GIGABYTE Technology has unveiled the G262-ZL0 and G492-ZL2, two new liquid-cooled HPC and AI training servers built to push the NVIDIA HGX A100 accelerators and AMD EPYC 7003 processors to the limit. GIGABYTE collaborated with CoolIT Systems to develop a thermal solution that helps prevent overheating and server downtime in compute-dense data centers via direct-liquid cooling technology.
GIGABYTE Technology has unveiled the G262-ZL0 and G492-ZL2, two new liquid-cooled HPC and AI training servers built to push the NVIDIA HGX A100 accelerators and AMD EPYC 7003 processors to the limit. GIGABYTE collaborated with CoolIT Systems to develop a thermal solution that helps prevent overheating and server downtime in compute-dense data centers via direct-liquid cooling technology.
This allows for a balance in optimal performance, high availability, and efficient cooling. The G262-ZL0 is a 2U GPU-centric server that supports the NVIDIA HGX A100 4-GPU baseboard, while the larger G492-ZL2 is a 4U GPU-centric server with the NVIDIA HGX A100 8-GPU baseboard.
GIGABYTE G262 ZL0 front
GIGABYTE’s new servers isolate the GPU baseboard from the other components, which means the accelerators (as well as the dual CPU sockets) can be liquid-cooled to maintain consistent, peak performance. The other chamber houses the CPUs, RAM, storage, and expansion slots. In addition, the servers have multiple 2.5” U.2 bays that support PCIe 4.0 x4 lanes and are equipped with a range of PCIe slots for faster networking using a SmartNIC.
GIGABYTE G492 ZLZ front
These new GIGABYTE servers are also built for the new AMD EPYC 7003 processor and GPU baseboard, NVIDIA HGX A100 80GB accelerator. This makes them ideal for innovators and researchers in HPC, AI, and data analytics, all of which require a high level of CPU and GPU compute.
GIGABYTE G262 ZL0 top view
The flexibility that the NVIDIA HGX A100 platform brings to the new G262-ZL0 and G492-ZL2 servers is important for a lot of use cases, as the new NVIDIA Magnum IO GPUDirect technologies mean faster throughput while offloading workloads from the CPU. As a result, you will see noticeable performance boosts
- GPUDirect RDMA allows for direct data exchange between GPUs and third-party devices like NICs or storage adapters.
- GPUDirect Storage grants a direct data path to move data from storage to GPU memory while offloading the CPU, thus resulting in higher bandwidth and lower latency.
GIGABYTE G492 ZLZ top view
The four NVIDIA A100 server incorporates NVIDIA NVLink to give high-speed interconnects, while the eight NVIDIA A100 server uses NVSwitch and NVLink to enable 600GB/s GPU peer-to-peer communication.
GIGABYTE G262-ZL0 and G492-ZL2 Specifications
Model | G262-ZL0 | G492-ZL2 |
CPU | Dual AMD EPYC 7003 | |
GPU | NVIDIA HGX A100 4-GPU | NVIDIA HGX A100 8-GPU |
Memory | 16 DIMMs, up to 3200MHz | 32 DIMMs, up to 3200MHz |
Storage Slots (Gen4 x4) | 4 x 2.5″ U.2 NVMe/SATA/SAS | 6 x 2.5″ U.2 NVMe/SATA |
3 x M.2 | 2 x M.2 | |
Expansion Slots (Gen4 x16) | 5 x low-profile | 10 x low-profile |
1 x OCP 3.0 | – | |
Network | 2 x 1GbE LAN, 1 x MLAN | 2 x 10GbE LAN, 1 x MLAN |
Power Supply | 2+0 3000W 80+ Platinum | 3+1 3000W 80+ Platinum |
Availability
Those looking to purchase the new G262-ZL0 and G492-ZL2 liquid-cooled servers can contact GIGABYTE directly.
Engage with StorageReview
Newsletter | YouTube | Podcast iTunes/Spotify | Instagram | Twitter | Facebook | TikTok | RSS Feed