Facebook has posted a job opening looking for an expert in ASIC and FPGA, two custom silicon designs that companies can gear toward specific use cases — particularly in machine learning and artificial intelligence.
There’s been a lot of speculation in the valley as to what Facebook’s interpretation of custom silicon might be, especially as it looks to optimize its machine learning tools — something that CEO Mark Zuckerberg referred to as a potential solution for identifying misinformation on Facebook using AI. The whispers of Facebook’s customized hardware range depending on who you talk to, but generally center around operating on the massive graph Facebook possesses around personal data. Most in the industry speculate that it’s being optimized for Caffe2, an AI infrastructure deployed at Facebook, that would help it tackle those kinds of complex problems.
FPGA is designed to be a more flexible and modular design, which is being championed by Intel as a way to offer the ability to adapt to a changing machine learning-driven landscape. The downside that’s commonly cited when referring to FPGA is that it is a niche piece of hardware that is complex to calibrate and modify, as well as expensive, making it less of a cover-all solution for machine learning projects. ASIC is similarly a customized piece of silicon that a company can gear toward something specific, like mining cryptocurrency.
Facebook’s director of AI research tweeted about the job posting this morning, noting that he previously worked in chip design:
While the whispers grow louder and louder about Facebook’s potential hardware efforts, this does seem to serve as at least another partial data point that the company is looking to dive deep into custom hardware to deal with its AI problems. That would mostly exist on the server side, though Facebook is looking into other devices like a smart speaker. Given the immense amount of data Facebook has, it would make sense that the company would look into customized hardware rather than use off-the-shelf components like those from Nvidia.
(The wildest rumor we’ve heard about Facebook’s approach is that it’s a diurnal system, flipping between machine training and inference depending on the time of day and whether people are, well, asleep in that region.)
Most of the other large players have found themselves looking into their own customized hardware. Google has its TPU for its own operations, while Amazon is also reportedly working on chips for both training and inference. Apple, too, is reportedly working on its own silicon, which could potentially rip Intel out of its line of computers. Microsoft is also diving into FPGA as a potential approach for machine learning problems.
Still, that it’s looking into ASIC and FPGA does seem to be just that — dipping toes into the water for FPGA and ASIC. Nvidia has a lot of control over the AI space with its GPU technology, which it can optimize for popular AI frameworks like TensorFlow. And there are also a large number of very well-funded startups exploring customized AI hardware, including Cerebras Systems, SambaNova Systems, Mythic, and Graphcore (and that isn’t even getting into the large amount of activity coming out of China). So there are, to be sure, a lot of different interpretations as to what this looks like.
One significant problem Facebook may face is that this job opening may just sit up in perpetuity. Another common criticism of FPGA as a solution is that it is hard to find developers that specialize in FPGA. While these kinds of problems are becoming much more interesting, it’s not clear if this is more of an experiment than Facebook’s full all-in on custom hardware for its operations.
But nonetheless, this seems like more confirmation of Facebook’s custom hardware ambitions, and another piece of validation that Facebook’s data set is becoming so increasingly large that if it hopes to tackle complex AI problems like misinformation, it’s going to have to figure out how to create some kind of specialized hardware to actually deal with it.
A representative from Facebook did not yet return a request for comment.
Read more TechCrunch