China’s analog AI chip is 3000 times faster than Nvidia’s A100 GPU.

A new paper from China’s Tsinghua University describes the development and performance of an ultra-fast and highly efficient artificial intelligence processing chip specialized in computer vision tasks. The all-analog combined electronic and optical computing (ACCEL) chip, as the chip is called, leverages photonic and analog computing in a specialized architecture capable of delivering more than 3,000 times the performance of the Nvidia A100 with four million times the power consumption. Bottom Line Yes, this is a specialized chip – but rather than seeing it as market fragmentation, we can see it as another step towards a future of heterogeneous computing, where semiconductors are increasingly designed to match a specific need. up to an “everything” configuration.

As published in Nature, ACCEL performed 4.6 trillion operations per second on the vision task—hence a 30,000-fold performance improvement over Nvidia’s A100 (Ampere) and its 0.312 quadrillion operations. According to the research paper, ACCEL can perform 74.8 quadrillion operations per second with 1 watt of power (what the researchers call “system energy efficiency”) and a computing speed of 4.6 peta-operations per second. Nvidia’s A100 has since been replaced by Hopper and the H100’s 80 billion transistor chip, but even against that the results look unremarkable.

Of course, speed is essential in any processing system. However, accuracy is essential for computer vision tasks. However, the range of applications and ways to use these systems to govern our lives and civilization is wide: from the wearable device market (perhaps in XR scenarios) through autonomous driving, industrial inspections, and other image recognition and detection systems worldwide. has it. General, such as facial recognition. The Tsinghua University paper says that ACCEL experimentally performed against the tasks of Fashion-MNIST, 3-class ImageNet classification, and time-lapse video recognition with “highly competitive” accuracy levels (at 85.5%, 82.0%, and 92.6%, respectively), while Was shown. Superior system robustness in low light conditions (0.14 fJ μm-2 per frame).

ah Traditional optoelectronic computing workflows, including large-scale photodiodes and ADC arrays. b, ACCEL workflow. A diffraction optical computing module processes the input image in the optical domain for feature extraction, and its output optical field is used to generate optical currents directly by the photodiode array for analog electronic computing. The EAC outputs successive pulses corresponding to several output nodes of the equivalent network. The binary weights in the EAC are reconfigured during each pulse by the SRAM by changing the connection of the photodiodes to the V+ or V- lines. The comparator outputs a pulse with the maximum voltage as the expected result of ACCEL. JACCEL schematic with an OAC integrated directly in front of an EAC circuit for processing high-speed, low-energy vision tasks. Interferometer MZI, Mach–Zehnder. D2NN, Deep Diffusion Neural Network” (Image credit: Tsinghua University/Nature)

In the case of ACCEL, the Tsinghua architecture works through optical-diffractive analog computing (OAC) assisted electronic analog computing (EAC) with scalability, non-linearity and flexibility on a chip – but 99% of its performance is implemented in the optical system. According to the paper, this helps combat limitations in other vision architectures such as Mach-Zehnder interferometers and deep diffraction neural networks (DNNs).

#Chinas #analog #chip #times #faster #Nvidias #A100 #GPU
Image Source : www.tomshardware.com

Leave a Comment