HighPoint just rolled out its new Rocket 7638D PCIe 5.0 switch card, and it’s built specifically to unlock Nvidia’s GPUDirect technology, linking AI GPUs directly with NVMe storage. The whole point here is to speed up AI training and inference workloads by skipping the CPU bottleneck, provided the software stack fully supports GPUDirect.
If you’ve been following NVIDIA’s roadmap, you’ll know that GPUs starting from the A100 already support GPUDirect, which allows GPUs to move data directly to SSDs or network cards without touching the CPU or system memory. The catch, though, is that not every PCIe Gen5 switch supports the P2P DMA capability required, and that’s where switch cards like this step in.
At the heart of the Rocket 7638D is Broadcom’s PEX 89048 switch, giving system builders 48 lanes of PCIe 5.0 connection with 16 going into the host, another 16 connect to external GPU boxes through a CDFP CopprLink connector, and the remaining 16 are dedicated to NVMe storage via MCIO 8i connectors. That storage side can scale up to 16 NVMe drives, pushing configurations as high as 2PB of fast storage.
When deployed, the Rocket 7638D allows GPUDirect Storage workflows to bypass the host CPU and RAM entirely, delivering predictable bandwidth up to 64 GB/s along with lower latency, which is especially valuable when training large datasets that chew through both compute and storage. And since the Broadcom switch inside carries its own Arm-based CPU, the card is self-managed and plays nicely with both Arm and x86 platforms across all major operating systems, no special drivers required.
HighPoint also added practical field-service features like VPD tracking for hardware/firmware matching and health monitoring tools for PCIe link performance, which make troubleshooting easier in big data center setups. Pricing hasn’t been revealed yet, and it’ll likely vary depending on order volumes and configurations.
