The Alcatel-Lucent Enterprise OmniCompute 8600 (OC8600) AI Server is a high-performance, scalable, GPU-based server ideally suitable for AI/ML (Artificial Intelligence/Machine Learning) and HPC (High-Performance Computing) applications. The server is ideal for training large language models, automation, object classification, and recognition use cases. The system is designed with eight AMD Instinct MI325X GPUs and dual AMD EPYC 9005/Turin series processors. The AMD MI325X GPU integrates 256 GB of HBM3 memory each.

Each MI325X offers a 16-lane PCIe Gen5 host interface and seven AMD Infinity FabricTM links for full connectivity between eight GPUs. AMD Infinity Fabric technology delivers excellent I/O efficiency, scaling and communication within and between industry-standard accelerator module (OAM) device packages.

• Featuring eight AMD InstinctTM MI325X GPUs based on CDNATM 3 architecture
• Dual 5th Gen AMD EYPC processors with advanced data center performance and efficiency
• 7 Infinity FabricTM links with 128 GB/s for full connectivity between 8 GPUs in a ring (scale-up)
• 16 PCIe Gen5 with 128GB/s bandwidth for scale-out network
• Up to 1.3X AI performance and up to 2.4X HPC performance vs. competitive accelerators
• AMD ROCmTM software optimized for generative AI and HPC applications, easy for migrating existing code

Preliminary picture of the OC8600 AI Server, final product may look slightly different.

• 8 RU

• CPU: AMD EPYCTM 9005/Turin series processor, 2 sockets EPYC 9575F 64C 3.3GHz 400W SP5
• Memory: 24 x 64GB RDIMM 2 RANK 5600
• Operating System: Linux
• BIOS: AMI BIOS (32 MB SPI Flash ROM)
• Front End Interface: 2 x BCM957608-P2200GQF00 Dual-Port 200GbE QSFP112 PCIe Ethernet NIC
• Storage: 2 x 1.92TB U.2 2.5 NVME4 1DWPD (non SED)

• AMD MI325X 8 GPUs with AMD Infinity Fabric

• USB: 2 x USB3.0 ports
• VGA: 1 x D-Sub 15-pin port
• RJ-45: 1 x GbE port dedicated for IPMI
• Buttons: UID button, power button, reset button
• LEDs:
  • UID led
  • PWR led
  • Baseboard management controller
  • Indicator (over temperature warning indicator/fan and PSU fail)

• Remote Direct Memory Access (RDMA), RoCEv2
• 8 x LP slots for switch board (PCIe x16)
• 8 x BCM957608-P1400GDF00 Single-Port 400G QSFP112-DD PCIe Ethernet NIC

• Front:
  • 6 x 7.68TB U.2 2.5 NVME4 1DWPD (non SED)

• AST2600

• 3300 W hot-swap RPSU, 80+ Titanium (4+2 redundancy)

• 15 hot-swappable fans

• W x D x H: 448 x 850 x 351 mm (17.63 x 33.46 x 13.82)

• 10°C - 35°C

• 8 half-height + 4 full-height PCIe 5.0 x16 slots

• ROCm: 6.2.4
• RCCL: 2.20.5
• PyTorch: 2.3, 2.2, 2.1, 2.0, 1.13
• TensorFlow: 2.16.1, 2.15.1, 2.14.1
• JAX: 0.4.26
• ONNX Runtime: 1.17.3