With last week’s big Altera acquisition Intel made an expensive bet on a future of data center hardware that uses significantly more customized designs than today’s monolithic racks of commodity x86 servers. As I wrote at the time, “The only justification for Intel’s move can be its perception of a secular technology shift from commodity processors to custom hardware purpose-built for specific applications,” since the financial numbers didn’t justify such an exorbitant price. The market apparently agreed, since Intel’s stock price has lagged the S&P 500 by more than 6% in the intervening week. Markets focus on the short-term whereas this deal is decidedly part of a long-term strategy, however incorporating the FPGA hardware is actually the easy part.
In justifying the acquisitions, Intel CEO Brian Krzanich highlighted the potential for dramatic performance improvements by integrating Altera FPGAs with Xeon processors and there are many proven cases, notably in scientific computing, where executing application code on customized hardware yields astounding improvements. The problem is that FPGAs and GPUs are more difficult to program, requiring specialized code using device-specific APIs and an understanding of the underlying peculiarities of the FPGA or GPU hardware. It will be impossible for Intel to realize Krzanich’s prediction that FPGAs are part of one-third of all cloud servers by 2020 without eliminating the software development hurdles between conventional application code and its execution on non-traditional hardware. What Krzanich didn’t highlight is the significant technical progress being made on this front. For an overview of promising solutions, ranging from high-level languages and APIs to new software layers that perform real-time code translation, see the full column.
As the column concludes, with Altera soon to be part of Intel’s hardware arsenal, expect the company to focus on solving the problems of using FPGAs to accelerate a wide variety of applications. In fact, Intel and Altera recently co-sponsored the Heterogeneous Architecture Research Platform (HARP) research program through the ACM “to spur research in programming tools, operating systems, and innovative applications for accelerator-based computing systems.” As the column highlights, startups like Bitfusion could make attractive acquisition for its hardware accelerator abstraction software, but Intel has also shown, through work on Hadoop, Cloudera, OpenStack and the Linux kernel, that it understands how to foster organic software development in support of its hardware. The race to simplify application development in a world of CPU accelerators will be interesting to watch.