of 2026.

4

However, in 2027, NVIDIA will unleash the NVIDIA Rubin Ultra NVL576 AI server platform, which will blow even the next-gen Vera Rubin NVL144. We have 15 Exaflops of FP4 inference power, and 5 Exaflops of FP8 Training capabilities, that represents a crazy 14x increase over the GB300 NVL72, with a whopping 4.6PB (petabytes, not terabytes) of HBM4 memory joined by 365TB of fast memory.