HighPoint Rocket 763…

Posted by

vnlnc

On September 26, 2025

For AI and similar enterprise workloads that are heavily reliant on GPU and storage resources, CPU bottlenecks can cause significant idle time, but enterprise storage leaders at HighPoint Technologies have found a solution with the new HighPoint Rocket 7638D, a 48-lane (32 downstream, 16 upstream) PCIe Gen 5 switch adapter.

Unlike its previous Rocket cards, which were RAID solutions designed to maximize sequential transfers for SSDs to as high as 56 GB/s, the HighPoint Rocket 7638D is not a RAID card. Instead, it exists solely to facilitate faster communication between the graphics hardware and NVMe storage, allowing files to skip the CPU entirely and be sent straight to the GPU using the Rocket 7638D as an intermediary.

According to HighPoint, this effectively eliminates CPU bottlenecks, overhead, and latency in impacted workloads, but especially data-intensive applications like deep learning, model training, inference, and data preprocessing. The card can support up to 2 petabytes of NVMe storage across 16 NVMe drives connected by MCIO 8i ports. HighPoint Technologies highlights its intended use with NVIDIA GPUs, but it’s possible the card will work with other accelerators as well. The company mentions it works with Intel, AMD, and ARM platforms, however.

“For decades, the path from storage to compute has been a bottleneck. Data has had to travel from an NVMe drive, through the host CPU, and into the GPU’s memory—a slow, inefficient process that wastes valuable compute cycles. The Rocket 7638D shatters this conventional model. It is the first 48-lane Gen5 PCIe switch adapter engineered with a dedicated x16 Gen5 pathway for both an external GPU and NVMe storage from a single slot. This architecture creates a direct, peer-to-peer data channel that bypasses the host CPU entirely,” HighPoint Technologies explains in a press release.

In short, the HighPoint Rocket 7638D is a highly specialized solution designed for specific use cases. The CPU bottleneck being eliminated is specifically in workloads where large amounts of data is being routed from storage to the GPU, with no extra processing needed from the CPU. This is obviously not for any consumer workloads and even some enterprise workloads, like CGI and video rendering, which are highly reliant on both GPU and CPU processing power. In those cases, this HighPoint Rocket card won’t really help. But for model training with large datasets and other AI chores, there are some intriguing benefits to be had.

Blog

Leave a Reply Cancel reply