Myrtle Software today announced that it has joined the expanded Intel® FPGA Design Solutions Network (DSN) programme to provide deep learning accelerator support for Intel’s latest hardware and software platform solution for data centres.
Intel’s comprehensive new hardware and software platform solution for workload acceleration based on FPGAs, including the first in a family of Programmable Acceleration Cards, the Intel® Programmable Acceleration Card (Intel PAC) with Intel® Arria® 10 GX FPGA, is designed to address a wide range of demanding applications in cloud and data centre services. The new Intel card combines reconfigurable silicon, and utilizes the Intel Acceleration Stack for Intel® Xeon® CPU with FPGAs to enable the acceleration of compute intensive workloads like deep learning at scale in data centres. Offloading workloads such as deep learning to an FPGA enables them to run at a much lower latency and consumes less power than a traditional CPU or GPU.
Peter Baldwin, Myrtle Software’s founder stated, “We have a fantastic team of talented engineers who have developed some unique deep learning capabilities here in Cambridge. We’ve been working with the Intel® Arria® 10 FPGA since 2016 and are excited to have the opportunity to provide a key component for Intel’s new cloud infrastructure enabling easier FPGA acceleration for data centre customers.
Myrtle’s expertise in scheduling and optimising deep learning algorithms for efficient FPGA silicon will enable cloud service providers and enterprises to quickly deploy advanced deep learning algorithms on the new hardware with a high degree of efficiency and performance.
Myrtle’s team of PhD mathematicians, computer scientists and hardware engineers have developed software to automatically realize deep-learning algorithms as efficient silicon designs, capable of very high performance and low power usage- ideal for data centre FPGA deployment. Myrtle offers a unique approach that produces bespoke designs for each specific network, without the need for FPGA hardware expertise, and avoids the inefficiencies of existing methods. The technology is also highly scalable in terms of the resources that can be targeted on the FPGA, an important factor in commercial applications where running multiple neural networks together is becoming common.
“Deep learning Inference carries a diverse set of requirements across many applications, so creating a strong partner ecosystem of solutions to service these requirements is paramount to ensure broad adoption,” said Reynette Au, vice president of marketing, Intel Programmable Solutions Group. “Our expanded DSN program offers customers an easier way to power FPGA acceleration drawing from a set of IP solutions such as Myrtle Software’s agile software compilation flow that enables users to build custom topologies with an efficient hardware implementation.”
Myrtle Software is a Cambridge UK-based technology company that has developed powerful new software to accelerate deep learning inference on FPGAs.
Myrtle’s technology significantly reduces the time it takes to realize deep learning algorithms on FPGAs and has successfully automated the process of taking pre-trained neural network descriptions and outputting efficient silicon designs.
Myrtle has realized multiple proprietary deep learning networks as silicon designs so that they execute at a latency and power point that makes them usable in real-world situations. Myrtle is also currently involved in a major collaboration to address the safety and verification challenges that are currently preventing sophisticated deep learning networks being used in road vehicles.