cageymaru
Fully [H]
- Joined
- Apr 10, 2003
- Messages
- 22,363
Xilinx is known as the creator of the field-programmable gate array (FPGA), but now they want to conquer the artificial intelligence (AI) field. Today they unveiled the Adaptive Compute Acceleration Platform (ACAP) which can be used with AI and many other technologies. As the largest player in the FPGA market, Xilinx has been competing against Intel and NVIDIA for awhile, but things will change in 2019. The new 7nm ARM-based Xilinx Versal family will be 2x - 8x times faster and 4x more efficient than Nvidia GPUs today in the inference benchmark. Versal completely destroys the Intel offerings as it is 43x - 72x faster than Intel Xeon processors in the same inference benchmark. The Xilinx Versal family will use a comprehensive sets of tools and software that will be revealed next year.
Xilinx also announced a partnership with AMD to combine two 32C/64T AMD EPYC 7551 server CPUs with eight of the freshly-announced Xilinx Alveo U250 acceleration cards for high-performance, real-time AI inference processing. The result is a world-record* 30,000 images per-second inference throughput on GoogLeNet; a widely used convolutional neural network. The Xilinx ML Suite allows compatibility with such machine learning (ML) technologies such as TensorFlow.
The Scalar Engines are built from the dual-core Arm Cortex-A72, providing a 2X increase in per-core single-threaded performance compared to Xilinx's previous-generation Arm Cortex-A53 core. A combination of advanced architecture and power improvements from the 7nm FinFET process yield a 2X improvement in DMIPs/watt over the earlier 16nm implementation. The ASIL-C certified(1) UltraScale+ Cortex-R5 Scalar Engines migrate forward to 7nm with additional systemlevel safety features based on learning from Xilinx's current automotive volume deployments.
Xilinx also announced a partnership with AMD to combine two 32C/64T AMD EPYC 7551 server CPUs with eight of the freshly-announced Xilinx Alveo U250 acceleration cards for high-performance, real-time AI inference processing. The result is a world-record* 30,000 images per-second inference throughput on GoogLeNet; a widely used convolutional neural network. The Xilinx ML Suite allows compatibility with such machine learning (ML) technologies such as TensorFlow.
The Scalar Engines are built from the dual-core Arm Cortex-A72, providing a 2X increase in per-core single-threaded performance compared to Xilinx's previous-generation Arm Cortex-A53 core. A combination of advanced architecture and power improvements from the 7nm FinFET process yield a 2X improvement in DMIPs/watt over the earlier 16nm implementation. The ASIL-C certified(1) UltraScale+ Cortex-R5 Scalar Engines migrate forward to 7nm with additional systemlevel safety features based on learning from Xilinx's current automotive volume deployments.