Nvidia Blackwell architecture doubles GPU performance

Recently, new rumors have emerged concerning NVIDIA’s next-generation GeForce RTX 50 series, anticipated to employ Blackwell architecture GPUs. Hitherto, it has been asserted that Blackwell would succeed Hopper, corresponding to the GB100 GPU, though it seems a consumer-oriented GB102 GPU may also be in the works.

At last year’s Arete Technology Conference, NVIDIA’s Vice President and Chief of Accelerated Computing, Ian Buck, reiterated the company’s commitment to updating major GPGPU architectures every two years, confirming the launch of Blackwell architecture GPUs in 2024. It is expected that GTC 2024 may witness the debut of Blackwell architecture, with server products likely preceding GeForce graphics cards.

According to a summary from Wccftech, Blackwell architecture GPUs for the GeForce RTX 50 series will maintain a monolithic design, while data center products will adopt MCM packaging. NVIDIA remains unfazed by AMD’s chiplet strategy; continuing with a monolithic design for consumer-grade products offers greater control over timelines and reduced risk, ensuring timely production.

Presently, Hopper and Ada Lovelace architecture GPUs are manufactured using TSMC’s custom 4nm process. NVIDIA plans to extend its collaboration with TSMC for next-generation Blackwell architecture GPUs, utilizing a 3nm process. Insiders suggest that TSMC’s 3nm process costs 25% more than its 5nm process, which may contribute to a price increase for next-generation GPUs.

NVIDIA will implement significant alterations to the Blackwell architecture; although not a completely new design, SM and CUDA will feature a novel structure, and aspects such as ray-tracing performance will be further optimized and enhanced. RT units may be replaced by PT units. Blackwell architecture GPUs are likely to support GDDR7 memory, offering increased efficiency compared to GDDR6X, although Samsung’s GDDR6W may also have a chance. A 512-bit memory bus width is possible, and if operating at 36Gbps, the memory bandwidth would exceed 2TB/s.

Additionally, the GeForce RTX 50 series will adopt the PCIe 5.0 interface, with GPU core frequencies reaching 3GHz or higher. The RTX 5090, based on the GB102, will boast 144 SM clusters, amounting to 18,432 CUDA cores, a 96MB L2 cache, GDDR7 memory, and a PCIe 5.0 x16 interface.