The chip competition has entered a new high as the demand for faster and more powerful chips in the field of artificial intelligence and machine learning continues to grow.
In a post on Meta's official website a few days ago, it was revealed that the company launched the second generation of its training and inference gas pedal. The chip is designed to provide the necessary computing power for Meta's AI infrastructure. Meta first introduced the chip last year and claimed that its second-generation product offered a significant boost in performance.
Unlike recent market competition between chipmakers such as Intel and Nvidia for faster, more powerful processors in artificial intelligence (AI) and high-performance computing, Meta isn't targeting its chip products at mass-market AI customers. Instead, Meta has chosen the path of developing customized silicon-based chips designed specifically to meet the company's own AI processing needs. The gas pedal consists of an 8×8 grid of processor elements, Meta said. These processor elements deliver significantly improved dense computation performance as well as sparse computation performance. Meta explains that these performance improvements are largely due to optimizations and enhancements to the sparse computation pipeline architecture.
Meta has made several improvements to the design of the new chip: first, tripling the local storage capacity per processor element (PE); second, expanding the static random access memory (SRAM) on the chip from 64MB to 128MB; third, increasing the chip's bandwidth by a factor of 3.5; and increasing the capacity of the low-power double-data-rate synchronous dynamic random access memory (LPDDR5). LPDDR5) capacity has been doubled. The new chip also runs at 1.35 GHz, up from 800 MHz, and is physically larger than its predecessor and utilizes a more advanced 5-nanometer process.
To support the next-generation chip, Meta has developed a large rack-mounted system that can accommodate up to 72 gas pedals. The system consists of three racks, each containing 12 boards with two gas pedals mounted on each board. On the software side, Meta said the company further optimized its software stack and created the Triton-MTIA compiler back-end to produce high-performance code for MTIA hardware.The Triton-MTIA back-end performs a number of optimizations to maximize hardware utilization and support high-performance core algorithms.
Similar to Meta, Google is developing custom internal silicon for its AI. At Google's Cloud Next computing conference, Google approved detailed information about several of its AI chips designed for data centers and announced a central processor based on the Arm architecture. It is claimed to have twice the performance of previous Google TPUs.
Not to be outdone at the event, Intel unveiled the Gaudi3 AI gas pedal.The Gaudi3 is designed to achieve AI compute speeds up to four times faster than its predecessor, with a 1.5x increase in memory bandwidth and a doubling of network bandwidth to allow for significant advances in large-scale system scaling.
Intel expects this chip to dramatically improve the performance and efficiency of AI training and inference when working with popular large language models and multimodal models.The Gaudi3 gas pedal is fabricated on an advanced 5nm process and is designed to allow all engines, such as matrix multiplication engines (MMEs), tensor processing cores (TPCs), and network interface cards, to operate in parallel, thus delivering the deep learning computational the fast and efficient acceleration required for deep learning computations.
As of now, the competition in the AI chip market has been very intense, with companies vying for market share through a variety of means, such as technological innovation, product optimization, and market strategies. With the continuous progress of technology and the expansion of application scenarios, the competition pattern of the AI chip industry will be more diversified and complex.
