{
  "what": "MIG partitioning on the H200 NVL: what was measured and what was not",
  "host": "h200-ref",
  "date_utc": "2026-08-03",
  "gpu": "NVIDIA H200 NVL, 141 GB",
  "operator_actions_required": [
    "stop the akili trio and the Sionna RT process (MIG needs a compute-free GPU)",
    "sudo nvidia-smi -i 0 -mig 1 (privileged; not available to the agent)"
  ],
  "profiles_available": {"1g.18gb": 7, "1g.18gb+me": 1, "3g.71gb": 2, "7g.141gb": 1},
  "instances_created": 3,
  "instance_profile": "1g.18gb",
  "measured": {
    "memory_visible_to_process_gb": 17.2,
    "memory_visible_vs_whole_gpu": "17.2 of 141, a HARD partition; MPS cannot do this",
    "fp16_4096_matmul_tflops_per_instance_sequential": [93.64, 93.38, 93.29],
    "fp16_4096_matmul_tflops_instance_1_concurrent": 93.6
  },
  "NOT_measured": {
    "three_way_concurrent_isolation": "Instances 2 and 3 produced NO output when all three ran at once, though each works sequentially. The concurrent non-interference claim is therefore UNPROVEN and is not made. Cause not diagnosed; the window had to close to restore a user-facing service.",
    "cuphy_ldpc_inside_a_mig_instance": "not attempted in this window"
  },
  "comparison_to_mps_on_gb10": {
    "mps_cap_100_tflops": 79.18,
    "mps_partitions_memory": false,
    "mig_partitions_memory": true,
    "reading": "MPS caps SM share among clients that joined its server and gives no memory isolation; MIG hard-partitions both. For an AI-on-RAN deployment where an L1 tenant must not be starved by an AI tenant, MIG is the stronger primitive where the hardware offers it, and the GB10 offers none."
  },
  "teardown": "compute instances and GPU instances destroyed, MIG mode returned to Disabled, akili trio restarted and verified serving in 56 s with 95772 MiB restored"
}