Two years ago, Intel acquired Altera, a FPGA chip supplier. What will such a big acquisition bring to Intel? Now, Intel is finally ready to show us its gains. FPGA, or field programmable gate array, is an integrated circuit that can be customized to perform specific functions. X86 only executes the x86 instruction set, and FPGA can be reprogrammed to perform the specified task. That’s why x86 processors are called general-purpose computing processors, and FPGAs are considered customizable.
The company’s strategy is interesting and seems to be putting itself into a competitive situation. In terms of large-scale floating-point operations, Intel has an Xeon Phi product line that competes with NVIDIA and AMD GPUs. Now, FPGA has also joined the battlefield for those large-scale floating-point computing use cases. Like GPU, FPGA will be applied in two ways: inline and offload. Inline means that the data first passes through the CPU and then transfers to the FPGA for processing. Offload, also known as look aside, means that the CPU is not affected, and the data is processed directly in and out of FPGA.
For specific tasks, FPGA can show excellent performance. Now Intel has positioned Altera FPGA as a coprocessor, and admits that they will compete with Xeon pHi in some aspects, but in some specific tasks, FPGA is more general and suitable than phi and GPU. According to Bernhard friedbe, senior director of the programmable Solutions Group for Intel Software solutions.
“The advantage of FPGAs is that GPUs only work in some areas, not all. If you look at the usage models of inlining and uninstallation, most of them are limited to uninstallation. Therefore, you can use FPGA to cover a wider application space. ” He said.
The integrated solution provides close coupling between CPU and FPGA, with very high bandwidth, while the external PCI-E card is not so tightly coupled. Integration is ideal for applications with ultra-low latency and high bandwidth, says friedbe. “Most of the differences between integration and dispersion are due to system architecture and data movement. In a data center environment, there are many different workloads running, and no one wants to bind them to a particular application, “he said.
Friedbe points out that the more specialized you are, the more performance you can squeeze out of the accelerator. As a multi-functional accelerator, FPGA will achieve good results in some applications. The essence of FPGA is highly parallel and programmable, which is suitable for accelerating the parallelizable workload. These include data analysis, artificial intelligence (AI) and machine learning, video conversion, compression, security, financial analysis, and genomics.
Intel has adopted a two pronged strategy, offering two hybrid cpu-fpga processors – such as its desktop CPU with GPU on the mold – and discrete aria or Stratix brand FPGA devices on PCI-E cards.
The hybrid cpu-fpga device will be based on skylake CPU and aria 10 FPGA, and will use faster UPI (ultrapath interconnect) link – the successor of Intel’s QPI (QuickPath interconnect). In addition to running at 9.6gt/s or 10.4gt/s, it will be more efficient than QPI because it will support multiple requests per message.
Intel also offers a complete developer toolset and API to design applications for integrated and discrete products, using the same tools, accelerators, and libraries. All of them are written in OpenCL, a C-like language. “The key is standardization and open source. This is a new generation of forward compatible processors, easy to migrate, and provides an abstraction for FPGA developers to target a larger user base. ” Friebe said.