Celera

About Us

While designing new processors to run ML models can lead to improved performance, making effective usage of processor features in a clean software toolchain can become a challenging endeavor. Celera has put together a team of compiler experts and software engineers who know how to exploit the best performance of your hardware and how to organize its toolchain to improve productivity. With Celera, your models run faster and more organized.

Our Technology

Optimizing program performance is a complex task, particularly for ML models, which can have large memory footprints. From list scheduling algorithms to double-buffering methods to hide the latency of tensor transfers, our engineers know all techniques needed to optimize your code.

We have designed LLVM-based compilers for various architectures, from regular x86 multicores to specialized Neuromorphic architectures. We know how to make the best usage of dedicated MAC units, sophisticated ISAs, and complex DMA engines.

One of the biggest challenges in designing ML models is empowering developers with a seamless toolchain integrated with industry standards like Tensorflow, JAX/XLA, Pytorch, Glow, and ONNX. Our team has already integrated solutions to all these toolchains.

Some of our cases

LG XLA Optimizer

LG Electronics Neuromorphic processor is a 32-core architecture where each core is composed of a RISCV extended with new instructions dedicated to ML operations. Celera used TensorFlow/XLA and our convolution slicing optimization algorithm to improve LG’s model performance.

Working with such a competent and high-quality team has been a great pleasure, and I would like to thank everyone involved for the successful cooperation.
(Michael Frank, Roadmap Architect, LG Silicon Valley Lab, 2019)

SilicoNeuro Glow Compiler

Celera has ported the Glow Compiler toolchain to the SilicoNeuro NMP processor. Quantization and convolution slicing optimization techniques were employed in the design to create a smooth path from ONNX to Glow and LLVM code generation.

Celera's team delivered an optimized Glow compiler for the NMP processor with professionalism and within the project schedule.
(Chang Soo Kim, CEO, AiM Future, 2022)

ETRI NEST Toolchain

ETRI has developed an in-house toolchain that uses partitioning and parallelism algorithms to improve the performance of ML models. Celera has integrated its convolution slicing and optimizing compiling technology into NEST, enabling improved performance in one of NEST’s neuromorphic acceleration engines.