The Xilinx-Intel conflict has heated up a notch this week, as Xilinx bought a 200-person Chinese startup called DeePhi. The industry consensus is that this is an acqui-hire, made to bring a batch of developers with strong experience with Xilinx’s FPGAs in-house – as hiring staff has become difficult.
However, Riot has explored that apparent misconception previously, noting that there’s a real lack of solid evidence to support these kinds of claims. There are shortages in some specific sectors, but the belief that there is a general shortage is sustained partly by the field itself and those with vested interests.
Nonetheless, this does smack of a conventional acqui-hire, as Xilinx looks to bring valuable talent in-house, likely in preparation for a drawn-out war of attrition with Altera – it’s chief rival that is now owned by Intel, after $16.7bn was paid to ensure Intel could secure a strong foothold in a market that was about to quickly ramp up.
In the three years since Intel bought Altera, it has aligned itself with Microsoft, which now installs an Altera FPGA in all the new servers it spins up. Xilinx, in turn, has sided with Microsoft Azure’s chief rival, Amazon’s AWS wing, with a deal in September 2017 seeing AWS offer Xilinx FPGA resources and software development tools.
Both camps are looking to capitalize on the growing demand for heterogenous computing (het-gen), which broadly describes the use of more than one type of processing core architecture – such as CPU and GPU. With the growth in demand for AI and machine-learning (ML) applications, GPUs were able to address a real gap in the market, by providing far greater performance and bang-for-the-buck than servers full of CPUs could manage.
But this market is already moving towards more specialist designs, which would use the Field Programmable Gate Arrays (FPGAs) and eventually custom Application-Specific Integrated Circuits (ASICs), to leverage the benefits of dedicated hardware. This is why Intel wanted to quickly expand via Altera, and why Xilinx has to try to keep pace, as Intel still prints money – with a market cap of $241.3bn, compared to Xilinx’s $17bn.
So, DeePhi is part of this strategy. Founded in 2016, by a group from Stanford and Tsinghua Universities, the company has raised around $40m in investments. Samsung is a backer, as is MediaTek, along with seven Chinese VC firms. It is thought that Samsung is using DeePhi’s technology in its Galaxy S9, via the Samsung Exynos 9810 SoC.
DeePhi specializes in neural networks, and to this end has two hardware architectures. Aristotle is designed for video and image recognition, and uses Convolutional Neural Networks (CNNs) to this end, while Descartes is intended for speech recognition. As an embedded neural networking specialist, DeePhi should fit right in at Xilinx.
The startup was known for its model compression expertise – the practice of taking the neural network model that interprets data and comes to a decision, and then compressing it into a smaller package. Crucially, DeePhi was good at doing this without the model losing accuracy.
But now some real battle lines have been drawn, between both Xilinx and Intel, and AWS and Microsoft. As more cloud computing customers demand improved AI and ML performance, these providers are going to lean more heavily on FPGA providers. The main benefit of the FPGA is that it can be altered to best suit the application at hand, before being changed to address the next application that the cloud computing instance needs – without having to swap out components in the server rack.
No prices are being discussed for the acquisition, but it is worth noting that Xilinx had invested an undisclosed amount in DeePhi in May 2017. That announcement was made shortly after a team from DeePhi won the Best Paper award at FPGA2017, after achieving 43x the performance and 40x the performance per watt compared to a CPU; 3x the performance and 11x the performance per watt compared to a GPU. DeePhi will continue operating out of China.