Intel announced this week it has begun shipping its 10nm Agilex FPGAs to early-access customers, including Microsoft, featuring the Compute Express Link (CXL), a cache and memory coherent CPUs-to-anything interconnect that has an industry consortium of more than 60 members. The company first announced the chips in April.
The Agilex FPGA is the product of the Altera group, which Intel bought in 2015 for $16.7 billion. It sold FPGAs under the Stratix brand name, but this line is the first to come out under Intel ownership. CXL replaces OmniPath Connect, a fabric Intel developed but no one else supported. The company ended support for OmniPath earlier this month in favor of CXL, which has wide industry support.
According to Intel, the new Agilex parts offer up to 40% higher performance or 40% lower power than the Stratix line of chips. AgileX also adds support for DDR4 and DDR5, which isn’t even out yet, high-bandwidth memory (HBM), and Optane DC Persistent Memory DIMMs. It’s also compatible with both PCIe 4.0 and PCIe 5.0, though PCIe 5.0 support will come at a later date.
With its 3D silicon-in-package (SiP) technology, Intel promises the chip will do virtually every kind of computing task. “This combination of advanced technologies allows Intel to integrate analog, memory, custom computing, custom I/O, and Intel eASIC device tiles into a single package along with the FPGA fabric,” the company said in a statement. Intel offers a “custom logic continuum” that allows developers to migrate designs from FPGAs to low-power ASICs.
Intel also said Agilex is the only FPGA supporting hardened BFLOAT16, a floating point instruction set popular in AI chips, with up to 40 teraflops of digital signal processor (DSP) performance (FP16), and that it supports up to 112 Gbps data rates for high-speed networking required for 400GE.
5G and IoT uses
Intel expects the customers will use Agilex FPGAs to develop networking, 5G, artificial intelligence, and accelerated data analytics solutions, saying the chip provides faster networking throughput and lower latencies.
The first place you will be able to use them is Azure. Microsoft said it plans to use the FPGAs for accelerating real-time AI, networking, and other applications/infrastructure across Azure Cloud Services, Bing, and other data center services.