Dual EPYC 9575F · 8× NVIDIA H200 NVL AI Training Server
CG480-S6053-EPYC9575F-H200X8Available to order6 week lead time
Built on the MSI CG480-S6053 4U MGX reference platform, this dual-socket AI training server pairs two AMD EPYC 9575F processors (64 cores / 128 threads each) with eight NVIDIA H200 NVL GPUs delivering 1,128GB of HBM3e memory and up to 900GB/s of NVLink-bridged GPU-to-GPU bandwidth. 1,536GB of 12-channel DDR5 RDIMM memory and a dual-tier NVMe boot and SSD data storage configuration support the largest LLM training and fine-tuning workloads without compromise. Every system ships fully assembled and validated through L6-L10 integration and test, backed by a 3-year system warranty.
- CPU
- 2x AMD EPYC 9575F (Zen 5 “Turin”), 64 cores / 128 threads each — 128 cores / 256 threads total
- Clock Speed
- 3.3GHz base, up to 5.0GHz max boost (4.5GHz all-core boost)
- Cache
- 256MB L3 per processor (512MB total), 64MB L2 per processor
- TDP
- 400W per processor (configurable 320W–400W)
- Socket
- SP5, dual-socket (2P) configuration
- Platform PCIe
- PCIe 5.0, up to 128 lanes per socket
- GPU
- 8x NVIDIA H200 NVL (Hopper architecture), PCIe Gen5, dual-slot, air-cooled
- GPU Memory
- 141GB HBM3e per GPU (1,128GB total), 4.8TB/s memory bandwidth per GPU
- GPU TDP
- Up to 600W per GPU (configurable)
- Tensor Performance (per GPU)
- 3,958 TFLOPS FP8 / 1,979 TFLOPS FP16-BF16 (with sparsity)
- Host Interface
- PCIe Gen5 x16 per GPU (128GB/s)
- NVLink Bridges
- 2x 4th-generation NVLink bridge (H200 NVL 2-/4-way bridge configuration)
- Bridged Bandwidth
- Up to 900GB/s GPU-to-GPU bandwidth within bridged groups
- Remaining Fabric
- PCIe 5.0 x16 host fabric for inter-GPU communication outside bridged groups
- System Memory
- 24x 64GB DDR5 RDIMM ECC — 1,536GB total
- Configuration
- 12 memory channels per socket, 24 DDR5 RDIMM slots fully populated
- Speed
- Up to 6400MT/s
- Per-Socket Bandwidth
- ~614GB/s
- Boot Drive
- 1x 960GB M.2 NVMe SSD (M.2 2280/22110, PCIe 3.0 x2 chassis slot)
- Data Storage
- 2x 3.84TB SSD, front hot-swap 2.5" SFF bay
- Expansion
- Chassis supports up to 8–12x front hot-swap 2.5" SFF bays (U.2 PCIe 5.0 NVMe / SAS / SATA) for additional storage
- Onboard LAN
- 2x 10GBase-T RJ-45 (Intel X710-AT2)
- Management LAN
- 1x dedicated 1000Base-T BMC port
- Power Supplies
- 4x hot-swap redundant, 3200W CRPS, 80 PLUS Titanium (3+1 redundant configuration)
- Form Factor
- 4U rackmount, NVIDIA MGX reference design (MSI CG480-S6053)
- Dimensions
- ~800mm (L) x 440mm (W) x 175mm (H)
- Weight
- ~35kg (bare chassis, excludes installed components)
- GPU Bay Support
- Up to 8x FHFL dual-width PCIe 5.0 GPUs, air-cooled
- BMC
- ASPEED AST2600, IPMI 2.0, DMTF Redfish
- Security
- TPM 2.0, dual BIOS
- Certification
- Platform listed in Red Hat's hardware catalog for RHEL 9.5+ and RHEL AI 1.5+ (x86_64)
- Integration
- Full L6-L10 system assembly, burn-in, and validation testing
- Warranty
- 3-year system warranty
Processor & Chipset
GPU Accelerators
GPU Interconnect
Memory
Storage
Networking
Power
Chassis
Management & Security
Assembly & Support
Priceavailable on request
Get a quote