NVIDIA Blackwell GPUs Power Production AI
NVIDIA's Blackwell architecture is being integrated into enterprise-grade servers for production AI workloads, with server vendor YADRO adopting the new H200 and RTX PRO 6000 accelerators. The hardware is also appearing in consumer laptops, such as MSI's Crosshair 16 HX AI, which features the GeForce RTX 5070 Laptop GPU. This adoption coincides with NVIDIA's collaboration with OpenAI, which has nearly doubled the output of the GPT-OSS 120B model.
- The Blackwell architecture, officially announced at GTC on March 18, 2024, is built on a custom TSMC 4NP process and features a dual-die design, connecting two chips with a 10 TB/s interconnect to create a single, unified GPU with 208 billion transistors. - A key architectural improvement is the second-generation Transformer Engine, which adds support for new, smaller data formats like 4-bit floating point (FP4). This allows Blackwell to double the performance and model size that can be supported in memory while maintaining accuracy, delivering up to 2.5 times the performance of the prior Hopper architecture in certain AI workloads. - For large-scale deployments, the liquid-cooled GB200 NVL72 rack system connects 36 Grace CPUs and 72 Blackwell GPUs, designed to function as a single massive GPU for trillion-parameter models. This system provides 1.4 exaflops of AI performance and up to 30TB of fast memory. - The H200 Tensor Core GPU is not part of the Blackwell architecture but is a recent Hopper-based GPU that serves as a bridge, being the first to offer 141GB of HBM3e memory with 4.8 TB/s of bandwidth. This provides nearly double the memory capacity and 1.4 times more bandwidth than the H100. - The professional-grade RTX 6000 series is built on the preceding Ada Lovelace architecture, not Blackwell, and features 48GB of GDDR6 memory, 18,176 CUDA cores, and 568 fourth-generation Tensor Cores. - The consumer-facing GeForce RTX 50 series, based on Blackwell, was announced at CES 2025. Laptops featuring these GPUs, paired with Intel's Arrow Lake-HX and AMD's Strix Point CPUs, are expected to be available in March 2025. - Looking beyond Blackwell, NVIDIA and OpenAI announced a strategic partnership in September 2025 for OpenAI to deploy at least 10 gigawatts of NVIDIA's next-generation "Vera Rubin" platform. The first phase of this large-scale AI infrastructure is planned to come online in the second half of 2026.
