Native HIP graph replay is correct but provides no substantial whole-step speedup in the measured fixtures:1.020× at32³ and1.002× at128³. It reduces host submission wall time while preserving the same112-kernel macrostep. Keep this backend experimental; the current bottleneck is not solved by replacing launches alone.
The pinned hbox Manifest selects /othersys/tmp/intmat-probe-20261014/AMDGPU-c77460e860b35f21277d8f1253e37304bd8b8993 (version2.8.0). Its src/hip/graph.jl, SHA256 082d96e3a1fe2fe39de63e3f552062504c972d9b0a052310d8b12880bb75dd89, provides AMDGPU.HIP.capture, instantiate, and launch. The official graph API describes the same sequence. Capture uses the current stream and temporarily disables garbage collection. We explicitly avoid @captured: that macro can execute a warm-up body and retries after capture invalidation, which is inappropriate for a stateful biological step unless its effects are separately restored.
Capturing device_cellular_step! directly is incorrect. It increments a host attempted_steps field once, and embeds that number into lifecycle selection, mutation, daughter placement and inherited catalyst regulation. Replaying it would repeat a random stream.
The prototype DeviceGraphStepIndex borrows device metadata. Inside existing lifecycle/random kernels its multiplication reads UInt32(metadata[6])+1, before the existing last dc_complete_step! node increments the completed-step count. On successful runs this equals the ordinary attempted-step index. A latched failure suppresses all later kernels, including completion; its partial state remains unresumable. Graph capture must not change host counters or World.time; only actual launches count as attempts. A replay batch must reject completed-counter overflow, preexisting failure, replaced buffers, changed parameters/dt, or mismatched ownership. Restore allocates new buffers and therefore requires a new graph.
The graph owner must retain the process, hook, captured graph, executable graph, and all borrowed arrays for their full lifetime. Capture occurs only after warming the exact kernel specializations outside capture on a disposable/restored fixture. No compilation, synchronization, status download, array allocation, or observation belongs inside capture.
studies/device_graph_clock.jl is an isolated device-index adapter. studies/device_graph_enqueue.jl is a private copy of the existing host launch schedule, calling only shared simulation kernels. After exact private parity, production device_cellular_enqueue! now contains that unchanged schedule and accepts a scalar or device index. The ordinary API retains its current host counters, checks, launch order and arithmetic. The graph clock and HIP owner remain experimental files outside production includes.
test/device_graph_clock.jl checks random values against ordinary UInt32 indices, then compares every full snapshot over a birth/mutation trajectory and strict restore. scripts/probe_amd_graph_capture.jl is an unlaunched compatibility probe for a granted exclusive GPU window: it checks capture has no execution side effects, repeated graph launches advance the random stream, and a failure latch suppresses replay. Passing that two-node gate would establish API compatibility only.
The completed performance experiment uses a complete-loop alternating comparison from the same initialization: ordinary queued steps versus graph replay, identical timestep/substeps/capacity/physics, compilation excluded, synchronized wall time and host enqueue time recorded separately, followed by primary-array, identity, RNG and checkpoint continuation gates. A small-kernel graph result alone is insufficient evidence of faster cellular ecology.
CPU gates passed 170 exact clock/RNG checks plus 42 full-snapshot and strict-restore assertions comparing the frozen pre-extraction schedule, ordinary API, and new device-clock enqueue over 20 three-dimensional birth/mutation steps. The callable 3D constructive habitat gate also passes after extraction. scripts/benchmark_device_graph.jl stages alternating whole-loop comparisons, host enqueue time and HIP event stream intervals; native GPU results are recorded below. The pinned graph file was also byte-identical to the older depot copy; the Manifest-selected source above is the authority.
The final CPU gate passed 227 assertions, including coherent cell-buffer/ownership replacement, habitat binding replacement, failed/incomplete state preflight, replay attempt bookkeeping, counter overflow, and empty-population audit. The graph owner retains the ownership objects captured at instantiation and rejects a coherently replaced process before launching the stale graph.
Recorded results and original per-pair measurements use immutable model commit 575c14b, Radeon RX6750XT, Julia1.12.4, and the pinned AMDGPU2.8 compiler at optlevel1. HIP capture left state unchanged, five replays advanced the random stream bitwise, and a failure latch suppressed replay. Every full macrostep captured112 nodes.
Both physical extent4³ cases used512 founders, capacity8192, exchange radius.18, density128, dt.001, and16 mechanical substeps. Each trajectory ran20 complete steps, ending at1022 cells with no blocked births. All compared primary arrays, identities, contact graph, and lifecycle counters were bitwise equal between native and graph execution; graph checkpoint restore was exact.
| Grid | Eligible alternating pairs | Native20 steps | Graph20 steps | Wall speedup | Native submission wall | Graph submission wall |
|---|---|---|---|---|---|---|
| 32³ | 2 | .262232s | .257203s | 1.020× | .134691s | .000772s |
| 128³ | 3 | 2.479238s | 2.473626s | 1.002× | 1.320251s | .000814s |
Graph replay provides no substantial ecology throughput gain in these fixtures. It shortens the host submission interval, while the stream remains busy with the same work. Submission wall time includes possible driver backpressure; process CPU time was not measured. Whole-stream events do not identify individual kernel costs. The next useful measurement is an explicitly instrumented stage/kernel breakdown, followed by reducing actual support, transport or mechanical work.
The first32³ graph sample compiled its host replay-wrapper keyword specialization inside the timer. The table excludes that complete alternating pair; original records remain unchanged and the machine-readable summary flags the caveat. The standalone probe's first attempt also exposed a missing namespace import; the external retry supplied it without editing the immutable source. The current harness fixes that import and warms an actual disposable graph replay before timing; those harness changes still need their own GPU replay.
The same native window passed the new habitat scalar observation methods: owned material 518.7399999880721 exactly matched the detached audit, valid state was finite, injected NaN expression/catalyst bits were detected, and named observations were detached. These checks validate the observation boundary, not ecological throughput. The GPU was released immediately after the bounded process completed.