AI chip race
Sci-Tech Snippets-8.
Synopsis based on papers from IEEE Computing Edge May 2023, Intelligent Systems, Vol. 37, №2, 2022, and CSET 2020.
The AI chip market forecast is $79.8 billion by 2027 (from $10.6 billion in 2021).
AI chips are hardware accelerators. They offer high efficiency for AI and ML applications, especially for transformers and DNN (deep neural networks) training and inference.
Types of AI chips include GPUs (graphical processing units), FPGAs (field-programmable gate arrays), and ASICs (application-specific integrated circuits).
CPUs (central processing units) are general-purpose chips efficient for strong sequential operations, whereas GPUs are designed for intensive parallel computing and high-bandwidth memory. FPGAs have better inference capabilities, whereas ASICs are relatively easier to design.
Thumb rule: GPUs are used for training, FPGAs for inference, and ASICs for both.
In general, AI chips provide 10x to 1000x improvements in efficiency and speed relative to CPUs, with GPUs and FPGAs at the lower end and ASICs at the higher end. DSL (domain-specific language) provides efficiency gains to ASIC. Inference accuracy is highest for GPUs (9899.7%); for ASICs, it is 90-98%.
Today’s race horses
- H100 NVIDIA: The GPU for training DNNs with a parallel computing platform compute unified device architecture.
- TPUv4 Google: ASIC. 2x performance over TPU v3 chips.
- VCK5000 AMD-Xilinx: FPGA. More customizable than ASICs.
- Agilex-M series Intel: Fastest FPGA with in-package high-bandwidth memory.
- Kunlun AI chips, Baidu, and Ascend 910 and 301 from Huawei.
Node race
The increase in transistor density on a chip correlates with the number of nodes. Each node corresponds to the transistor size (in nm). A node change doubles the density.
TSMC (Taiwan Semiconductor Manufacturing Company), Samsung, and Intel are in the 7 nm and 5 nm fab (fabrication) races. Nikon actively supplies for <= 90 nm fab. Only ASML Photolithography from the Netherlands sells equipment for 5 nm fabs.
Notable incumbents in the race
LG (S. Korea), Cambricon (China), Iluvatar CoreX (China), Biren (China), MetaX (China), SambaNova US, Cerabras US, Graphcore (England), PEZY Computing (Japan), and TensTorrent Canada).
