Internet Giants and the Need for Faster AI: Part 2 Dear Daily Prophecy Reader, Yesterday, I began this discussion of the need for speed when it comes to artificial intelligence (AI). If you missed it, you can click here to catch up. Now, Alibaba got into the business of making its custom AI chips by acquiring an ASIC (application specific integrated circuit) maker called C-Sky, which it renamed Pingtogue. Pintogue now drives all semiconductor development for Alibaba, a way of ensuring that Pintogue will develop the chips Alibaba needs. That's integration. However, buying capabilities is not the only path to technology integration. After all, Pintogue itself does not incorporate all the competencies required to build the array of AI chips Alibaba will need. Among those it is missing is: competency in processors whose logic paths can be reset in the field, based on experience, often known as field programmable gate arrays (FPGAs). The Need for Adaptable Chips At its core, artificial intelligence is a learning machine, continuously adapting and course-correcting as it encounters new data. This requires that the chips themselves also be adaptable. Programmability is particularly important in AI applications because the algorithms and neural networks are constantly evolving, getting better, faster, more efficient, and more accurate day-by-day. Pintogue's specialty is building application specific integrated circuits (ASICs) chips for custom applications. This includes leading-edge AI performance, but once an ASIC is created its logic functions — however customized — are locked in. If an ASIC needs updating, tweaking, or bug fixing it means a new chip, new design, new fabrication line. It's not only back to the drawing board, it's back to the bank. This cost and lack of flexibility are big reasons for the tremendous appeal of FPGAs (produced by the likes of Xilinx, Intel, and Achronix). FPGAs include logic gates that are not specified at the factory. These chips can be electronically re-programmed in the field so that, for instance, a piece of "and" logic becomes an "or" or a "nor" or a "nand." Transistors that pass current can be re-programmed to block it. FPGAs can accommodate virtually any logic design, from networking to artificial intelligence, and data signal processing. They can be quite expensive on a per-unit basis, but the non-recurring engineering costs involve only design efforts and not the costly fabrication (and long schedule times) of a fully custom ASIC device. Heretofore, engineers had to choose: ASICs, cheap in volume, but fixed in function? Or FPGAs, individually more expensive but flexible. Why not combine them instead? |
No comments:
Post a Comment