RTX Spark 18C vs M5 Pro/Max Benchmarks: 32% Faster Prompt Processing, But Apple Wins on Decode
Temporary-Size7310 · reddit · 2026-11-23
Hands-on benchmarks of Qwen3.8 27B (NVFP4/INT4) on RTX Spark 18C vs Apple M5 Pro/Max:
- Prompt processing: Spark dominates and stays consistent at large context (1,098 tok/s at 2k-4k tokens on TRT-LLM-EDGE, 32% faster than M5 Max; 888 tok/s at 60k tokens — about 2x the M5 Max)
- Decode: bandwidth-bound, on par with llama.cpp and likely Apple-favored (no MLX data points yet); on text generation the M5 Max with speculative decoding matches Spark (31.1 vs 31.7 tok/s), without it both Macs lead (18.2/16.5 vs 15.7 tok/s)
- TRT-LLM-EDGE is more power-efficient than llama.cpp; lowest PSU mode keeps tg stable but lowers pp
- Gaming: roughly 4070 mobile / 5060 mobile / 3060 desktop level, 10-15% below 5070 mobile (3DMark linked)
- eGPU appears feasible under certain conditions (Level1Techs forum case)
The author welcomes community MLX measurements across the whole pipeline.
More from Infra
- Bittensor's Subnet 92 claims a 'frontier JEV model' 200x faster than LLMs — markjeffrey · 2026-11-23
- Vercel CEO Rauch: 'free' ZDR stealth models look like industrial-scale data exfiltration — AccBalanced · 2026-11-23
- Lenovo's enterprise AI server sales surge 200%, outpacing cloud giants' 170% — shashib · 2026-11-23
- Australia's Spy Agency Now Treats Data Center Ownership as Foreign Interference Risk — rohanpaul_ai · 2026-11-23
- RustingStrata 1.3.0: a Rust GGUF engine that runs models bigger than your VRAM — Serious_Scarcity7818 · 2026-11-23
- Apple's appenz: agent swarms are filling in NVIDIA's CUDA moat — appenz · 2026-11-23