{
    "submitter": "Crusoe",
    "division": "closed",
    "status": "available",
    "system_type": "datacenter",
    "system_type_detail": "cloud",
    "system_name": "Crusoe Cloud 64x8 MI355X (512 GPU) — gpt-oss-120b / vLLM",
    "number_of_nodes": 64,
    "host_processor_model_name": "AMD EPYC 9575F",
    "host_processors_per_node": 2,
    "host_processor_core_count": "64",
    "host_processor_frequency": "3.3 GHz base / 5.0 GHz boost",
    "host_processor_caches": "",
    "host_processor_interconnect": "",
    "host_memory_capacity": "3000 GB",
    "host_memory_configuration": "DDR5",
    "host_storage_capacity": "30 TB (8x 3.84TB NVMe)",
    "host_storage_type": "NVMe SSD",
    "filesystem": "Crusoe Shared Disk (NFS)",
    "host_networking": "8x AMD Pollara 400G RoCE per node (3200 Gbps aggregate/node)",
    "host_networking_topology": "Ethernet/RoCE; multi-node = TCP token streams (no cross-node RCCL)",
    "host_network_card_count": "8 x 400Gbit/s per node",
    "accelerators_per_node": 8,
    "accelerator_model_name": "AMD Instinct MI355X 288GB HBM3e",
    "accelerator_memory_capacity": "288GB",
    "accelerator_memory_configuration": "HBM3E",
    "accelerator_host_interconnect": "PCIe Gen5 x16",
    "accelerator_interconnect": "XGMI (intra-node)",
    "accelerator_interconnect_topology": "intra-node 8-GPU XGMI; inter-node RoCE (token streams only)",
    "accelerator_frequency": "",
    "accelerator_on-chip_memories": "",
    "cooling": "Direct liquid-cooled",
    "framework": "vLLM 0.22.1 (ROCm 7.2.2), PyTorch 2.10",
    "operating_system": "Ubuntu 24.04 (node amdgpu driver 6.14.x; container ROCm 7.2.2 userspace)",
    "other_software_stack": "AITER, hipBLASLt; ZMQ distributed SUT (1 dedicated head + 64x 8-GPU nodes = 512 single-GPU vLLM replicas)",
    "other_hardware": "",
    "hw_notes": "Crusoe Cloud mi355x-288gb-roce.8x, 64 nodes = 512 MI355X GPUs. First-ever 512-GPU gpt-oss-120b (Offline + Server).",
    "sw_notes": "AMD MLPerf Inference v6.1 gpt-oss-120b harness (vLLM backend). Native-MXFP4 model openai/gpt-oss-120b@b5c939d (NOT the v6.0 Quark repack). tensor_parallel_size=1 (one GPU per replica, 8 replicas/node); device_count=512. Dedicated ZMQ head dispatches Offline sample-ranges / Server per-query and relays token streams."
}