Updated
Updated · HPCwire · Jul 23
DriveNets, AMD Publish MI350 AI Blueprint Claiming 5% Higher Throughput
Updated
Updated · HPCwire · Jul 23

DriveNets, AMD Publish MI350 AI Blueprint Claiming 5% Higher Throughput

3 articles · Updated · HPCwire · Jul 23

Summary

  • DriveNets released a validated reference architecture for AI clusters built on AMD Instinct MI350 GPUs and its AI Fabric, alongside a deployment guide covering design, rollout and tuning.
  • Benchmarking on MI355X clusters showed about 5% higher throughput and 10% to 15% lower time to first token than publicly available industry results, with sub-20 millisecond inter-token latency and at least 50 output tokens per second per user.
  • Resiliency tests found stable collective-communication performance under concurrent RDMA traffic and transient link disruptions, with no observable fabric recovery delays.
  • The companies said the architecture is meant to support open, multi-vendor AI infrastructure across training and inference, while improving GPU utilization and lowering cost per token.
  • AMD's recent participation in DriveNets' $410 million Series D round and a new joint lab for proof-of-concept testing underscore a broader push to win production AI and NeoCloud deployments.

Insights

AMD and DriveNets are challenging a titan. Does their open alliance have the software maturity to truly win the AI race?
This new AI design excels at small scales, but can it maintain its edge in the massive GPU clusters of tomorrow?