Serving tech enthusiasts for over 25 years.
TechSpot means tech analysis and advice you can trust.
Forward-looking: Despite facing a setback in the rollout of its Blackwell GPUs for data centers last year due to a design flaw, Nvidia has swiftly rebounded and is poised to deliver its next series of products over the next few years. CEO Jensen Huang confirmed during the company's earnings call that the next major release, dubbed Blackwell Ultra (B300-series), is on track for the second half of this year.
This mid-cycle refresh of the Blackwell architecture promises significant improvements over its predecessors. The B300-series is expected to offer higher compute performance and eight stacks of 12-Hi HBM3E memory, providing up to 288GB of onboard memory. Although unofficial, there are estimates of a 50 percent performance uplift compared to the B200-series.
To complement these powerful GPUs, Nvidia will introduce the Mellanox Spectrum Ultra X800 Ethernet switch, boasting a radix of 512 and support for up to 512 ports. This networking upgrade will further enhance the capabilities of AI and HPC systems built around the B300-series.
Image credit: Constellation Research
Looking beyond Blackwell, Nvidia is already working on its next-generation architecture, codenamed Vera Rubin. Set to debut in 2026, the Rubin GPUs represent a significant step toward achieving artificial general intelligence (AGI).
The Rubin platform will feature eight stacks of HBM4E memory, offering up to 288GB of memory, along with a Vera CPU, NVLink 6 switches operating at 3600 GB/s, CX9 network cards supporting 1,600 Gb/s, and X1600 switches. Huang has hinted at the transformative potential of the Rubin architecture, describing it as a major leap forward in terms of performance and capabilities.
Nvidia has also indicated its readiness to discuss post-Rubin products at the upcoming GPU Technology Conference (GTC) in March. One potential breakthrough on the horizon is the rumored Rubin Ultra, projected for release in 2027. This product could push the boundaries of GPU design even further, potentially incorporating 12 stacks of HBM4E memory. This is a substantial increase from the 8 stacks used in previous generations, potentially offering up to 576GB of total memory. The use of HBM4E technology would provide unprecedented memory bandwidth and capacity, crucial for handling increasingly complex AI models and computations.
To achieve this, Nvidia would need to master the use of 5.5-reticle-size CoWoS interposers and 100mm × 100mm substrates manufactured by TSMC. This represents a significant increase from the current 3.3-reticle-size interposers used in today's most advanced GPUs. The larger interposer size would allow for more components to be integrated onto a single package, enabling the inclusion of additional memory stacks and potentially more GPU tiles.
English (US) ·