3 ms·
This is a weird article. Headline says they are building an ASIC. Provides absolutely no details other than facebook are hiring hardware engineers. It could be
by phire 8y ago
This is a weird article.
Headline says they are building an ASIC. Provides absolutely no details other than facebook are hiring hardware engineers. It could be a CPU (CPUs are ASICs too), it could be a GPU. It could be something related to machine learning, or it could be enterprise great switching fabric. Who knows?
Then the article goes off on a huge nig tangent about how facebook is working on a compiler for generic machine learning accelerators. The article explictly says that this compiler targets all machine learning accelerators, especially a bunch that are already in development or released by other companies.
These are completely independant chunks of news, nothing facebook has said links one to the other. Yet by putting both news items in a single article, the author implies that facebook are working on a machine learning chip, while providing absolutly no proof.
- Scaevolus 8y agoWhy would Facebook build a CPU or a GPU? Custom routers generally use off the shelf switching fabric chips with custom software stacks, but it's board-level hardware design, not ASIC design. Machine learning consumes a huge amount of compute, and GPUs waste a lot of power executing the workloads. Specialized accelerators for it pay back their investments quickly.
- tyingq 8y agoAmazon is using a custom ASIC in their network gear: https://www.geekwire.com/2017/amazon-web-services-secret-weapon-custom-made-hardware-network https://www.geekwire.com/2017/amazon-web-services-secret-wea...
- phire 8y agoFacebook is a large company with many projects going on. While I admit that machine learning accelerators is a likely candidate, there are a huge number of possibilities for potential ASIC projects. Facebook owns Oculus, I could easily see Oculus needing a custom ASIC, maybe a low-latency display driver. Or a high-powered SoC if they decide to do an untethered Rift. This is also where I could see them potentially making a GPU. Making a whole SoC is hard, why not take an off the shelf one and attach it to a custom mobile GPU chip. It's also possible that Facebook might have a smartphone project on the go (again). A custom SoC could be useful for that. Or facebook might be considering post smartphone devices. There are lots of places where a SoC could be useful. Facebook has been making their own open servers (and switches) for a while. But they still built around "backdoored" intel CPUs. I could easily Facebook wanting a high-performance server grade open CPUs (probably RiscV based).
- raverbashing 8y ago> It could be something related to machine learning, or it could be enterprise great switching fabric These are the ones that make sense, and my money is on ML