The Gigabyte Server team has some really interesting GPU compute servers on display at GTC 2018. These span from 10x GPU servers which are in some ways similar to DeepLearning11 to 2U 8x GPU solutions that are similar to some of the offerings by HPE. New for the show is the Gigabyte G190-G30 server which offers 4 GPUs in a 1U with NVLink topology.
Gigabyte G190-G30 Server – 4x GPUs in 1U with NVLink
New for the show is the Gigabyte G190-G30 server. This is a 1U solution that offers dual Intel Xeon E5 V3/ V4 CPUs along with four SXM2 GPUs connected by NVLink. The SXM2 is a form factor primarily used by the NVIDIA Tesla P100 and NVIDIA Tesla V100 GPUs. SXM2 packaging is parallel to the motherboard which aids in cooling. It also allows for NVLink topologies.
We were slightly surprised that the solution utilizes Intel Xeon E5 V3/ V4 CPUs. As we mentioned in other deep learning server articles, E5 V4 chips are in a supply constrained situation right now as Intel is pushing customers towards the Xeon Scalable solution that it launched three quarters ago.
For most deep learning applications, this is not an issue. A large number of deep learning servers, especially PCIe based versions are still on Intel Xeon E5 generation CPUs. In terms of NVLink and GPU topology, this is a single root NVLink solution that supports Pascal and Volta GPUs.
Also interesting on the server is the front I/O. Gigabyte places the 1GbE and IPMI out-of-band network ports in the front of the chassis. High-speed I/O is destined for the rear of the chassis where there are multiple PCIe 3.0 x16 slots for networking and NVMe storage expansion.
The NVIDIA Tesla V100 and Tesla P100 SXM2 modules take well over 250W each including chassis cooling. With 16x DDR4 RDIMMs and two CPUs along with high-speed networking, the power supply needs are significant. The Gigabyte G190-G30 server uses two 2kW power supplies.
Beyond the Gigabyte G190-G30, the company showed off a number of other options including a unique 10x GPU solution along with a 2U 8x GPU solution, both which we have seen before. On a broader scale, we have some new-generation Intel and Cavium-based 2U servers in the STH lab, and Gigabyte has cabling for multiple GPUs in their standard CPU chassis.
More on those solutions soon on STH.