Loading benchmark filters…
Build with ZETIC
Share your model, hardware, and goals for latency, energy per inference, and memory.
We’ll agree on scope, then benchmark, optimize, and validate against your baseline and quality targets.
6 models · 1 board · 12 ZETIC comparisons in view
Selected baseline and ZETIC build on the same board, per request, for each scenario. Test conditions in Details.
Latency · p50Baseline / ZETIC · speedup on the right
| Model | Board | Latency | Improvement | Quality result | Details |
|---|---|---|---|---|---|
| AGX OrinSingle question |
per request · p50 | 3.1×faster | Agreement with original
Answer accuracy −0.7 pp vs original · DecisionBench 1,000 Answer accuracy −0.7 pp vs original · Robot frames, Scene understanding Answer accuracy −0.1 pp vs original · Robot frames, Success detection Answer accuracy +0.1 pp vs original · Robot frames, Safety judgment | ||
| AGX OrinRobot frames |
per request · p50 6,342 requests per round | 1.8×faster | Agreement with original
Answer accuracy −0.7 pp vs original · DecisionBench 1,000 Answer accuracy −0.7 pp vs original · Robot frames, Scene understanding Answer accuracy −0.1 pp vs original · Robot frames, Success detection Answer accuracy +0.1 pp vs original · Robot frames, Safety judgment | ||
| AGX OrinSingle question |
per request · p50 | 4.9×faster | Agreement with original
Answer accuracy −0.1 pp vs original · DecisionBench 1,000 Answer accuracy +0.1 pp vs original · Robot frames, Scene understanding Answer accuracy −0.3 pp vs original · Robot frames, Success detection Answer accuracy +0.3 pp vs original · Robot frames, Safety judgment | ||
| AGX OrinRobot frames |
per request · p50 6,038 requests per round | 2.1×faster | Agreement with original
Answer accuracy −0.1 pp vs original · DecisionBench 1,000 Answer accuracy +0.1 pp vs original · Robot frames, Scene understanding Answer accuracy −0.3 pp vs original · Robot frames, Success detection Answer accuracy +0.3 pp vs original · Robot frames, Safety judgment | ||
| AGX OrinSingle question |
per request · p50 | 7.5×faster | Agreement with original
Answer accuracy +0.5 pp vs original · DecisionBench 1,000 Answer accuracy −0.6 pp vs original · Robot frames, Scene understanding Answer accuracy +0.2 pp vs original · Robot frames, Success detection Answer accuracy −0.2 pp vs original · Robot frames, Safety judgment | ||
| AGX OrinRobot frames |
per request · p50 6,147 requests per round | 1.5×faster | Agreement with original
Answer accuracy +0.5 pp vs original · DecisionBench 1,000 Answer accuracy −0.6 pp vs original · Robot frames, Scene understanding Answer accuracy +0.2 pp vs original · Robot frames, Success detection Answer accuracy −0.2 pp vs original · Robot frames, Safety judgment | ||
| AGX OrinSingle question |
per request · p50 | 2.5×faster | Agreement with original
Answer accuracy +0.2 pp vs original · DecisionBench 1,000 Answer accuracy +1.1 pp vs original · Robot frames, Scene understanding Answer accuracy −0.8 pp vs original · Robot frames, Success detection Answer accuracy 0.0 pp vs original · Robot frames, Safety judgment | ||
| AGX OrinRobot frames |
per request · p50 2,028 requests per round | 1.6×faster | Agreement with original
Answer accuracy +0.2 pp vs original · DecisionBench 1,000 Answer accuracy +1.1 pp vs original · Robot frames, Scene understanding Answer accuracy −0.8 pp vs original · Robot frames, Success detection Answer accuracy 0.0 pp vs original · Robot frames, Safety judgment | ||
| AGX OrinSingle question |
per request · p50 | 2.4×faster | Agreement with original
Answer accuracy −0.1 pp vs original · DecisionBench 1,000 Answer accuracy −1.1 pp vs original · Robot frames, Scene understanding Answer accuracy 0.0 pp vs original · Robot frames, Success detection Answer accuracy −1.2 pp vs original · Robot frames, Safety judgment | ||
| AGX OrinRobot frames |
per request · p50 2,028 requests per round | 1.5×faster | Agreement with original
Answer accuracy −0.1 pp vs original · DecisionBench 1,000 Answer accuracy −1.1 pp vs original · Robot frames, Scene understanding Answer accuracy 0.0 pp vs original · Robot frames, Success detection Answer accuracy −1.2 pp vs original · Robot frames, Safety judgment | ||
| AGX OrinSingle question |
per request · p50 | 2.1×faster | Agreement with original
Answer accuracy +0.1 pp vs original · DecisionBench 1,000 Answer accuracy +0.6 pp vs original · Robot frames, Scene understanding Answer accuracy −0.5 pp vs original · Robot frames, Success detection Answer accuracy −1.5 pp vs original · Robot frames, Safety judgment | ||
| AGX OrinRobot frames |
per request · p50 1,990 requests per round | 1.9×faster | Agreement with original
Answer accuracy +0.1 pp vs original · DecisionBench 1,000 Answer accuracy +0.6 pp vs original · Robot frames, Scene understanding Answer accuracy −0.5 pp vs original · Robot frames, Success detection Answer accuracy −1.5 pp vs original · Robot frames, Safety judgment |
Method · Latency is p50 per request, from the raw input to the option probabilities. Agreement with original is the share of questions where the ZETIC build picks the same answer as the original model; answer accuracy is scored against gold answers, shown as ZETIC minus original in percentage points. Both belong to the build, so each scenario on a board shows the same values. How decision models are measured
ZETIC optimization overview
View all resultsLatency per request on the same model and board.
7.5×
faster
4.9×
faster
3.1×
faster
2.5×
faster
2.4×
faster
2.1×
faster
2.1×
faster
1.9×
faster
1.8×
faster
1.6×
faster
1.5×
faster
1.5×
faster
