Skip to content

Latest commit

 

History

History
61 lines (33 loc) · 3.79 KB

File metadata and controls

61 lines (33 loc) · 3.79 KB

Experiment Log

Question

Can remote KV placement make a synthetic long-context decode configuration feasible, and under what modeled conditions does near-memory partial-state attention beat returning cold KV over the link?

All four bundled profiles explicitly declare hardware_profile: synthetic and measurement_scope: none. Their knobs are not inferred from Apple M4 data or vendor specifications.

H1 — HBM-only capacity

Input: 7B FP16 weights, FP16 GQA KV, 8,192 prompt tokens, 64 generated tokens, 16 synchronized sequences, 24 GiB HBM, 2 GiB reserve, 16-token KV pages.

Test: page-allocate the post-final-write KV at length 8,256 and admit reserve, weights, and hot KV.

Result: rejected. The runtime reserve plus weights and full page-allocated KV exceed HBM.

Decision: any performance number for this placement would be misleading, so the result remains an explicit capacity failure.

H2 — Remote sliding window

Change: keep only 1,024 tokens per sequence in HBM; place older pages in the remote tier.

Result: feasible. Across decode, the cold scan reads about 900.88 GiB from remote media and about 945.92 GiB from the interconnect after the 5% protocol factor. The modeled link is the most frequent bottleneck.

Decision: capacity relief does not imply adequate latency; media and link service must be evaluated separately.

H3 — Near-memory partial-state attention

Change: preserve the same placement and media pages, but send per-layer queries to remote compute and return (output, row_max, row_sum) per query head.

Result: feasible. Remote-media reads remain about 900.88 GiB, while remote-to-accelerator partial-state link reads are about 0.53 GiB. With the synthetic 12 TOPS peak / 60% efficiency knob, remote media becomes limiting and mean decode latency is lower than ordinary tiering.

Decision: near-memory reduces link movement, not remote-media scanning. The win is conditional on compute, media, link, fixed latency, overlap, page size, and precision.

H4 — Counterexample

Change: reduce only peak near-memory throughput to 0.1 TOPS in the committed stress scenario. Workload, pages, capacities, link/media rates, efficiencies other than near-memory compute, and placement remain fixed.

Result: feasible but slower than ordinary tiering; near-memory compute is the bottleneck.

Decision: reject the universal claim “near-memory always wins.” The analysis also computes the conditional break-even and constructs a deterministic point below it.

Sensitivity and break-even

The generated analysis evaluates four peak link bandwidths, seven peak near-memory throughputs, twelve one-at-a-time input axes, and five declared multipliers. A monotone geometric bisection locates the peak near-memory TOPS where the two policies tie. The exact thresholds are regenerated in evidence/benchmark-summary.md.

This sweep is not sampling and its min/max is not a confidence interval. It answers “what does this equation do under these declared perturbations?” rather than “how likely is this outcome on hardware?”

Separate MPS equation checks

Committed Apple M4 MPS artifacts contain median/p95 aggregate summaries, not timing iterations. Held-out sizes/contexts check transfer and one-layer attention equation shapes. They do not exercise the scenario's HBM reserve, page allocator, CXL/remote paths, PIM, multi-layer decode, CUDA, energy, or end-to-end serving. No MPS fitted value is used above.

Reproduction

make verify
.venv/bin/memoryflow simulate scenarios/7b-long-context-tiered.json --steps
.venv/bin/memoryflow analyze scenarios/7b-long-context-tiered.json

make verify confirms exact input/source hashes, deterministic generated files, schema boundaries, static analysis, and test coverage.