NVIDIA Completes Blackwell B300 Tape-Out, 50% Performance Boost

According to SemiAnalysis, NVIDIA has successfully completed the tape-out of its upgraded Blackwell GPU, the B300 chip, and unveiled the improved Grace-Blackwell component, the GB300.

The report highlights that the design goals of the B300 are consistent with those of the previous B100 and B200 chips, utilizing TSMC's 4NP 4nm process. However, adjustments to the design have led to a performance boost, with the B300 achieving up to 50% higher FLOPS compared to the B200. The B300 GPU also incorporates 12-layer HBM3E DRAM, increasing the memory capacity per GPU to 288GB, while maintaining a memory bandwidth of 8TB/s per GPU.

NVIDIA is set to launch the B100, B200 Blackwell GPUs, and the GB200 component in March 2024. The GB200 features an ARM-based CPU paired with two GPUs. The initial launch of the Blackwell components faced challenges due to overheating of GPUs in server racks, which were addressed through complex configurations involving liquid cooling, internal thermal monitoring, and performance management technologies. These solutions require extensive configuration and iteration from large-scale customers and system integrators.

SemiAnalysis notes that the thermal design power (TDP) for the B300 Blackwell components has been increased to 1.2kW per CPU, compared to 1kW for the B200, alongside a larger memory capacity, making it ideal for inference applications.

Cutting-edge semiconductor wafers typically spend about three months in production, and the full cycle, including wafer-level testing, packaging, and component testing, usually takes around six months from tape-out to chip shipment.

The B300 components will feature slight changes in packaging compared to the B200/GB200, with NVIDIA expected to enhance the "SXM Puck" module and BGA packaging for the Grace CPU, allowing customers to source additional components for their compute boards. The recommended host management controller (HMC) will come from the startup Axiado, while the second-layer memory may be sourced from Micron.

NVIDIA CEO Jensen Huang is scheduled to deliver a keynote at CES 2025 in Las Vegas on January 7, where he is likely to confirm the details of the B300 release.