Ma showed an AI processor with 128 cores and 25W

M**a showed an AI processor with 128 cores and 25W power consumption and announced the development of its improved successor

Meta* Platforms yesterday shared new details on its projects to modernize its data centers to better support AI systems. Among other things, the self-developed MTIA AI chip family was reported.

    Image source: ***

Image source: Meta*

Facebook owner* and Instagram* announced in a series of blog posts that he had developed the first generation chip in 2020 as part of the Meta program.* Training and Inference Accelerator (MTIA). The goal was to improve the performance of the recommendation models the company uses to suggest ads and other content in newsfeeds. However meta* has changed its mind about the wide adoption of its first proprietary AI chip and is now working on its successor. In a blog post, the company describes the first MTIA chip as a “learning opportunity.”

The first MTIA chip focused solely on the work of already trained neural networks, for example to run algorithms that decide whether, for example, a dance video or a meme about a cat should be displayed as the next message in the user’s feed. it says in the blog. Joel Coburn, software engineer at Meta*said during the presentation of the new chip that Meta* Initially, people turned to graphics processing units (GPUs) to run neural networks, but found that they were not well suited for such tasks. “Despite significant software optimizations, their performance for real models is low. This makes their practical application difficult and expensive. That’s why we need MTIA‘ said Coburn.

Meta last year* was involved in a major project to modernize its AI infrastructure after management realized it lacked the hardware and software to meet the demands of teams building AI-powered capabilities. As a result, the company abandoned plans for a large-scale implementation of its own chip to run AI and began work on a more ambitious chip that can also train neural networks.

Meta* admitted that its first MTIA chip cannot handle high-complexity AI models, but found that it can handle low- to medium-complexity models more efficiently than competitor chips. The MTIA chip consumes just 25 watts of power, a fraction of what market-leading chips from vendors like NVIDIA consume, and uses an open-source chip architecture called RISC-V. The MTIA v1 processor is manufactured using 7 nm process technology, has dimensions of 19.34 × 19.1 mm and contains 64 processing units, each of which contains two RISC-V cores, for a total of 128 cores. The clock frequency is only 800 MHz. Supports up to 64GB RAM connected via 16 channels.

In addition to MTIA, Meta* is developing another chip to handle specific types of computing workloads, the company announced at today’s event. Called Meta* Scalable video processor or MSVP, the chip is Meta’s first in-house ASIC solution*, designed for video-on-demand processing and video streaming. Remember this meta* started developing server video chips a few years ago and announced ASICs for video transcoding and conferencing in 2019. MSVP is the result of some of those efforts, as well as a renewed attempt to gain a competitive advantage in the sweat video space.

Only on Facebook* People spend 50% of their time in the app watching videos. For example, to serve a wide range of devices around the world, videos are uploaded to Facebook* or Instagram*, transcoded into multiple bitstreams with different encoding formats, resolutions and quality. MSVP is a programmable and scalable chip and can be configured to efficiently support both the high quality transcoding required for VOD and the low latency and faster processing time required for live streaming.Co-author of Meta Executives* Harikrishna Reddy and Yunqing Chen.

The company added that it has an artificial intelligence-based system that helps its engineers create computer code, similar to the tools offered by Microsoft Corp, Inc and Alphabet Inc.

* It is included in the list of public associations and religious organizations for which the court made a final decision to dissolve or prohibit activities on the grounds provided for in Federal Law No. 114-FZ of July 25. 2002 “On Countering Extremist Activities”.

About the author

Dylan Harris

Dylan Harris is fascinated by tests and reviews of computer hardware.

Add Comment

Click here to post a comment