HighPoint Introduces P2P and GPU Direct Storage Hardware for AI Infrastructures

In traditional server configurations, data transferred from NVMe flash arrays must pass through the host system’s memory (RAM) and CPU before reaching the GPU. This routing process introduces latency and limits throughput, which can reduce the efficiency of GPU accelerators. Although the GDS software enables more direct communication between memory and GPUs, a physical hardware fabric is required to isolate these pipelines from one another and achieve maximum performance.

HighPoint’s new product line utilizes Broadcom’s PEX89048 (Gen5) and PEX88048 (Gen4) PCIe switch engines. These components support a transfer bandwidth of up to 64 GB/s with microsecond-level latency and are available in two different hardware topologies.

CPU-Routed Memory Feeders on the Motherboard

The first topology is designed for platforms that use integrated PCIe slots on the motherboard. This approach employs high-density storage add-in cards (AICs), adapters, and enclosures to consolidate flash storage under a unified switch architecture and utilize local GPUs via a shared CPU root complex.

PCIe Gen 5 Infrastructure (up to 64 GB/s):

• NVMe AICs: Rocket 1608A (8x M.2 slots) and Rocket 1604A (4x M.2 slots).
• NVMe adapters: Rocket 1628A (4x internal MCIO 8i) and Rocket 1624A (4x internal MCIO 4i). PCIe Gen 4 infrastructure (up to 32 GB/s):
• NVMe AICs: SSD7749M2 (16x M.2 slots), SSD7749M (8x M.2 slots), Rocket 1508 (8x M.2 slots), and Rocket 1504 (4x M.2 slots).
• NVMe adapters: Rocket 1528D (4x internal SlimSAS 8i).
• External storage enclosures: RocketStor 6542AW (8-bay U.2/U.3) and RocketStor 6541AW (4-bay M.2/U.2).

These arrays are compatible with various form factors, including standard NVMe, M.2, E1.S, and U.2/U.3 modules.

Hardware-Isolated Compute and Storage Circuits

The second topology was designed for virtualized environments where business-critical isolation is required. In this configuration, the GPU and NVMe storage are hosted on the same physical PCIe switch. This architecture enables data transfers to occur locally on the switch adapter, completely bypassing the host CPU. This method eliminates common performance bottlenecks caused by hypervisor IOMMU or ACS security locks in paths routed through the motherboard.

Key Hardware Components:

• Rocket 1628A: A single-slot GDS matrix with four internal MCIO-8i ports. When combined with an MCIO PCIe x16 G5 bridge card, it provides Gen5 x16 bandwidth to a GPU while simultaneously supporting up to 16 NVMe devices via UBM backplane connectivity.
• Rocket 7638D and Rocket 7634D: These external bridges use CDFP CopprLink connectivity to connect local high-speed storage arrays to external Gen5 x16 eGPU enclosures such as the RocketStor 8631D and the RocketStor 8631C.

Product Availability

The Gen4 and Gen5 switch AICs, adapters, and connectivity accessories optimized for P2P and GDS are available through HighPoint’s global distributor network and system integration partners.