Hardware Rig Config
● Highlighted sizes are standard factory tiers for this chip; any capacity can be simulated.
macOS caps GPU Metal allocation (~67–75% default). Override iogpu.wired_limit_mb to unlock 90% VRAM.
sudo sysctl iogpu.wired_limit_mb=29491
LLM & Quantization
SMOOTH — Faster than you can read.
VRAM & Memory Footprint Telemetry
CPU / System RAM Offload Pain Meter — Hybrid Split Active
MAC WIRED MEMORY UNLOCK CAN RESCUE THIS MODEL!
Exceeds macOS's default Metal cap, but fits inside 90% Unified Memory.
Quantization Compatibility Matrix (Qwen 32B)
| QUANT | BPW | TOTAL RAM | MODE | EST. SPEED | VERDICT |
|---|
Flex or Roast My Rig — Benchmark Permalink
Permalink syncs your exact rig, model, quantization & context window.
Weights are modeled as (Params_B × BPW) / 8 × 1.05 (5% GGUF header/embedding overhead). KV cache uses each model's GQA/MLA architecture (2 × layers × kv_heads × head_dim) scaled by selected KV precision (FP16 / Q8_0 / Q4_0) plus 0.8 GB compute buffer.
Batch-1 token generation is memory-bandwidth bound (~0.72 sustained bus efficiency). Every token streams active weights + active KV cache. For MoE models (Qwen3-30B-A3B, Llama-4 Scout, DeepSeek 671B), all weights reside in RAM while only active experts stream per token.
Apple Silicon shares Unified Memory between CPU & Metal GPU (default Metal wired cap is ~67–75%, unlockable to 90% via sysctl iogpu.wired_limit_mb). PC rigs that exceed dedicated VRAM split layers into DDR4/DDR5 RAM, where CPU DRAM bandwidth bottlenecks throughput.