Seven years after its $16.7 billion acquisition of FPGA maker Altera, Intel is expanding the technology it gained into new areas.
While the primary use for an FPGA processor has been for smartNICs that offload tasks from server CPUs, Intel is now looking to broaden its application from the data center to remote, edge computing, and embedded systems.
It’s not as if the Altera processors languished over the last several years, however. One major change is manufacturing. When Intel purchased Altera, its chips were made by TSMC. Now they are made by Intel, so hopefully that’s one less supply-chain headache to worry about.
“Supply has just become critical,” said Patrick Dorsey, vice president of Intel’s data center and AI group and general manager of the programmable solutions group. “The number one feature in the market is ‘do you have my part?’ So having the supply chain is just going to become critical.”
As for the chips, Intel outlined a roadmap that breaks down the new and future FPGAs into four separate product categories, all under the Agilex brand name.
Agilex D-Series
The new Agilex D-Series FPGAs join the older Cyclone, Arria, and Stratix brands in the midrange space. The D-Series will be physically smaller overall, with lower thermal requirements and a lower price.
Dorsey says the D-Series will be sold for use in areas such as 5G baseband radios, embedded applications such as the medical instrument industry, and manufacturing. The D-Series will also find its way into the data center via applications such as storage acceleration, where the FPGA would handle data movement instead of the CPU, according to Dorsey.
Multi-chip FPGA packages
The second product line consists of Agilex FPGAs with support for Compute Express Link (CXL) and PCIe 5. Designed for demanding processor workloads, these will also find a home in data centers for network processing. This line of FPGAs will feature a multi-chip package – you’ll see the FPGA paired with an ASIC, a CXL chip, and PCIe 5, all in a single package tied together by high-speed interconnect.
This will be part of Intel’s smartNIC, or as Intel calls it, the infrastructure processing unit (IPU) strategy.
One reason for the multichip package strategy is that it allows Intel to upgrade pieces of the chip without requiring a full revision. The initial Agilex FPGAs this year will come with CXL 1.2 support, for example. But next year, they will be upgraded to CLX 2.0. With this design, Intel can simply replace the CXL component in the package without having to redesign and remake the entire chip.
Sundance Mesa
Another part of Intel’s data center strategy focuses on artificial intelligence and supporting AI applications both within the network as well as on the edge. The Sundance Mesa line of new FPGAs will come with advanced AI capabilities related to INT8 processing, which is used for the inference portion of AI.
“You’re not going to do training in an FPGA, but you can do really fast inference, and you can innovate really quickly with that programmable fabric,” said Dorsey.
Intel Direct RF portfolio
The final product line is the Intel Direct RF portfolio, which are processors dedicated to analog and RF signaling applications.
Another “huge focus” for Intel is developers, according to Dorsey. FPGAs are notoriously difficult to write apps for, he says. So Intel has more than 35% of its engineering team creating software to take the load of writing basic code off developers. Intel is also making compilers and other development tools available, and it’s including FPGA in its oneAPI strategy, which Intel designed to simplify programming by creating one API that supports a variety of processors, from CPU to GPU to FPGA and more.
