Dutch chip startup Axelera AI has launched a platform that combines its Metis AI Processing Unit (AIPU) chip, RISC-V controller and Voyager SDK software stack.
The Metis platform is available as embedded cards, boards and vision systems to accelerate computer vision processing at the edge.
The Metis AIPU provides peak performance of 214 TOPs with an energy efficiency of 15 TOPs/W at INT8, thanks to a novel Digital In-Memory Computing (D-IMC) engine. FP32 iso-accuracy is achieved without the need to retrain the neural network models.
Axelera’s software stack, Voyager SDK, enables rapid deployment of neural networks in the field allowing adopters of Axelera technology to focus on solving the business problems.
A four processor PCIe card with up to 856 TOPs and a $149 M.2 format card will be available for selected customers in early 2023. Separately, Axelera AI is working with partners to develop system ready solutions.
- Axelera shows DIANA analog in-memory computing chip
- Axelera AI raises $27m for in-memory computing at the edge
The software stack, Voyager SDK, includes an integrated compiler, a runtime stack and powerful optimization tools that enable customers to import their own models delivering class-leading performance on the Metis AI platform. Application templates speed up development, with neural networks in Axelera’s curated Model Zoo.
The Metis AIPU is a quad core architecture. Each core can execute all layers of a standard neural network without external interactions delivering up to 53.5 TOPs of AI processing power.
The cores can either be combined to boost the throughput of a complex workload or operate independently on the same neural network to reduce latency or process concurrently different neural networks for applications featuring pipelines of neural networks.
Central to the operation of each core is Axelera’s D-IMC engine for accelerating matrix-vector multiplication operations, offering an unprecedented high energy efficiency of 15 TOPs/W at INT8 precision.
The four cores are integrated into a System-on-Chip (SoC), comprising RISC-V controller, PCIe interface, LPDDR4X controller, and a security complex connected via a high-speed Network-on-Chip (NoC).
The NoC also connects the cores to a multi-level shared memory-hierarchy of more than 52MiB of on-chip high-speed memories, while the LPDDR4 controller connects to external memory, enabling support for much larger neural networks. Finally, the PCIe interface provides a high-speed link to an external host, which will offload full neural network applications to the Metis AIPU.
The Voyager SDK provides turnkey pipelines for state-of-the-art models, which a user may customize for their application, generate optimized code for, and deploy to Metis-enabled edge devices with one-click simplicity.
The platform supports frameworks for a wide range of environments, ranging from fully embedded use cases to distributed processing of multiple 4K video streams across networked devices without expertise in deep learning.
Via the Axelera AI Model Zoo, developers can access Neural Networks to customize, fine-tune, deploy as-is, or import their own pre-trained neural networks. The Voyager SDK automatically quantizes and compiles neural networks that have been trained on different frameworks, generating code which runs on the Metis AI platform. Optimized networks running on Metis AIPU are indistinguishable from those running on systems with floating-point units.
“The Metis AIPU not only delivers record performance and energy efficiency but also achieves FP32 iso-accuracy without retraining. It demonstrates the effectiveness of Axelera’s D-IMC engine by achieving matrix-vector multiplications with high-precision and constant time-complexity, which in tandem with the advanced quantization algorithms, the large on-chip SRAM and the RISC-V technology are well-suited to address the stringent operational requirements of the AI applications at the Edge,” said Axelera AI co-founder and CTO Evangelos Eleftheriou.
“Axelera’s end-to-end integrated software stack, built on open, industry-standard frameworks and APIs, addresses the user pain points of AI application development at the Edge”.
“The Metis AI Platform represents an inflection point for businesses looking to deploy computer vision applications at the Edge,” said Paul Neil, VP Product Management at Axelera AI. “Tightly integrated hardware and software capability, a curated model zoo and customizable application templates enable customers to accelerate through the evaluation and development cycle and shorten time-to-revenue.”
Axelera AI has also announced an Early Access Programme to help companies co-design real-world Edge AI systems using Metis.
“We’re on a mission to democratize access to AI, starting with the Early Access Program. The program gives selected customers the advantage of leveraging our disruptive technology ahead of competitors, while providing crafted support for their specific requirements, with a set of tangible benefits including hardware co-design, dedicated software and machine learning support, and co-marketing initiatives,” said Flavio Devidé, Axelera AI Head of Go-To-Market.
“In the past, customers have suffered with immature Edge AI offerings, have struggled to port their software and failed to reproduce declared performance in their real applications. The Metis AI Platform has been designed from the ground up to address these issues, and the Early Access Program removes all the barriers to its adoption.”