Real artificial intelligence requires the capability to perceive the environment and make decisions based on observations. Over the past few decades, perception tasks such as image/voice recognition have made significant progress thanks to deep learning models. However, Probabilistic Graphical Models (PMs), which can enhance machine learning with reasoning and decision-making abilities, require more attention.
To support more general PMs, sampling-based Markov Chain Monte Carlo (MCMC) algorithms are widely used to address the intractable problems. Unfortunately, MCMC is compute-intensive and challenging to parallelize, leading to inefficient execution on modern CPU/GPU platforms. Thus, we propose AIA, an Approximate Inference Accelerator, designed to accelerate inference of PMs at the edge.
AIA comprises a RISC-V host and a 2D mesh of 16 customized RISC-V cores optimized for efficient PM inference. Each core features: (i) a novel non-normalized Knuth-Yao sampler and interpolation unit, and (ii) core-to-core direct data access via the register file, which addresses compute-intensive operations. To fully exploit the parallel potential of MCMC algorithms, a customized compiler chain has been developed for effective spatial mapping and scheduling on the chip. AIA can generate 1277 MSamples/s at 0.9V and 20 GSamples/s/W at 0.7V, which is up to 2x faster and 1.45x more energy efficient compared to the previous state-of-the-art Markov Random Field (MRF) accelerator. We further demonstrate the flexibility of our design by mapping Bayesian Networks benchmarks onto AIA.
28/6/2024 11:00 - 12:00
ESAT Aula L