Thanks tls,
A few months ago, the big boys announced that they were going to provide a set of AI/NN standards, even though they are demonstrably unqualified to do so. It seems the same has been happening with benchmarking AI/NNs, the essential tool for comparing performance of NNs, so brainChip is setting the record straight.
[I recall the porcine aviator or someone else over at the other place bleating (oinking) about the lack of benchmarks]
The following graphs from the white paper compare 2 nodes of Akida with Silicon labs, Reneas, and STM MCUs, while those further down show comparisons with Nvidia Jetson nano and Google Coral.
Then there are graphs for Performance (Speed)/Efficiency (energy) which BrainChip proposes as being a more relevant benchmark for edge devices.
The white paper then goes on to compare the times to load the model libraries which is important for NNs running multiple networks.
T
he benchmark data below is divided into two sets of graphs, each comparing BrainChip’s Akida processor using three representative applications. These applications – keyword spotting, object detection, and anomaly detection – play critical roles in enabling the edge-specific verticals highlighted in the previous chapter. The first set of graphs are based on publicly available MLPerf datasets from MLCommons that benchmark Akida and leading MCU vendor offerings.* As the first chart illustrates, Akida-based performance delivers extremely low latency using just a fraction of the energy consumed by conventional MCUs.
View attachment 27180
T
he contrast is even more pronounced when benchmarking Akida against conventional MCUs for visual wake words – a popular edge application that demands complementary levels of efficiency and performance in power-conscious and thermally constrained environments.
View attachment 27181
S
imilarly, the anomaly detection benchmark shows a substantial gain in latency and energy versus the MCU solutions.
View attachment 27182
T
he second set of graphs are also based on publicly available MLPerf datasets from MLCommons that benchmark Akida and leading DLA and TPU vendor offerings. These compare Akida with two higher end edge AI processors: NVIDIA Jetson Nano (a deep learning accelerator) and Google Coral (a tensor processing unit). Even at a significantly lower frequency, Akida matches or outperforms both processors while consuming significantly less power.
This performance and efficiency data is based on a two-node configuration – which we compare here to ensure consistency with earlier MCU comparisons. It should be noted that higher node configurations further reduce latency – and are potentially more efficient as they enable faster compute. However, even with a two-node configuration, Akida’s latency fits well within the 5ms target, and the energy consumed per frame by Akida is only a fraction of the power drawn by larger processors.
View attachment 27183
View attachment 27185
View attachment 27186
T
he relative efficiency of the anomaly detection benchmark highlights the benefits of an efficient AI engine over traditional MCUs even for small workloads.
View attachment 27187