Nvidia caused a shift in high-end computing more than a decade ago when it introduced its general-purpose GPUs and CUDA development platform to work with CPUs to increase the performance of compute-intensive workloads in HPC and other environments and drive greater energy efficiencies in datacenters.
Nvidia and to a lesser extent AMD, with its Radeon GPUs, took advantage of the growing demand for more speed and less power consumption to build out their portfolios of GPU accelerators and expand their use in a range of systems, to the point where in the last Top500 list of the world’s fastest supercomputers, 86 used Nvidia GPU accelerators.
Intel over the years has looked to extend its reach into the accelerator space. Its attempt at developing GPUs through what was called the Larrabee project, and that evolved around 2010 into its x86-based many-core initiative that became Xeon Phi coprocessors that initially were used as offload engines, and later as standalone processors. As we at the The Next Platform have noted, the initial “Knights Landing” Xeon Phis began shipping in 2015, but last year Intel ended development of the coprocessor versions following customer demand. At the same time, Intel has since created its own GPUs that it uses with its CPUs for a number of systems, including low-end servers and workstations, but not much beyond that.
However, Intel made a major step in the accelerator game with its 2015 $16.7 billion acquisition of Altera, a company that made field-programmable gate arrays (FPGAs), chips that can be programmed and reprogrammed on the fly to adapt to whatever workloads that are running on the system at that moment, a feature highly valued by hyperscalers and cloud providers. Intel had a partnership dating back to 2013 with Altera, one of two dominant FPGA makers – the other being Xilinx, with Lattice Semiconductor being a distant third – and Intel saw the programmable chips as a way of making sure that tasks getting offloaded from its Xeon processors are moved to another Intel chip rather than GPUs from Nvidia or AMD.
Over the past couple of years, Intel engineers have worked to incorporate the Altera technology into the Intel portfolio. In 2016, Intel began shipping a development multi-chip module that paired a 14-nanometer Xeon “Broadwell” CPU with Arria10 FPGAs from the Altera acquisition. Last year the chip maker introduced a programmable acceleration card with the Arria10 GX FPGA and the Acceleration Stack for Intel Xeon Scalable processors with FPGAs (see below), a collection of firmware, software, APIs and tools aimed at making it easier to develop for FPGAs and deploy them in the datacenter.
Over the past year, some enterprises have adopted FPGAs through the CPU-FPGA PCIe cards that could be used with Intel-based servers, according to Intel. However, adoption of the chip maker’s FPGA technology could take a leap forward now that Dell EMC and Fujitsu are incorporating the programmable chips into some of their mainstream servers.
Intel and the system makers announced the deal this week, marking the first time top-tier OEMs have adopted the chips for their servers. Intel’s Programmable Acceleration Card with the Arria 10 GX FPGAs and the acceleration stack are now available in Dell EMC’s 2U dual-socket PowerEdge R740 and R740XD (below) rack servers that are powered by Intel’s Scalable processors and will support up to four FPGAs and the R640, a 1U, two-socket system that will support one FPGA.
Dell EMC told The Next Platform that embracing the programmable chips is a way of offering customers a broader range of accelerators – they also support GPUs – for compute- and data-intensive workloads. Specifically, FPGAs will be useful for mainstream datacenter environments that are struggling with the huge amounts of data that is more quickly being generated through such trends as the Internet of Things (IoT) and emerging workloads like machine learning. The accelerators bring versatility through the rapid programmability they offer, which can lead to higher performance and greater power efficiency, the OEM said.
For its part, Fujitsu is making Intel’s FPGAs available as “early access for priority customers,” with available in its Primergy RX2540 M4, a 2U dual-socket rack server.
Intel said that while Dell EMC and Fujitsu are the first major OEMs to incorporate the FPGAs into their systems, they won’t be the last. The chip maker said it is working with other system makers, system integrators and value-added resellers (VARs) and expect to make more announcements in the months to come.
The performance and efficiency boosts that come with FPGAs will make them attractive to other vendors. Intel said tests with the FPGAs have shown that the accelerators can improve the performance in a range of workloads, including a two-fold speedup of options trading simulations. For relational databases using Swarm64 acceleration technology, FPGAs can deliver more than 20 times faster real-time data analytics, two times faster data warehousing and three times the storage compression, according to the chip maker.
Beyond financial acceleration and data analytics, Intel is targeting a number of workloads for its FPGAs, including artificial intelligence (AI), genomics, cybersecurity and video transcoding.