There's a pretty interesting debate in the AI space right now on whether FPGAs or ASICs are the way to go for hardware-accelerated AI in production. To summarize, it's more about how to operationalize AI - how to use already trained models with millions of parameters to get real-time predictions, like in video analysis or complex time series models based on deep neural networks. Training those AI models still seems to favor GPUs for now.
Google seem to be betting big on ASICs with their TPU. On the other hand, Microsoft and Amazon seem to favor FPGAs. In fact Microsoft have recently partnered with Xilinx to add FPGA co-processors on half of their servers (they were previously only using Intel's Altera).
The FPGA is the more flexible piece of hardware but it is less efficient than an ASIC, and have been notoriously hard to program against (though things are improving). There's also a nice article out there summarizing the classical FPGA conundrum: they're great for designing and prototyping but as soon as your architecture stabilizes and you're looking to ramp up production, taking the time to do an ASIC will more often be the better investment.
So the question (for me) is where AI inference will be in that regard. I'm sure Google's projects are large scale enough that an ASIC makes sense, but not everyone is Google. And there is so much research being done in the AI space right now and everyone's putting out so many promising new ideas that being more flexible might carry an advantage. Google have already put out three versions of their TPUs in the space of two years
Which brings me back to Xilinx. They have a promising platform for AI acceleration both in the datacenter and embedded devices which was launched two months ago. If it catches on it's gonna give them a nice boost for the next couple of years. If it doesn't, they still have traditional Industrial, Aerospace & Defense workloads to fall back on...
Another wrinkle is their SoCs are being used in crypto mining ASICs like Antminer, so you never know how that demand is gonna go. As the value of BTC continues to sink there is constant demand for more efficient mining hardware, and I do think cryptocurrencies are here to stay. While NVDA has fallen off a cliff recently due to excess GPU inventory, XLNX has kept steady.
XLNX TTM P/E is 28.98
Semiconductors - Programmable Logic industry's TTM P/E is 26.48
Thoughts?
Submitted December 12, 2018 at 03:00PM by neaorin https://ift.tt/2BdXLlr
No comments:
Post a Comment