The G492 is GIGABYTE’s second-generation 4U G-series server. Based on the primary era G481 (Intel structure) / G482 (AMD structure) servers, the user-friendly design and scalability have been additional optimized. In addition to supporting two 280 W 2nd Gen AMD EPYC 7002 processors, the 32 DDR4 reminiscence slots assist as much as eight TB of reminiscence and preserve knowledge transmission at 3200 MHz. The G492 has built-in PCIe Genfour switches, which might present extra PCIe Genfour lanes. PCIe Genfour has twice the I/O efficiency of PCIe GenThree and absolutely permits the computing energy of the NVIDIA A100 Tensor Core GPU, or it may be utilized to PCIe storage to assist present a storage improve path that’s native to the G492.
With NVIDIA GPU acceleration turning into the mainstream expertise in knowledge facilities, scientists, researchers, and engineers are dedicated to utilizing GPU-accelerated HPC and AI to satisfy the vital challenges of at this time’s world. According to NVIDIA, the A100 Tensor Core GPU has delivered the very best efficiency leap in comparison with earlier generations. The A100 PCIe GPU additionally maintains the identical 250 W TDP profile and mechanical design because the earlier era V100 GPU, but will increase the HBM2 reminiscence capability to 40 GB. Without altering the code, the velocity of utilizing TensorFloat-32 (TF32) for AI mannequin coaching is six instances the efficiency of the V100. The NVIDIA A100 can deal with AI mannequin processing that’s maturing and quickly rising in each measurement and complexity.
The G492 is effectively designed to assist NVIDIA A100 PCIe GPUs. Considering A100 GPUs utilization, GIGABYTE has constructed PCIe Genfour switches within the system to offer high-speed PCIe mesh networks to assist GPUDirect peer-to-peer (P2P) communications between GPUs and RDMA expertise to parallel a fair bigger computing cluster. Through GPUDirect P2P, every GPU can immediately entry the reminiscence of different GPUs by the PCIe bus, thereby avoiding the switch of knowledge to the server’s system reminiscence and decreasing the delay of knowledge trade. Taking deep studying for instance, well-known open supply deep studying frameworks, equivalent to TensorMovement and MXNet, present assist for GPUDirect P2P, and the NVIDIA Collective Communication library (NCCL) can be optimized for GPUDirect P2P.
The growth of PCIe Genfour lanes by the PCIe Genfour switches additionally makes the expandability of the G492 a lot greater than earlier G481 / G482 servers. In addition to containing 10 dual-slot A100 GPUs within the chassis, three PCIe x16 slots and one OCP 3.zero slot are reserved on the entrance and rear of the chassis, offering customers with a further 4 add-on card improve choices for SAS playing cards or NVIDIA Mellanox InfiniBand playing cards. Low price to efficiency ratio and adaptability are the primary product appeals of the G492. The G492 collection servers present excessive flexibility for customers to configure on their very own and broaden the computing functionality primarily based on demand. GIGABYTE will quickly broaden its NGC-Ready techniques providing with NVIDIA A100 GPUs. The NGC-Ready techniques are constructed for AI purposes and are examined for performance and efficiency of deep studying and machine studying workloads,…
