Thesis just out of KTH.
Haven't downloaded to read it but just the abstract was enough for me on the confirmation of how just our AKD1000 First Gen stacked up and clearly showed its "edge" in the edge space over NVIDIA GPU.
Think it's a given that as complexity increases the processing power shifts however the power usage is still an advantage.
Be interesting to see how it goes when they focus on fully customised models.
Would also be great to see what step up AKD1500 / Gen 2 / TENNs etc is capable of in a comparison given this was just Gen 1.
kth.diva-portal.org
Comparison of Akida Neuromorphic Processor and NVIDIA Graphics Processor Unit for Spiking Neural Networks
Chemnitz, Carl
KTH, School of Electrical Engineering and Computer Science (EECS).
Ermis, Malik
KTH, School of Electrical Engineering and Computer Science (EECS).
2025 (English)Independent thesis Basic level (degree of Bachelor), 10 credits / 15 HE creditsStudent thesisAlternative title
Jämförelse av neuromorfisk processor Akida och NVIDIA grafikkort för Spiking Neural Networks (Swedish)
Abstract [en]
This thesis investigates the latency, throughput and energy efficiency of the BrainChip Akida AKD1000 neuromorphic processor compared to a NVIDIA GeForce GTX 1080 when running two different spiking neural network models on both hardwares. Spiking neural networks is a subset of neural networks that are specialized for neuromorphic processor. The first model is a simple image classification model (GXNOR on MNIST), and the second is a more complex object detection model (YOLOv2 on Pascal VOC). The models were trained and quantized to 2-bit and 4-bit weight precision, respectively,
enabling spiking execution both on Akida AKD1000 and on GTX 1080, for the GPU CUDA was used. Results show that Akida achieved significant reductions in energy consumption and clock cycles for both models, consistent with prior findings within the field. Specifically,
for the simple classification model the AKD1000 achieved 99.5 % energy reduction with 76.7 % faster inference times, despite having a clock rate 91.5 % slower than the GPU. However,
for the more complex object detection model, the Akida took 118.1 % longer per inference, while reducing the energy expenditure by 96.0 %. For the MNIST model the AKD1000 showed no correlation in both cycles & time and cycle & energy. While for the YOLOv2 model it had a 0.2 correlation for both previous mentioned ratios.
Suggesting that as model complexity increases, the Akida’s behaviour converges toward the GPU’s linear correlation patterns. In conclusion,
the AKD1000 processor demonstrates clear advantages for low-power, edge-oriented applications where latency and efficiency are critical. However, these benefits diminish with increasing model complexity, where GPUs maintain superior scalability and performance. Due to limited documentation of the chosen models, a 1-to-1 comparison was not possible.
Future work should focus on fully customized models to further explore the dynamics.
thestockexchange.com.au
www.linkedin.com
so let’s hope our hopes will come true! Hopefully
