We are excited to announce the immediate availability of a powerful new node configuration to accelerate your GPU workloads on Sherlock: SH4_G8TF64.1. Featuring 8x NVIDIA H200 Tensor Core GPUs, this new configuration delivers cutting-edge performance for AI, machine learning, HPC, and data-intensive research.

The H200 GPUs bring the latest advances in NVIDIA’s Hopper architecture, providing exceptional compute power, memory bandwidth, and AI acceleration. With 8 of these GPUs in a single server, the SH4_G8TF64.1 node is designed to handle large-scale parallel applications, advanced neural network training, and the most demanding GPU-accelerated HPC codes.

An updated SH4_G8TF64 configuration

This updated node configuration builds on the previous SH4_G8TF64 model, ensuring familiarity and consistency for users. This includes:

• 2x Intel 8462Y+ Sapphire Rapids CPUs, 64 physical cores total

• 2 TB of total system RAM

• 4x 7.68 TB NVMe SSDs for ultra-fast node-local storage

• 4x Infiniband NDR (400 Gb/s) network connections for high-bandwidth, low-latency communication

• 8x NVIDIA H200 GPUs, SXM5, each with 141GB of HBM3e memory

And most importantly, this GPU model upgrade comes at no additional cost, and the SH4_G8TF64.1 price remains the same as the previous SH4_G8TF64 configuration.

H200 advantages over H100

The H200 GPU offers key improvements over the previous H100 generation:

• Memory Capacity: 141 GB of HBM3e memory, almost double the H100’s 80 GB, allowing larger models and datasets.

• Memory Bandwidth: Peak bandwidth of 4.8 TB/s, roughly 43% faster than the H100’s 3.35 TB/s, accelerating data-intensive workloads.

• Performance: MLPerf Inference v5.0 results show the H200 delivers up to 28-45% higher throughput compared to H100 on workloads like the Llama 2 70B benchmark, aided by better memory and thermal designs.

These enhancements make the SH4_G8TF64.1 model a best-in-class option for cutting-edge AI research, large-scale simulations, and high-performance data processing.

For more details and pricing, please check out the Sherlock catalog (SUNet ID required).
The new SH4_G8TF64.1 configuration is available for purchase today, and can be ordered online though the Sherlock order form (SUNet ID required).

Stay tuned for further updates as we continue advancing the Sherlock cluster to meet the evolving needs of Stanford’s research community. And as usual, please don’t hesitate to reach out if you have any questions!