Feature: Programmable devices
Adaptive platforms are enabling the next
computing era By Ivo Bolsens, Chief Technology Officer, Xilinx W
e’re at an inflection point in the semiconductor industry, with intelligent,
connected devices becoming pervasive in cars, homes, offices, factories, cities, and so on. The cost for this pervasive Artificial Intelligence (AI) is an exponential increase in the data- processing requirements placed on the semiconductors that power these systems.
Industry challenges Much of what has driven general- purpose processors and computing in general for the last 50 years has run out of fuel. Dennard transistor scaling (when transistors shrink, power density remains constant for a given area of silicon) stopped working in the early 2000s, about the same time RISC architectural improvements were saturating. Ten we hit single-core performance and the power/thermal walls.
26 February 2021
www.electronicsworld.co.uk Multi-core processors with multi-
threading helped deliver more throughput performance, but that started earlier last decade due to Amdahl’s Law, which holds that latency can be taken out of a performance task through parallel computing. At the same time, Moore’s Law – which states that the number of transistors in a dense integrated circuit (IC) doubles about every two years – began to slow. Here, we are using the common concept of Moore’s Law, meaning every two years a processor moves to a new process node, yielding higher performance, lower power and halved area. Today, Moore’s Law is no longer delivering improvements in performance, power and area. You meet one, or maybe two of the three, depending on the application and how hard you work at it. It’s also taking more than two years to get to the next smaller process node. As advances in semiconductor
manufacturing technology have slowed, we’ve needed something new to keep advancing.
Xilinx Vitis chip structure
Architectural innovation The decline of the traditional industry drivers is encouraging many new innovations across the semiconductor industry. An increasingly important one is the use of Domain Specific Architectures (DSAs), which are optimised for a particular domain; AI inference, for example. Such an architecture typically contains logic, memory and interconnect, pre-configured to optimally implement the type of processing that AI inference requires – typically MAC-heavy math operations. The architecture will be flexible to allow different types of neural network layers to be implemented efficiently in hardware. In contrast, general-purpose processors
can execute a broad range of applications, but they generally don’t deliver the best performance or best power efficiency, particularly for applications that can be parallelised, like image processing. The workloads that computing needs today and in the future are increasingly parallel and frequently streaming or ‘live’ and, hence, need to be processed with low latency.
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44
