Specifications
Model / MPN
G593-SD1-AAX3 · ordering part number 6NG593SD1DR000ABX3
Form Factor
5U rackmount — W 447 mm x H 219.7 mm x D 945 mm
GPU Module
NVIDIA HGX H200, 8x SXM GPUs
GPU Bandwidth
900 GB/s GPU-to-GPU via NVIDIA NVLink and NVSwitch
GPU Memory
8x 141 GB HBM3e = 1.1 TB total
Processors
2x 5th Gen or 4th Gen Intel Xeon Scalable, TDP up to 350 W; Intel Xeon CPU Max Series supported
Socket / Chipset
2x LGA 4677 (Socket E), Intel C741 chipset
Memory
32x DDR5 RDIMM, 8-channel — up to 5600 MT/s (1DPC), 4400 MT/s (2DPC)
Storage
8x 2.5 inch Gen5 NVMe / SATA / SAS-4 hot-swap front bays
PCIe Expansion
4x FHHL PCIe Gen5 x16 plus 8x low-profile PCIe Gen5 x16
Networking
BlueField-3 DPU/SuperNIC compatible; 2x 10GbE via Intel X710-AT2; dedicated management LAN
Management
ASPEED AST2600 BMC with GIGABYTE Management Console web interface
Power
4+2 3000 W 80 PLUS Titanium redundant (100-240 V AC, C19 cords)
Cooling
GPU tray 6x 60x60x76 mm plus 11x 80x80x80 mm fans; board fans plus PCIe-zone fans
Motherboard
MSB3-G41 with dual ROM architecture
Operating Environment
10 to 35 degrees C, 8% to 80% non-condensing
In the Box
Server, 2x CPU heatsinks, 6x carriers, L-shape rail kit
Target Workloads
LLM training and inference, HPC, digital twins, scientific computing
Overview
The GIGABYTE G593-SD1-AAX3 is a 5U dual-socket server built on NVIDIA's HGX H200 eight-GPU baseboard. The SXM module delivers 900 GB/s of GPU-to-GPU bandwidth over NVLink and NVSwitch and carries 1.1 TB of HBM3e across the eight accelerators, which is the memory capacity tier large-model training and long-context inference actually need to keep weights and optimiser state on-device.
Host compute is supplied by two 5th or 4th Gen Intel Xeon Scalable processors at up to 350 W TDP on the LGA 4677 Socket E platform with the C741 chipset. Thirty-two DDR5 RDIMM slots run eight channels per socket at up to 5600 MT/s in single-DIMM-per-channel configurations. Expansion covers both needs in a GPU server: four full-height half-length Gen5 x16 slots for host adapters and eight low-profile Gen5 x16 slots, with BlueField-3 DPU and SuperNIC compatibility for scale-out networking.
Storage is eight front-accessible 2.5 inch Gen5 NVMe bays with SATA and SAS-4 support, and power comes from a 4+2 configuration of 3000 W 80 PLUS Titanium supplies — redundancy that matters when a training run is measured in weeks. Thermal design separates the GPU tray from the board zone, and the AST2600 BMC with GIGABYTE Management Console provides full out-of-band control.
Key Benefits
1.1 TB of HBM3e on-module: eight H200 GPUs keep frontier-scale models resident without host memory spill. 900 GB/s NVLink domain: tensor-parallel and expert-parallel training communicate at full speed. Titanium 4+2 power: duty-configured redundancy plus high efficiency lowers power cost per training FLOP. Gen5 NVMe feed: eight hot-swap Gen5 bays keep data pipelines fed during checkpointing. Documented MPN: the ordering part number 6NG593SD1DR000ABX3 makes BOM and RFQ work unambiguous.
Applications
Large language model pretraining, fine-tuning and inference serving, high-performance computing simulation, digital twin and CFD workloads, scientific research clusters, and enterprise AI factory build-outs where an eight-GPU HGX H200 node with documented SKU-level ordering is required.
Request a Quote — GIGABYTE G593-SD1-AAX3 — 5U 8X NVIDIA HGX H200 AI SERVER (6NG593SD1DR000ABX3)
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →