The NVIDIA® H100 Tensor Core GPU is an AI accelerator designed to enhance AI workloads in data centers. It is based on the NVIDIA Ampere architecture and features fourth-generation Tensor Cores, which provide up to 3x faster FP64 performance and up to 6x faster FP32 performance compared to the previous generation. The H100 also supports FP16 and BF16 precision, enabling up to 4x faster throughput for AI workloads. Additionally, the H100 offers support for Multi-instance GPU (MIG) and NVLink, making it well-suited for a variety of server applications.
The standout feature of the Tyan Transport HX FT65-B8050 (sku name: B8050F65TV8E2H-2T-N) server is the compatibility with Nvidia's H100 GPUs. The testing results showcase the remarkable performance and capabilities of this combination:
The standout feature of the Tyan Transport HX FT65-B8050 (sku name: B8050F65TV8E2H-2T-N) server is the compatibility with Nvidia's H100 GPUs. The testing results showcase the remarkable performance and capabilities of this combination:
- GPU_BURN Test:
The Nvidia H100 GPUs showcased their prowess by achieving an impressive 27,000 to 29,000 GFlops/s over a 24-hour period using the GPU_BURN test. These GPUs demonstrated their ability to handle intensive workloads with ease while maintaining a temperature of under 63 degrees Celsius. - Host to Device Bandwidth:
The server exhibited exceptional data transfer speeds with a host to device bandwidth of 56.2 GB/s for pinned memory transfers. This capability ensures that data can be efficiently loaded onto the GPU for processing. - Device to Host Bandwidth:
The Tyan server continued to shine with a device to host bandwidth of 55.1 GB/s for pinned memory transfers. This high-speed data transfer capability is essential for quickly retrieving processed results from the GPU. - Device to Device Bandwidth:
The Tyan Transport HX FT65-B8050 (sku name: B8050F65TV8E2H-2T-N) , equipped with Nvidia's H100 GPUs, showcased its remarkable prowess with a device to device bandwidth of 1,434.0 GB/s for pinned memory transfers. This capability is crucial for scenarios involving complex data transfers and processing within the GPUs themselves. - Matrix Multiplication Performance:
The matrixMultiplyPerf test demonstrated the server's versatility in handling various matrix multiplication operations. The Nvidia H100 GPUs delivered impressive results across different matrix sizes, with the UMhintAs operation standing out as the best performer. - Transpose Test:
The server's ability to efficiently perform matrix transpose operations was highlighted. Different methods were tested, and the results showcased the throughput and time for each operation.