Nvidia announces GB200 Grace Blackwell superchip merging CPU and GPU with four B200 GPUs and two Grace CPUs.
Nvidia has announced two new GPU products targeting different segments of the data center market. The GB200 Grace Blackwell NVL4 Superchip represents the high-performance end, featuring four B200 Blackwell GPUs connected via NVLink and two Grace ARM-based CPUs all on a single motherboard. This configuration delivers 1.3TB of coherent memory and is aimed at HPC and AI-hybrid workloads. According to Nvidia's performance claims, the GB200 NVL4 achieves 2.2X the simulation, 1.8X the training, and 1.8X the inference performance compared to the Nvidia GH200 NVL4 Grace Hopper Superchip, its direct predecessor. The GB200 NVL4 will be available in 2H 2025 from multiple providers including MSI, Asus, Gigabyte, Wistron, Pegatron, ASRock Rack, Lenovo, HP Enterprise, and others.
On the opposite end of the spectrum, Nvidia is releasing the H200 NVL, a dual-slot air-cooled GPU designed for existing data center infrastructure. The H200 NVL features PCIe 5.0 connectivity with 128 GB/s bandwidth and uses a flow-through cooling design optimized for rack-mount solutions, with intake air running from right to left and no blower-style fan. The GPU delivers 30 TFLOPS of FP64 and 60 TFLOPS of FP32 performance, with tensor core capabilities rated at 60 TFLOPS of FP64, 835 TFLOPS of TF32, 1,671 TFLOPS of BFLOAT16, 1,671 TFLOPS of FP16, 3,341 TFLOPS of FP8, and 3,341 TFLOPS of INT8.
Compared to its predecessor, the H200 NVL significantly outperforms the H100 NVL it replaces, offering 1.5X the memory capacity and 1.2X the memory bandwidth. This translates to up to 1.7X faster inference performance and 1.3X faster performance for HPC workloads. When benchmarked against Ampere's equivalent GPUs, the H200 NVL is 2.5X faster. The GPU also supports NVLink, enabling up to 900 GB/s of bandwidth per GPU and allowing system providers to connect up to four GPUs in a single rig.
The H200 NVL is optimized for the majority of existing data center configurations. According to Nvidia's survey data, roughly 70% of enterprise racks use air cooling with 20kW of power or less. As a PCIe GPU, the H200 NVL allows data center providers to reuse existing racks and replace only the GPUs, reducing waste and significantly lowering hardware upgrade costs. The H200 NVL will be available from Dell, HP Enterprise, Lenovo, and Supermicro, as well as from additional platform providers including Aivres, ASRock Rack, Asus, Gigabyte, Ingrasys, Inventec, MSI, Pegatron, QCT, Wistron, and Wiwynn.