Implementation Review Summary

Intent

Add UMA (Unified Memory Architecture) support for NVIDIA DGX Spark (GB10) / Blackwell GPUs, falling back to system memory when NVML memory reporting is unavailable and suppressing misleading PCIe metrics.

Findings Addressed

• CRITICAL: Swapped total_memory/used_memory for discrete GPUs in NVML path -- The fix commit (af86da0) corrected UMA destructuring order (total_memory, used_memory) but left the else branch mapping m.used to total_memory and m.total to used_memory. This would have caused all discrete NVIDIA GPUs to display used memory as total and vice versa. Fixed in e511090.

Remaining Items

• MEDIUM: Redundant NVML memory_info() calls -- is_uma_device() calls device.memory_info() once, then the non-UMA branch calls it two more times (one for total, one for used). A single call with destructuring would reduce 3 NVML round-trips to 1. Not blocking since NVML calls are fast, but worth a follow-up.
• LOW: Redundant "Memory Type: Unified" insertion -- The detail map gets "Memory Type: Unified" inserted in both create_device_detail() (cached static info) and again in get_gpu_info_nvml() (per-refresh). The second insertion is harmless but unnecessary since the detail is cloned from the cached version which already contains it.
• LOW: No architecture check in nvidia-smi fallback -- The nvidia-smi fallback path only matches by device name ("gb10", "dgx spark") and cannot check DeviceArchitecture::Blackwell. This is acceptable since the nvidia-smi fallback is inherently limited, but future Blackwell UMA products with different names would need name additions here.

Verification

• All stated requirements implemented
• No placeholder/mock code remaining
• Integrated into project code flow (both NVML and nvidia-smi fallback paths)
• Project conventions followed
• Existing modules reused where applicable (follows Jetson's /proc/meminfo pattern)
• No unintended structural changes
• Tests pass (cargo test: 19 passed, cargo clippy: clean)

PR Finalization Complete

Summary

• Tests: Added 10 new unit tests (nvidia module now has 17 tests, up from 7). Coverage additions:
  • is_uma_device_name: gb10/DGX Spark matching (case-insensitive), non-UMA GPU rejection, empty string
  • read_meminfo_memory: successful parse from a temp file, graceful handling of a missing file path
  • parse_meminfo_content: saturating_sub correctness when available > total, malformed line handling
• Refactor: Extracted is_uma_device_name(name: &str) -> bool pure helper, eliminating the duplicated name-check logic between is_uma_device_with_mem and the nvidia-smi fallback path
• Lint/Format: Collapsed nested if let + inner if into a single if let ... && ... per clippy collapsible_if rule; all other checks were already passing

All checks passing (245 tests, cargo clippy clean, cargo fmt clean). Ready for merge.