Benchmarks

This section provides benchmarks for the Vollo Trees accelerator for a variety of decision tree models.

Performance figures are given for a 256-unit configuration for the V80 accelerator card. If you require a different configuration, please contact us at vollo@myrtle.ai.

All these performance numbers can be measured using the vollo-trees-sdk with the correct accelerator card by running the provided benchmark script.

V80: 256 units

Raw buffer API

This is using buffers allocated with vollo_rt_get_raw_buffer which lets the runtime skip IO copy.

modelnum treesmax depthinput featuresfully populatedmean latency (us)99th percentile latency (us)
single-decision-t1-d1-f321132No0.870.89
example-small-t256-d5-f64256564No1.001.02
example-small-t256-d5-f64-full256564Yes1.001.01
example-medium-t1024-d8-f12810248128No1.121.13
example-medium-t1024-d8-f128-full10248128Yes1.121.14
example-large-t4096-d10-f10244096101024No1.821.83

User buffers

modelnumtreesmax depthinput featuresfully populatedmean latency (us)99th percentile latency (us)
single-decision-t1-d1-f321132No0.880.89
example-small-t256-d5-f64256564No1.001.01
example-small-t256-d5-f64-full256564Yes1.001.01
example-medium-t1024-d8-f12810248128No1.121.13
example-medium-t1024-d8-f128-full10248128Yes1.121.13
example-large-t4096-d10-f10244096101024No1.831.84