Were you unable to attend Transform 2022? Check out all of the summit sessions in our on-demand library now! Watch here.
Intel claims its new Agilex M has the world’s highest bandwidth field-programmable gate array.
The company formally launched its 10nm Agilex I- and F-series FPGAs about a year ago, and has now completed its portfolio with the introduction of the M-series. At the time, Intel claimed its FPGAs delivered up to two times the performance per watt compared to its competitor Xilinx 7nm Versal FPGAs, which was recently acquired by AMD. Intel claims the M-series maintains this advantage. Intel said the newest FPGAs are based on the Intel 7 process, which is a further enhancement of the previous 10nm SuperFin process used in the other Agilex FPGAs.
Agilex M is all about the memory
The main focus of Agilex M is memory. The FPGA features in-package high bandwidth memory (HBM) DRAM, which allows it to achieve the highest bandwidth among FPGAs. The M-series further has hardened support for the other common types of memory, including double data rate 5 (DDR5) and Optane, and also has fast on-chip SRAM.
MetaBeat will bring together thought leaders to give guidance on how metaverse technology will transform the way all industries communicate and do business on October 4 in San Francisco, CA.
Intel further claims that it has the highest digital signal processing (DSP) compute density compared to other FPGAs with HBM. Enabled through 12,300 variable-precision DSP blocks, Agilex M achieves 18.5 TFLOPS of FP32 throughput, or 88.6 tera operations per second (TOPS) at INT8 precision. For comparison, Xilinx Versal HBM achieves respectively 17.5 TFLOPS and 74.9 TOPS.
To be specific, the M-series features two HBM2e stacks, for a maximum of 32Gb and 820Gbps bandwidth, which is 60% higher than the Stratix 10 MX and the same as Xilinx Versal HBM. Combined with 8 DDR5 DIMMs, total bandwidth becomes 1.1TB/s. This is a bit higher than Xilinx Versal HBM’s 1.06Tbps or the Achronix Speedster 7t’s 0.5 TB/s.
To support such a high bandwidth, Intel said it developed a hardened dual-memory network on chip (NoC) with a peak aggregate bandwidth of 7.5TB/s. In addition, to move data between the FPGA and external systems, the FPGA has 72 SerDes transceivers, including eight that can operate at 116 Gbps, which is the fastest commercially available transceiver data rate. This allows support for 400G ethernet, and there is also support for PCIe 5.0 and CXL.
Intel says these specifications make the FPGA suitable for a range of applications. These include next-generation firewalls for offloading compute-intensive workloads from the CPU, 8K broadcasting, 800GbE test and measurement, 5G RF, and cloud applications such as AI and cryptocurrency.
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.