Nvidia sets timeline for Blackwell Ultra GPUs, teases Vera Rubin architecture for 2026

Nvidia plans Vera Rubin for 2026, after Blackwell Ultra in late 2025.

: Nvidia has set a timeline for the release of its Blackwell Ultra GPUs, slated for the second half of 2025, following the resolution of a previous design flaw. This series, known as the B300-series, will enhance compute performance and memory capacity, featuring up to 288GB of HBM3E memory. Looking ahead, Nvidia plans to introduce the Vera Rubin architecture in 2026, aiming for significant advancements towards achieving AGI. A future product, Rubin Ultra, may include even greater memory upgrades, pushing technological boundaries further.

In a recent earnings call, Nvidia CEO Jensen Huang confirmed that the company is on track to release its Blackwell Ultra GPUs, known as the B300 series, in the second half of 2025. This series marks a mid-cycle refresh of the Blackwell architecture and promises substantial enhancements over previous versions. These GPUs are anticipated to deliver enhanced compute performance and will be outfitted with eight stacks of 12-Hi HBM3E memory, amounting to 288GB of onboard memory. Although Nvidia has not officially confirmed specific performance metrics, industry observers estimate these GPUs will boast up to a 50% performance increase compared to their B200-series predecessors.

Accompanying the Blackwell GPUs is the Mellanox Spectrum Ultra X800 Ethernet switch. This networking device is anticipated to significantly augment AI and high-performance computing (HPC) systems built on the B300-series by providing a radix of 512 and supporting up to 512 ports. This switch's capabilities are expected to further optimize the performance of data center infrastructures, enhancing Nvidia's footprint in this industry segment.

Looking beyond 2025, Nvidia is developing a future generation of GPUs under the Vera Rubin codename, targeting a 2026 release. Vera Rubin is seen as a pivotal step toward achieving artificial general intelligence (AGI), a concept representing a revolutionary leap in AI development. This architecture is expected to include eight stacks of HBM4E memory with up to 288GB of capacity. Other significant components include a Vera CPU, NVLink 6 switches with an impressive 3600 GB/s operational capacity, and CX9 network cards offering 1,600 Gb/s.

Excitement surrounds Huang's mention of the Rubin Ultra, a product potentially arriving in 2027, which could feature twelve HBM4E stacks, thus doubling the memory capacity to 576GB. The introduction of HBM4E technology is expected to provide outstanding memory bandwidth and capacity, crucial for managing increasingly sophisticated AI models and computations. Achieving this will require managing new technological challenges, such as utilizing larger CoWoS interposers and substrates compared to current-generation GPUs.

Nvidia's ambition does not stop with the Vera Rubin architecture. They have indicated that future-generation products will be discussed at the forthcoming GPU Technology Conference (GTC) in March. This conference will likely elaborate on Nvidia's strategies and technological roadmaps, positioning the company to push the boundaries of GPU design and AI capability further.

Sources: TechSpot, Constellation Research

hardwareAIdata center