Skip to content

Latest commit

 

History

History
89 lines (74 loc) · 5.23 KB

File metadata and controls

89 lines (74 loc) · 5.23 KB

GPU sharing support matrix (calque#104): live hardware verification

Status: Authoritative current behavior. Verified through: v0.4.0. Resolves the g7 MIG ambiguity that NVIDIA's own datasheets could not settle (Workstation Edition datasheet omits MIG; developer blog confirms MIG only for Server Edition; the MIG User Guide's Supported-GPUs table lists the card family with no edition qualifier). Verified 2026-08-08 by acquiring one real instance per family via calque's own plan.Acquirer (spot, cmd/gpuprobe) and running nvidia-smi -L, nvidia-smi mig -lgip/-lgi, nvidia-smi -mig 1, and a live nvidia-cuda-mps-control -d start on each — ground truth, not vendor literature.

The headline finding

AWS deploys the Server Edition of both RTX PRO 4500 Blackwell (g7) and RTX PRO 6000 Blackwell (g7e)nvidia-smi -L on both reports "Server Edition" explicitly. The earlier datasheet research assumed g7 shipped the Workstation Edition (matching AWS's marketing description, "1 x NVIDIA RTX PRO 4500 GPU," which doesn't name an edition) — that assumption was WRONG. Once the edition is known, the ambiguity dissolves: NVIDIA's own developer blog already said Server Edition supports MIG, and live hardware confirms it.

Results

Card Instance Edition (nvidia-smi -L) MIG MPS Live profiles (mig -lgip)
L4 g6.2xlarge — (no edition name; consumer/OEM-style naming) NoUnable to enable MIG Mode ... Not Supported Yes — daemon started (nvidia-cuda-mps-control -d, pid confirmed via ps) n/a — "No MIG-supported devices found"
L40S g6e.2xlarge — (no edition name) NoUnable to enable MIG Mode ... Not Supported Yes — daemon started n/a — "No MIG-supported devices found"
RTX PRO 4500 Blackwell g7.2xlarge Server Edition YesEnabled MIG Mode for GPU ... succeeded Yes — daemon started 1g.16gb (+gfx/+me variants), 2g.32gb (+gfx) — max 2 instances, matches the MIG User Guide's documented profile set for this card
RTX PRO 6000 Blackwell g7e.2xlarge Server Edition YesEnabled MIG Mode for GPU ... succeeded Yes — daemon started 1g.24gb (+gfx/+me/-me variants), 2g.48gb (+gfx/+me/-me), 4g.96gb (+gfx) — max 4 instances, matches "Universal MIG" up to 4x24GB/2x48GB/1x96GB

Every card's driver reports Driver Version: 595.91.07, CUDA Version: 13.2 — the same stock driver across all four families (via the AMI Deep Learning Base OSS Nvidia Driver GPU AMI (Ubuntu 22.04) 20260804, which AWS's own image description explicitly lists as supporting G4dn/G5/G6/Gr6/G6e/G7/ G7e/P4d/P4de/P5/P5e/P5en/P6). No custom AMI or driver bake is needed for either MIG or MPS on any of the four families — the stock, AWS-published DLAMI already has MIG-capable drivers where the silicon supports it.

SharingMode table (ready for internal/target)

var sharingModeByFamily = map[string]SharingMode{
    "g6":  MPS, // L4:  no MIG (confirmed), MPS works
    "g6e": MPS, // L40S: no MIG (confirmed), MPS works
    "g7":  MIG, // RTX PRO 4500 Blackwell Server Edition: MIG confirmed (max 2 instances, up to 2g.32gb)
    "g7e": MIG, // RTX PRO 6000 Blackwell Server Edition: MIG confirmed (max 4 instances, up to 4g.96gb)
}

Every family supports MPS as a fallback (confirmed live on all four) — a card entered as MIG in the table can still run in MPS mode as an explicit opt-in (per M13's institutional-sharing design), it's just that MIG is the DEFAULT/preferred mode where hardware-isolated slicing is available.

Method note: MPS-before-MIG ordering

The first probe pass attempted nvidia-smi -mig 1 before testing MPS, and the MPS daemon start failed on that run (no process, no error in the log — silently didn't start). Re-running with MPS tested FIRST (before any MIG mode change, which requires a GPU reset) succeeded on every card. This is consistent with NVIDIA's own documented behavior: enabling MIG mode requires the GPU to be reset to take effect, and a card left in that transitional state can reject a subsequent plain (non-MIG) CUDA context that MPS needs. Anyone re-running this probe (or building the real MIG/MPS switching logic in M13) should test/use MPS BEFORE toggling MIG mode on a given boot, or reset between tests.

What this unblocks in M13

  • The per-card SharingMode seam issue can use the table above directly — no further hardware research needed, no "unconfirmed" caveat for g7.
  • The MIG fixed-layout provisioning issue has real, driver-confirmed profile sets to pick from for both g7 and g7e (not just g7e as originally scoped once g7 was thought unconfirmed).
  • The MPS trusted-tenant issue is now confirmed viable on all four families, not just "should work per NVIDIA's docs, no per-GPU restriction stated" — actually observed starting cleanly on every card.

Verification cost

4 spot instances (g6.2xlarge, g6e.2xlarge, g7.2xlarge, g7e.2xlarge), each held for the duration of one SSM diagnostic script (a few minutes), terminated immediately after. No instances leaked (confirmed via a post-hoc describe-instances sweep across the regions used). Tool used: cmd/gpuprobe (calque#104), a one-off diagnostic — not part of calque's product CLI surface.