CosmoEdge 1.1 Multi-Platform Video Analytics Benchmark

BM1688 · CV186X · RK3576 · RV1126B

This report publishes 49 controlled small-model cases, four completed 72-hour fixed-profile results, and formal VLM validation results for 3 platforms.

Benchmark scope: small-model capacity results come from 30-second local-loop steps, VLM performance from formal 60-second gated steps, and the separate 72-hour test covers the listed fixed-channel profiles under controlled local-loop input. See the methodology for complete conditions.

Concurrent mixed-workload matrix

Each channel contains two business tasks across three model stages: one person-detector stage plus detector and classifier stages for no-safety-helmet analysis.

PlatformWorkload per channelModel stages/chTarget FPS/taskPassing channelsBusiness-task bindings
BM1688Person detection + no-safety-helmet analysis35≥1632/32
CV186XPerson detection + no-safety-helmet analysis35≥816/16
RK3576Person detection + no-safety-helmet analysis35≥816/16
RV1126BPerson detection + no-safety-helmet analysis35≥48/8

Single-task capacity matrix

Values are last-passing channels; ≥ means the configured maximum still passed, * means the next binding was blocked, and † means expansion was blocked before measurement by storage preconditions.

PlatformWorkload24 FPS10 FPS7 FPS5 FPS
BM1688Person detector≥8≥16≥16≥16
BM1688No-safety-helmet analysis≥7*≥14*≥16≥16
CV186XPerson detector≥8*≥15*≥16≥16
CV186XNo-safety-helmet analysis6≥13*≥16≥16
RK3576Person detector612≥16≥8†
RK3576No-safety-helmet analysis61012≥16
RV1126BPerson detector2≥4≥4≥4
RV1126BNo-safety-helmet analysis2≥4≥4≥4

72-hour controlled long-run observation

All four platforms completed the configured dual-CV local-loop workload for 72 hours at 5 FPS per task.

PlatformConfigured channelsTask bindingsTarget FPS/taskDurationObserved / expected samplesCoverageMinimum / average / maximum FPSMaximum sampling gapDisk first → lastResult
BM1688816572 h4316 / 432099.91%4.68 / 5.086 / 5.4960.058 s96% → 96%PASS
CV186X816572 h4316 / 432099.91%4.54 / 5.085 / 5.2960.061 s96% → 96%PASS
RK3576816572 h4316 / 432099.91%5 / 5.098 / 5.1760.058 s14% → 15%PASS
RV1126B48572 h4316 / 432099.91%4.85 / 5.23 / 5.3760.067 s46% → 47%PASS

Open the 72-hour multi-platform report · canonical data

Validated VLM performance

All three platforms use one protocol, with the 80% throughput threshold participating directly in runtime PASS/FAIL.

PlatformTarget FPS/chExecuted gateLast passing channelsFirst failure
BM16880.180%67 channels: minimum active-route FPS ratio 69.5%, below the 80% gate
CV186X0.180%67 channels: minimum active-route FPS ratio 69.2%, below the 80% gate
RK35760.180%45 channels: minimum active-route FPS ratio 69.7%, below the 80% gate
RV1126B———Outside this VLM validation

BM1688, CV186X, and RK3576 also pass model load, task creation, valid inference, event/alarm output, and task recovery after service restart, each on the same fixed candidate package as its capacity run. CosmoEdge 1.1 supports VLM on all three platforms within the recorded package, model, and protocol scope.

Evidence entry points

PlatformOverviewCasesSingle-taskMixed workload72-hourVLM
BM1688Platform overview9 casesSingle-taskMixed workload72-hour long runVLM
CV186XPlatform overview16 casesSingle-taskMixed workload72-hour long runVLM
RK3576Platform overview15 casesSingle-taskMixed workload72-hour long runVLM
RV1126BPlatform overview9 casesSingle-taskMixed workload72-hour long run—

Test environment

PlatformDeviceOSRuntime / mediaCosmoEdge
BM1688BM1688 reference deviceUbuntu 22.04.5 LTSlibsophon/BMRT 0.4.12; Sophon FFmpeg 2.0.0 / Sophon GStreamer 2.0.0V1.1.0.0
CV186XCV186X reference deviceUbuntu 22.04.5 LTSlibsophon/BMRT 0.4.12; Sophon FFmpeg 2.0.0 / Sophon GStreamer 2.0.0V1.1.0.0
RK3576RK3576 EVBLinuxRKNN; Rockchip MPP / RGAV1.1.0.0
RV1126BRV1126B reference deviceLinuxRKNN; Rockchip MPP / RGAV1.1.0.0

Method and reproduction