Stefan Mileschin

Arm first announced Project Trillium machine learning IPs back in February and we were promised we’d be hearing more about the product in a few months’ time. Project Trillium is unusual for Arm to talk about because the IP hasn’t been finalised yet and won’t be finished until this summer, yet Arm made sure not to miss out on the machine learning and AI “hype train” that has happened over the last 8 months in both the semiconductor industry and as well as particularly in the mobile industry.

Today Arm details more of the architecture of what Arm now seems to more consistently call their “machine learning processor” or MLP from here on now. The MLP IP started off a blank sheet in terms of architecture implementation and the team consists of engineers pulled off from the CPU and GPU teams.

With the MLP Arm set out to provide three key aspects that are demanded in machine learning IPs: Efficiency of convolutional computations, efficient data movement, and sufficient programmability. From a high level perspective the MLP seems no different than many other neural network accelerator IPs out there. It still has a set of MAC engines for the raw computational power, while offering some sort of programmable control flow block alongside a sufficiently robust memory subsystem.

Starting off at a more detailed view of the IP’s block diagram, the MLP consists of common functional blocks such as the memory interconnect interfaces as well as a DMA engine. The above graphic we see portrayal of the data flow (green arrows) and control flow (red arrows) throughout the functional blocks of the processor. The SRAM is a common block to the MLP sized at 1MB which serves as the local buffer for computations done by the compute engines. The compute engines each contained fixed function blocks which operate on the various layers of the neural network model, such as input feature map read blocks which pass onto control information to a weight decoder.

https://www.anandtech.com/show/12791...p-architecture