Video summary
Qwen3.6-27B on M5 Max: 4-bit vs 8-bit vs BF16 Benchmarked
Main summary
Key takeaways
Video Summary: Qwen3.6-27B on Mac (MLX) — 4-bit vs 8-bit vs BF16 Benchmarks
Goal / Setup
- Benchmarks Qwen3.6 27B (latest Alibaba release) running on Mac using MLX.
- Compares quantization levels: 4-bit, 8-bit, and BF16.
- References an inference engine called OMLX, including a mention of a prior tutorial for it.
- Uses a “server” UI panel to configure and run benchmarking, including:
- model selection
- prompt evaluation
- context length
- batching settings
Mac Hardware
- The video references a specific Mac model (“Here is the Mac spec I’m using”), but the actual specs are not shown in the subtitles.
Benchmark Tool / How Testing Is Done (UI Features)
In the benchmarking interface:
- Select model variant
- e.g., 4-bit / 8-bit / BF16
- Single request mode
- choose prompt / prompt evaluating
- adjust context length (examples include 15k and higher)
- Continuous batching mode
- runs parallel sessions by increasing batch size (e.g., 2x, 4x)
- Tracks benchmark metrics including:
- GPU utilization (often reaches ~100%)
- Unified RAM usage
- Tokens per second (tokens/s) split into:
- Prompt evaluation (processing the prompt)
- Text generation
Results: 4-bit Quantized
Model Size / Memory
- Uses about ~26.7 GB unified RAM.
Performance Behavior
- GPU utilization reaches about ~100%.
- Increasing context length reduces speed:
- At 16,000 context length, generation drops to ~17.5 tokens/sec
- Prompt evaluation drops to ~861.9 tokens/sec (as stated)
Continuous Batching (Parallel Sessions)
- 2 sessions: ~23.1 tokens/sec
- 4 sessions: ~40 tokens/sec
- Key takeaway: throughput improves as batch size / parallelism increases.
Results: 8-bit Quantized
Model Size / Memory
- Uses about ~33.9 GB unified RAM (roughly 10 GB more than 4-bit).
Performance Behavior
- GPU utilization again reaches about ~100%.
- Slower than 4-bit.
- Reported generation speed while varying context ranges roughly:
- ~15 → 13.7 tok/s
- (trend varies slightly across tested lengths)
Continuous Batching
- Shows the same general trend: higher batch size improves speed.
Results: BF16 (Highest Precision)
Model Size / Memory
- Model footprint is about ~53.5 GB.
- Hardware requirement mentioned:
- Mac should have at least 64 GB unified RAM
Performance Behavior
- GPU utilization is about ~99–100%.
- Significantly slower than 4-bit and 8-bit.
- Generation speed decreases as context increases:
- ~9.5 → 9.4 → 8.9 → 8.7 tok/s
Continuous Batching
- Still benefits from increased batch size (speedups observed, consistent with other variants).
Overall Conclusions (Implied by the Benchmarks)
- 4-bit is the fastest in these tests, with the best generation throughput and strong scaling under continuous batching.
- 8-bit requires more RAM than 4-bit, with slightly lower throughput.
- BF16 needs the most memory and is the slowest, though it still improves with continuous batching.
- Across all variants:
- GPU utilization tends to saturate near maximum
- Longer context reduces generation speed
- Higher parallel/batch size improves throughput
Main Speakers / Sources
- Video narrator / channel host (name not given in the subtitles)
- Mentions OMLX and MLX on Mac (platform/tools referenced)