Dell has unveiled the PowerEdge C410x, a 3U PCI-Express expansion chassis that can house up to 16 external devices and connect up to as many as 8 servers. The first iteration will be used for GPGPUs, although PCIe flash memory or network adapter cards could presumably be added in later versions.
The initial C410x will support NVIDIA’s Tesla GPU modules: the older 10-series M1050 as well as the new 20-series (Fermi) M2050 (and M2070, when it becomes available). Full loaded with M2050 GPUs, the chassis delivers 16.5 double precision teraflops of peak performance in a 3U space. Dell says this is the first commercial 16-GPU chassis on the market and is 25 percent denser than its closest competitor.
The C410x was originally designed for one of Dell’s Data Center Solutions (DCS) oil and gas customers, who wanted to use GPU acceleration for their seismic codes. But since the customer wasn’t sure how many GPUs they wanted to match with their CPU servers, a static solution (where the GPU modules are integrated into the server itself) wasn’t acceptable. With the C410x design, each server can be configured to talk to between one and four GPUs. Redundant power and cooling is built-in and the individual PCIe modules can be swapped in and out without taking the system down.
The new expansion chassis was announced via Dell’s Inside Enterprise IT blog (link below). The writeup is a tad markety, but it does include a video with DCS systems architect Joe Sekel, who does a nice job of explaining the C410x design. For those wanting more technical detail, the system spec is available here. No pricing has been made public.