Per-Capability Quantization Degradation on Blackwell: A Throughput–Retention Study

As quantization drops BF16 to FP8-native to AWQ-4bit on an RTX 5090 (sm_120),throughput rises, but not monotonically and not in the same order at everyconcurrency. Measured under locked clocks with all other server flags heldidentical, AWQ-4bit reaches 2.2x BF16 at concurrency 1, yet FP8-native overtakesit at concurrency 64 (3,176 vs 2,639 output tok/s) as the regime shifts fromweight-bound to compute-bound and native FP8 tensor cores beatdequantize-then-compute. On quality, paired per-passage perplexity over 82,282 held-out tokens finds FP8indistinguishable from BF16 (ratio 1.0011, 95% CI [0.9997, 1.0024]), whileAWQ-4bit costs 6.0% (ratio 1.0601, CI [1.0548, 1.0657]). The decision-relevantresult for a serving team is that FP8-native buys 44% more throughput at highconcurrency for no measurable quality cost. The study's primary hypothesis — whether the three APEX capability dimensions(factual recall, instruction following, salience) degrade non-uniformly underquantization — could not be answered on this model. Only one of three dimensionsexhibits a positional curve at full precision, and bootstrapped curve-strengthintervals overlap across all three schemes. That outcome is reported as untestedrather than rejected, and the root cause is traced to model-relative probedifficulty calibration. Methodology was pre-registered before execution. All configurations, environmentpins, and raw results are committed athttps://github.com/vshortt73/blackwell-quant-tradeoff