A Chinese research paper reveals that “ACCEL”, an in-house analog AI processing chip, can deliver 3000 times faster performance than NVIDIA’s A100 & A800 GPUs.
Chinese ACCEL Analog AI Chip Reportedly Provides “3000 Times” Faster Performance Than NVIDIA’s A100 & A800
With China under the influence of global sanctions, it seems like the country is rapidly improving its “homegrown” solutions, in an attempt to maintain its existing pace of industry growth. A paper published by Tsinghua University, China reveals that the institute has devised a new technique for AI computing performance and developed a chip named ACCEL ( All-Analog Chip Combining Electronic and Light Computing), which basically harnesses the power of photonics and analog technology to provide exceptional performance, and the numbers revealed are quite shocking.
According to the publication via Nature, the AI chip ACCEL has the capability to deliver 4.6 peta-operations per second, which is indeed way ahead of what the current industry solutions offer but that isn’t all. The chip is designed to maintain power efficiency, since without doing so, it wouldn’t be applicable to the industry. The ACCEL employs a “systemic energy efficiency” architecture, which is able to deliver 74.8 peta-operations per second per watt. Hence, as the numbers disclose, the chip deviates from the industry trends, where high computing power is directly proportional to more power consumption.
Without any sort of real-time benchmark, labeling a chip as the “industry’s fastest” is justice, however, ACCEL was experimentally put up against the Fashion-MNIST, 3-class ImageNet classification and time-lapse video recognition scenarios to test the limits of “deep-learning” performance of the chip. It was able to deliver accuracies of 85.5%, 82.0%, and 92.6%, respectively, which depicts that the chip has widescale industry applications, and is not just limited to a particular segment. This makes things more exciting with ACCEL, and we can’t wait to see what the chip brings to the future.
Now let’s talk about how ACCEL actually works. The chip combines the capabilities of diffractive optical analog computing (OAC) and electronic analog computing (EAC) with scalability, nonlinearity, and flexibility. To achieve such efficiency numbers, the chip features an optoelectronic hybrid architecture in an all-analog way to reduce massive ADCs (Analogue-Digital Conversions), in large-scale workloads, which results in a much improved performance. The research paper published covers the mechanism of the chip quite extensively, hence you can have a look at it here, to get an idea of how things work with ACCEL.
For state-of-the-art GPU, we used NVIDIA A100, whose claimed computing speed reaches 156 TFLOPS for float32 (ref. 33). ACCEL with two-layer OAC (400 × 400 neurons in each OAC layer) and one-layer EAC (1,024 × 3 neurons) experimentally achieved a testing accuracy of 82.0% (horizontal dashed line in Fig. 6d,e). Because OAC computes in a passive way, ACCEL with two-layer OAC improves the accuracy over ACCEL with one-layer OAC at almost no increase in latency and energy consumption (Fig. 6d,e, purple dots). However, in a real-time vision task such as automatic driving on the road, we cannot capture multiple sequential images in advance for a GPU to make full use of its computing speed by processing multiple streams simultaneously48 (examples as dashed lines in Fig. 6d,e). To process sequential images in serial at the same accuracy, ACCEL experimentally achieved a computing latency of 72 ns per frame and an energy consumption of 4.38 nJ per frame, whereas NVIDIA A100 achieved a latency of 0.26 ms per frame and an energy consumption of 18.5 mJ per frame.
How will ACCEL and similar analog AI chip developments reshape the industry? Well, answering this question right now isn’t easy, given that the adoption of analog-based AI accelerators is still something for the future. While the performance numbers and statistics are quite optimistic, an important fact to note is that “deployment” of them in the industry isn’t as easy as it seems, given that it requires more time, greater financial resources, and in-depth research work. However, none can argue that the future looks bright for computing, and it is only a matter of time before we see such performance in the mainstream industry.
News Source: Tom’s Hardware