A wafer of Xeon E5 v3 processors. Source: Intel.
Earlier this year, chip giant Intel (INTC -0.38%) announced its intent to acquire programmable logic chip maker Altera (NASDAQ: ALTR). One of the key strategic benefits of this acquisition, according to Intel, is that the combined company will be able to offer processors that integrate Intel's CPUs with Altera's FPGAs.
FPGAs are of interest to Intel's customers primarily because they can be configured to act as workload specific accelerators on-the-fly. Although FPGAs can be paired in stand-alone fashion with Intel's processors, Intel has said that integration of FPGAs with its processors deliver significant efficiency and cost improvements over using stand-alone solutions.
A question worth pondering, then, is when investors should expect Intel to launch products with FPGAs integrated into the same piece of silicon as its Xeon processors. Let's take a closer look.
Expect on-package integration with Purley platform; on-die integration with Tinsley platform
Every two years or so, Intel introduces a new server platform. The company's last major server platform, known as Grantley, launched in late 2014. The next platform, known as Purley, is expected to arrive during the first half of 2017.
I believe that with the Purley platform, we will see products from Intel that integrate Xeon processors with Altera FPGAs onto the same package as a number of leaked slides have indicated. This should bring some efficiency benefits relative to an FPGA connected via an add-in board, but the maximum efficiency and cost benefit should come when the FPGA is integrated on-die.
The first chips compatible with the Purley platform, known as Skylake-EP and Skylake-EX, should launch alongside the platform in the first half of 2017. The follow-on chips for the platform, known as Cannonlake-EP and Cannonlake-EX, should launch in the first half of 2018.
Given that the Altera acquisition isn't expected to close until early next year, and given that it will take time to design an FPGA on Intel's 10-nanometer technology (Altera has yet to begin shipping stand-alone FPGAs built on Intel's 14-nanometer technology), I believe that we won't see such integration until Intel launches the follow-on platform to Purley.
The follow-on to Purley should launch in the first half of 2019, or roughly three years from now. This should give Intel and Altera enough time to not only put together a 10-nanometer FPGA but to integrate it into the Icelake-EP/EX processors that will launch with that platform.
What could integrated FPGAs mean for Intel's data center business?
Intel likely went ahead and spent $16.7 billion to acquire Altera because its customers indicated significant interest in FPGA acceleration. Indeed, Intel's claim during the transaction announcement that as many as a third of cloud service provider compute nodes could use FPGAs by 2020 was probably at least partially driven by customer feedback and/or requests.
Ultimately, if Intel is successful in delivering good integrated parts, the integration of FPGAs with Xeon processors should lead to a material increase in platform average selling prices in its data center group as the value of stand-alone FPGA accelerators are absorbed into Intel's Xeon processors.
Further, if FPGA acceleration turns out to be a big deal for Intel and its customers, then having FPGA capabilities in-house should prove quite an interesting competitive advantage. Indeed, it seems as though it will be very difficult for Intel's competition in the data center to develop such tightly integrated products to compete with Intel's CPU-plus-FPGA solutions.
Although Intel should see a benefit from selling Xeon processors with FPGAs on package beginning in the first half of 2017, I don't think that things will get really interesting until Intel delivers a solution with a Xeon processor and an FPGA on the same piece of silicon.
