AMD Targets High-End Signal Processing With New Versal Premium With AI Engines

Tech Industry

In the first major product announcement since the acquisition of FPGA leader Xilinx, AMD (the former Xilinx Group) announced a new Versal Premium Adaptive Compute Accelerated Platform (ACAP) with AI Engines for high-end signal processing applications like radar, wireless network testing equipment, semiconductor automated test equipment, and signal intelligence applications. The new ACAP is a System-on-Chip (SoC) that combines the performance and flexibility of the Versal Premium ACAP with the AI engines of the AI Core ACAP series, resulting in a new product that’s ideally suited for extreme signal processing applications.

While none of the silicon technology in this new Versal Premium with AI Engines (Premium AIE) ACAP is new, the new product demonstrates the value and flexibility of FPGAs, especially when combined with the multi-chip packaging technology used to create new products optimized for specific applications. Note that both AMD and Xilinx have been leaders in leveraging Chip-on-Wafer-on-Substrate (CoWoS) multichip packaging technology and have both been developing the technology in partnership with TSMC, further demonstrating why AMD’s acquisition of Xilinx was a great match. The Versal Premium AIE ACAP combines processing cores and I/O interfaces that optimize the high throughput required for signal processing.

As with the other Versal ACAPs (refer to previous articles) the Premium AIE packs a lot of different processing and I/O technologies into a single package. In addition to the AI engines, the Premium AIE includes dual ArmCortex-A72 CPU cores, dual Arm Cortex-R5F cores, DSP engines, and a Versal adaptable FPGA block all connected through a programmable on-chip network. In terms of I/O, the Premium AI supports PCIe Gen5, CCIX, CXL, 100G and 600G ethernet, MIPI, LVDS, and GPIO support. In addition, this ACAP sports on-chip 400G cypto engines. The Premium AI also includes four memory controllers for DDR4 or LPDDR4. That all translates into 9TB/s of serial I/O bandwidth, 5TB/s of ethernet bandwidth, 1.6TB/s of line rate encryption capabilities, and up to 123TB/s of memory bandwidth. Impressive numbers for signal processing application.

The company is introducing two new products, the Versal Premium VP2502 and VP2802. Both have the same CPU complex and 472 AI engine tiles but the VP2802 offers up to twice the number of adaptable engines and more I/O cores.

In addition to being able to run AI models more efficiently, the addition of the AI engines allows the signal processing workloads to be split between the DSP and AI engines, using each most efficiently for the relative tasks with improved overall performance, but still having the flexibility to run those tasks on either set of cores.

According to the company, the improvement in overall performance is significant. When compared to the previous generation Virtex UltraScale+ VU13, the Versal Premium VP2802 ACAP increases performance by 3.3x for RF signal processing, 4.7x for Radar data cube processing, and 5.8x for AI/ML inference processing while reducing the number of FPGAs needed for a design. In some cases, the number of FPGAs needed drops from as many as eight to two, which also reduces the size and complexity of the system board required. AMD claims that the new products will also exceed the performance other competing products in the market, but Tirias Research is unable to verify those test results.

The company provided two examples that demonstrate the dramatic increases in performance and efficiency while reducing system size and power. The first example was adaptive signal processing for radar, in which moving from using three Virtex UltraScale+ FPGAs to only one Versal Premium VP2802 ACAP doubles performance while lowering power consumption by 43%. The second example was a 5G mobile phone tester, where the design was simplified from two Virtex UltraScale and one Zynq UltraScale+ FPGAs to a single Versal Premium VP2802 ACAP. The result was a 1.6x increase in memory bandwidth, 2.1x increase in serial I/O bandwidth, and 3.3x increase in compute density while using 37% less silicon area and 22% less power.

As with the other Versal ACAP products, AMD will be introducing a new toolset that is based on the existing tools but optimized for the Premium AIE. As a result, developers can begin designing applications today using the current Versal Premium VPK180 and Versal AI Core VCK190 evaluation kits and associated tools. The documentation for the new VP2502 and VP2802 is available now, the optimized tools will be available in the second half of 2022, and the new devices will be available starting in the first half of 2023. Pricing was not available.

Like the other Versal ACAP products, the new Versal Premium with AI could be adapted to other applications, but we are also likely to see other specialized/optimized designs from AMD. The combination of multiple different types of processing engines, especially the Adaptive FPGA and AI engines, with different configurations of I/O allows for almost endless combinations that can be optimized for a wide variety of applications.

Advertisement

Leave a Reply

Your email address will not be published.