Once upon a time, building a server for high-performance computing meant finding a motherboard with as many CPU sockets and memory slots as possible and topping it off with an ATI Rage Pro with just enough graphics capability for troubleshooting and maintenance. The balance of computing power has shifted, and filling a compact chassis with as many accelerators as possible is now the preferred approach for many problems. Gigabyte developed its latest G190-G30 GPU server for buyers looking to stuff a 1U box with four Nvidia Tesla cards connected by NVLink.
Gigabyte says the G190-G30 can house up to four Nvidia Tesla P100 or Tesla V100 SMX2 modules. Integrated NVLink technology provides the accelerators with a hefty chunk of bandwidth for intercommunication. Tesla modules have up to 300 GB/s to talk with their neighbors over NVLink when conditions are right, compared to a maximum of 32 GB/s over 16 lanes of PCIe 3.0.
The G190-G30 has a pair of LGA 2011-v3 CPU sockets ready to accept Intel Xeon E5-2600 v3 or v4 processors. Each of the 16 memory slots can accept a 3DS LRDIMM of up to 128 GB capacity. The server has four 2.5" SATA drive bays and a pair of low-profile PCIe 3.0 x8 expansion slots. A pair of Gigabit Ethernet ports and an Aspeed AST2400 remote-management controller are built into the motherboard. The machine comes with a 2000-W 80 Plus Titanium-rated power supply.
Gigabyte didn't provide any pricing or availability information, but we think it's safe to expect that building a fully-kitted server based on the G190-G30 GPU server will be an expensive proposition.