MI325X: FP8 vs BF16 Precision Comparison
How FP8 and BF16 precision affect Qwen 3.5 397B-A17B inference on MI325X (AMD CDNA 3). Throughput, latency, and cost across LLM workloads. Use the chart controls below to switch sequences and metrics — same interactions as the main inference chart.
Near the low end of the 41–72 tok/s/user interactivity band, at 49 tok/s/user on Qwen 3.5 397B-A17B (MI325X): FP8 runs 347 tok/s/chip at $0.88/M tokens, BF16 runs 462 at $0.66/M. BF16 is 33% cheaper per token; BF16 delivers 33% more tok/s/chip. Precision changes affect both inference speed and model quality — consult the evaluation tab for accuracy benchmarks.
At 57 tok/s/user on Qwen 3.5 397B-A17B (MI325X), FP8 delivers 211 tok/s/chip at $1.45 per million tokens; BF16 delivers 382 tok/s/chip at $0.80. BF16 is 81% cheaper per token; BF16 delivers 81% more tok/s/chip. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact.
FP8 posts 124 tok/s/chip for $2.46 per million tokens at 65 tok/s/user on Qwen 3.5 397B-A17B (MI325X); BF16 posts 307 tok/s/chip for $1.00. BF16 is 147% cheaper per token; BF16 delivers 147% more tok/s/chip. Quantization-level accuracy differences are tracked on the evaluation tab. (Numbers reflect the default 1k/1k selection for this URL — table and chart below update if you change sequence or model in the controls. Each side uses the best available serving configuration for that precision, which may include speculative decoding such as MTP where recipes exist — the same convention as the other comparison pages.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | FP8:346.6BF16:461.5 | FP8:211.0BF16:382.0 | FP8:124.3BF16:306.9 |
| Cost ($/M tok) | FP8:$0.882BF16:$0.662 | FP8:$1.448BF16:$0.800 | FP8:$2.458BF16:$0.996 |
| tok/s/MW | FP8:205060BF16:273085 | FP8:124837BF16:226025 | FP8:73546BF16:181594 |
| Concurrency | FP8:~30BF16:~41 | FP8:~16BF16:~30 | FP8:~8BF16:~20 |
Inference Performance
Agentic inference metrics from the AgentX scenario and fixed-sequence inference metrics across models, hardware configurations, and serving parameters.