Summary

Derive the GPU accelerator SKU without nvidia-smi, then remove the SMI collector entirely. This lets the snapshot/agent image drop its CUDA base (nvcr.io/nvidia/cuda) and ship as a pure-Go static image built with ko, eliminating the recurring CUDA-image CVE surface.

Motivation / Context

The only data the rest of AICR consumes from the SMI collector is the GPU product name → accelerator SKU (gpu.smi.gpu.model, pkg/fingerprint/from_measurements.go#L41). Driver/CUDA/vbios/MIG/memory fields are collected but consumed by nothing (no fingerprint dimension, no validator code, no recipe constraint).

The SKU is already obtainable without SMI:

• GFD node label nvidia.com/gpu.product (set by GPU Feature Discovery / GPU Operator) is already captured by the topology collector and already feeds reconcileAccelerator (from_measurements.go#L102). In the motivating OKE L40S snapshot, SMI returned 0 GPUs yet the fingerprint still resolved an accelerator from this label.
• PCI device ID is already discovered by the NFD PCI source — extractHardwareInfo iterates the device instances and reads vendor/class (pkg/collector/gpu/nfd.go#L113-L119) but discards dev.Attributes["device"], the device ID needed to identify the SKU driver-free.

nvidia-smi never covered the day-0 case anyway (no driver image ⇒ no nvidia-smi); PCI enumeration is the genuinely driver-free path. So removing SMI in favor of label + PCI is strictly better for AICR's pre-deployment use case, and removes a heavy, CVE-prone base image.

The CUDA dependency is wired into the build, not just the agent pod:

.ko.yaml:15         defaultBaseImage: nvcr.io/nvidia/cuda:13.2.1-runtime-ubuntu24.04
.goreleaser.yaml:152               base_image: nvcr.io/nvidia/cuda:13.2.1-runtime-ubuntu24.04
.settings.yaml:150  snapshot_agent_cuda_image: nvcr.io/nvidia/cuda:13.1.0-base-ubuntu24.04

Fixes: N/A
Related: #1349, #1350 (token-boundary SKU matching — prerequisite; the PCI/label paths must reuse that same normalization + unknown-sku logic)

Proposed approach (phased)

1. (done in fix(fingerprint): match GPU SKUs on token boundaries not substrings #1350) Token-boundary SKU matching so product names / labels classify correctly and unknown SKUs fail safe to unknown-sku.
2. PCI device-ID → accelerator mapping. In extractHardwareInfo, read dev.Attributes["device"], map it to the AICR accelerator enum via a curated device-ID table, and surface it on HardwareInfo (pkg/collector/gpu/hardware.go#L30) + the hardware subtype. Add a fingerprint source for it. Reuse fix(fingerprint): match GPU SKUs on token boundaries not substrings #1350's normalization; unknown device IDs → unknown-sku.
3. Make the GFD label the primary SKU source and remove the SMI collector (collectSMI, the NVSMIDevice XML model, and the nvidia-smi exec in pkg/collector/gpu). Fingerprint precedence becomes: GFD label → PCI device-ID → unknown-sku.
4. Drop the CUDA base image. Flip .ko.yaml / .goreleaser.yaml base to distroless/static, remove snapshot_agent_cuda_image, and build the agent/server image with ko like aicr.

Open decisions (call out before implementation)

• Device-ID table ownership / freshness. Where it lives and how it's kept current (CI check vs periodic review). New device IDs fail safe to unknown-sku until added; NVIDIA-employee maintainers have launch visibility. AICR's enum is coarse (~7 buckets), so the table maps device-ID → bucket, not exact marketing names.
• GFD availability assumption. nvidia.com/gpu.product requires GPU Operator/GFD to have labeled the node. The PCI path is the fallback for nodes before that.
• Residual host access. PCI/module detection still needs host /sys (+ /proc) mounts (pkg/snapshotter/doc.go#L195). Removing SMI drops the driver / CUDA image / --runtime-class nvidia requirement, not the sysfs mount.

Acceptance criteria

• extractHardwareInfo maps PCI device IDs to the accelerator enum; covered by a table-driven test for every supported SKU's known device IDs, with unknown IDs → unknown-sku.
• Fingerprint resolves the SKU with nvidia-smi absent, on both a GFD-labeled node (label path) and an unlabeled node (PCI path).
• SMI collector code removed; no remaining consumers of SMI-only fields.
• Image base switched to static/distroless; nvidia-smi and the CUDA base no longer required to fingerprint; grype scan no longer surfaces CUDA-image CVEs.
• Docs updated (pkg/collector/gpu/doc.go, snapshotter agent docs, any image/runtime-class guidance).

Type of Change

• New feature / enhancement (driver-free SKU detection; dependency removal)
• Bug fix

Risk Assessment

• Low
• Medium — touches GPU collection, fingerprint precedence, and image build; phased and reversible (SMI removal is the last, gated step).