The company is showing off the servers at the GPU Technology Conference in Washington DC, which opens today.

Its leading product is the SuperServer 4028GR-TXR(T). This box used eight Tesla P100 SXM2 accelerators for high bandwidth for high performance computing (HPC) clusters.

Supermicro said this box uses Nvidia NVLink GPU interconnect tech which it claims delivers over five times the bandwidth PCI-E 3.0. It’s got an independent GPU thermal zone design and claims it will deliver up to 170 TFLOPS FP16 performance.

The company is also showing off its 4U SuperServer 4028-TR(T)2 system which supports up to 10 PCI-E Tesla P100 accelerators. It claims its single root complex PCI-E design will improve GPU peer-to-peer communication efficiency over QPI with up to 21 percent higher throughput and 60 percent lower latency. Both 1U and 4U SuperServers can support dual Intel Xeon processor E5-2600 v4/v3 processors, and will support up to three terabytes of DDR4-2400MHz memory and optional dual onboard 10FBase-T ports.