Performance

Honest first-party timing for the research-bar export path. No peer comparisons. No stale “Nx faster” marketing numbers.

What we measure

Wall time for export_sma_sweep on synthetic OHLC with device="auto" vs device="cpu_numba" (and explicit metal when useful).

• Primary job: fee-aware next-bar grids on a local Mac
• As of v0.17.5, research auto prefers cpu_numba for wall clock on typical grids (sets fallback_reason=auto_prefer_cpu_numba when Metal would otherwise be eligible). Explicit device="metal" still uses the size-gated Metal path.
• Times below are after a short Numba warmup (first call pays JIT)

Reproduce

# from a clone with editable / installed monte-neo
uv run python scripts/bench_research_bar.py --bars 100000 --combos 64
uv run python scripts/bench_research_bar.py --bars 100000 --combos 64 --run-1m

Script: scripts/bench_research_bar.py.

Measured numbers (2026-09-17) — MacBook Pro M1 Pro 16 GB

Host	Notes
MacBook Pro M1 Pro 16 GB (macOS arm64, Python 3.13.14, repo .venv)	Primary reference — Metal available
Linux executor (x86_64, ~16 GB class)	Cross-check only — Metal unavailable → auto ≡ cpu_numba

Command:

.venv/bin/python scripts/bench_research_bar.py --bars 100000 --combos 64 --run-1m

Primary table (M1 Pro 16 GB, post-warmup)

Before v0.17.5 (auto selected Metal when size-ok)

Bars	Combos	device requested	Device used	fallback_reason	Wall (s)	Notes
100 000	64	auto	metal	none	0.4844	auto selected Metal
100 000	64	cpu_numba	cpu_numba	none	0.0074	much faster wall clock on this grid
1 000 000	32	auto	metal	none	3.3941	auto selected Metal
1 000 000	32	cpu_numba	cpu_numba	none	0.0597	much faster wall clock on this grid

After v0.17.5 (auto → cpu_numba for wall clock)

Bars	Combos	device requested	Device used	fallback_reason	Wall (s)	Notes
100 000	64	auto	cpu_numba	auto_prefer_cpu_numba	~0.007	matches Numba row
100 000	64	cpu_numba	cpu_numba	none	~0.007	unchanged
100 000	64	metal	metal	none	~0.48	explicit Metal still allowed (may be slow)
1 000 000	32	auto	cpu_numba	auto_prefer_cpu_numba	~0.06	matches Numba row
1 000 000	32	metal	metal	none	~3.39	explicit Metal still allowed

Reading: On these grid sizes on M1 Pro, Numba wins wall clock. Research auto now picks the faster safe path. Use device="metal" only when you want the Metal economics path regardless of wall clock. Restore pre-0.17.5 prefer-Metal auto with MONTE_NEO_RESEARCH_AUTO_PREFER_METAL=1.

Secondary note (Linux executor, no Metal)

Earlier same-day run on a Linux host without Metal (auto ≡ cpu_numba): ~0.016–0.117 s post-warmup for the same 100k/64 and 1M/32 shapes. Useful only as a portable Numba cross-check — not an Apple Silicon claim.

How to read these numbers

• Research bar only — not an OMS event-loop claim, not live multi-venue throughput
• Wall clock includes signal build when timing.includes_signal_build is true (export schema)
• Cold start (first Numba compile) is excluded from the table; your first run will be slower
• Combos / bars change wall time roughly with work size — use the script, do not extrapolate marketing multipliers
• On oversized Metal jobs, expect fallback to cpu_numba with an explicit fallback_reason (see FAQ)
• Research auto speed decline uses auto_prefer_cpu_numba (distinct from size/budget gates)

Acceleration notes (scope language)

1. Numba — default portable research-bar path; research auto (v0.17.5+) prefers it for wall clock on typical grids
2. Metal economics — Apple Silicon path when explicitly requested (device="metal") or when auto is opted back into prefer-Metal via env; still gated for 16 GB-class hosts
3. Memory planner — plan_research_bytes soft budgets / tile hints
4. No-hang gate — oversized Metal jobs fall back instead of unbounded GPU wait
5. Research auto wall-clock policy — decide_research_accelerator may set auto_prefer_cpu_numba so batch/export inherit Numba without changing OMS resolve_device

Older docs that quoted “125x” / “300k ops/sec” / “v0.0.4 targeting” without a reproducible method are retired. Prefer this page + scripts/bench_research_bar.py.