GTX 1650: disambiguate GDDR5/GDDR6 variants by memory clock (measured 192 GB/s)

[cms-pm]

Summary

GPU_BANDWIDTH["GTX 1650"] = 128.0 is correct for the original GDDR5 card,
but the GTX 1650 also shipped in a later GDDR6 revision at 192 GB/s
(12 Gbps × 128-bit). Both use the same TU117 die and PCI device id 0x1F82, and
the driver reports the same name, so the curated single value under-states the
GDDR6 variant by 50% — which under-predicts tok/s and under-recommends models for
a card that is still widely sold.

This PR disambiguates the two by max memory clock, the only reliable
discriminator, and leaves 128 as the conservative default when the clock is
unknown.

Change

• data/gpu.py: GPU_MEMORY_CLOCK_VARIANTS — GTX 1650 → 192 GB/s when max
  memory clock ≥ 5500 MHz (GDDR6 ≈ 6001), else 128 (GDDR5 ≈ 4001). The 5500 split
  sits well clear of both regimes.
• gpu_db.py: resolve_detected_bandwidth(name, vram_bytes, mem_clock_mhz=None)
  consults the variant table first, then the existing curated/dbgpu path.
  Behaviour is byte-identical when mem_clock_mhz is unknown.
• nvidia.py: capture max memory clock via NVML (NVML_CLOCK_MEM) and the
  nvidia-smi fallback (clocks.max.memory). The smi query retries without the
  clock field if a driver rejects it, so an optional field can never reduce
  detection to zero GPUs.

Validation (measured, not inferred)

Clock-locked, sole-tenant, llama.cpp build 2cbfdc6, on a confirmed GDDR6
board (VBIOS 90.17.4D.00.1E, nvidia-smi clocks.max.memory = 6001 MHz):

model (Q4_K_M)	measured tok/s	est @128	est @192
Qwen3-1.7B	75.4 ± 0.12	52 (m/p 1.45)	78 (m/p 0.97)
Llama-3.2-3B	49.6 ± 0.03	35 (m/p 1.42)	52 (m/p 0.94)

At 128, whichllm under-predicts by ~45%; at 192 it lands within ~6%. The patched
detector returns ('NVIDIA GeForce GTX 1650', 192.0) on the card.

Data & reproduction. Methodology and the full four-model calibration are in
the accompanying paper (SSRN 6941538, The kv4 Trade-off Is Workload-Dependent);
the measured evidence + clock-locked harness are public and reproducible at
https://github.com/cms-pm/kv4-edge-inference (analysis/analyze.py reprints the
tok/s used here from the shipped data). This mirrors how PR #75 cited a paper for
a legacy GPU — but with measured, independently re-runnable numbers.

Tests

tests/test_gtx1650_variants.py — 15 tests: resolver disambiguation incl.
threshold boundary and back-compat (unknown clock → 128); nvidia-smi parse incl.
[N/A] clock and the 3-field-fails→2-field-retry regression guard; and a
measured-calibration check (192 estimate scales 1.5× over 128 and lands near the
measured 75.4). Full suite green.

Scope and honest limitations

• Cross-platform via NVML/nvidia-smi; only the bare-WMI fallback lacks a clock.
  detect_hardware() calls detect_nvidia_gpus() on every OS, and that path uses
  pynvml (nvml.dll) with an nvidia-smi fallback — both of which ship with the
  NVIDIA driver on Windows. So a GDDR6 GTX 1650 on a normal Windows box is
  disambiguated by the same memory-clock code as Linux. The 128 default is only
  reached in the narrow case where NVML and nvidia-smi both fail and detection
  falls through to the pure-WMI path (hardware/windows.py), which has no memory
  clock (Win32_VideoController does not expose VRAM clock). Teaching that
  fallback to shell nvidia-smi is a small, hardware-free follow-up; flagged here
  so the one remaining gap is explicit, not accidental.
• Generality. The mechanism is intentionally generic — the same name+clock
  ambiguity affects other cards, most notably the GT 1030 (DDR4 ~16 GB/s vs
  GDDR5 ~48 GB/s), a 3× error. This PR populates only the GTX 1650 (the variant
  measured here); GT 1030 and similar can be added to the same table as data
  without further code.
• The GDDR5 clock (~4001 MHz) is from NVIDIA's 8 Gbps spec, not independently
  measured here; only the GDDR6 board was measured. The 5500 threshold is safe
  given the 8-vs-12 Gbps gap regardless.

Backward compatibility

Additive only. All pre-existing tests pass; the new mem_clock_mhz parameters
default to None, preserving exact prior behaviour for every code path that does
not supply a clock.