SiFive will soon demo a Linux-driven, RISC-V based “FU740” SoC powering a PC and will unveil a new superscalar AI core. Interest in RISC-V is likely to increase if Nvidia completes its $40 billion acquisition of Arm.
At the Linley Fall Virtual Processor Conference starting Oct. 20, SiFive CTO Yunsup Lee will unveil a Linux-driven Freedom U740 (FU740) SoC based on its high-end U740 core and demonstrate it running on “the world’s first RISC-V PC.” In addition, SiFive Chief Architect and RISC-V architecture co-developer Krste Asanovic (see image at right) will announce a SiFive Intelligence AI core that appears to be related to the new U8-series.
As semiconductor companies weigh the competitive threat from Nvidia’s pending, $40 billion acquisition of Arm, they are likely to accelerate their transition to the open source RISC-V architecture. For the record, Nvidia claims it will remain vendor neutral in its licensing of Arm cores (see farther below).
SiFive preps a RISC-V “FU” to take on proprietary chip industry
At the Linley event, SiFive will announce its first Linux-ready system-on-chip (SoC) since the U5-series based Freedom U540 (FU540) SoC, which powers its HiFive Unleashed SBC. The Freedom U740 SoC will be based on its 64-bit, Linux-capable U7-series of Cortex-A55 like cores, which feature 2.5 DMIPS/MHz and 4.9 CoreMark/MHz performance. Other RISC-V based SoCs include Microchip’s FPGA enabled PolarFire SoC, which is also based on the U54 cores.
Core IP 20G1 improvements vs. earlier 19.08 release (left) and U74-MC block diagram
(click images to enlarge)
The U7-series currently includes the single-core U74 and the quad-core U74-MC. The MC model adds a fifth real-time monitor core based on its similarly, 64-bit Core IP 7 Series architecture S7 MCU core. In July, SiFive upgraded its U7- and earlier U5-series CPUs with a “20G1” release that boosted performance and reduced power.
The FU740 SoC will offer “modern PC expansion capabilities and form factor with a suite of included tools to facilitate broad professional software development,” says SiFive. The SoC will also “enable professional developers to create RISC-V applications in a bare-metal environment, from OS to end-user application” and support “porting of existing applications.” The SoC will be demo’d on a Linux-based PC prototype, which suggests the possibility that a mainboard such as the HiFive Unleashed may also be introduced.
While the FU740 name suggests the SoC is based on the U74 or U74-MC, SiFive mentions a “mix+match core complex,” which is the name of the heterogeneous multi-core technology that SiFive announced as part of it next-gen, Cortex-A72 like U8-Series. Much like Arm’s Big.Little with DynamIQ, mix+match is likely being backported from the upcoming, 7nm capable U8, which was announced last October.
Given the mention of mix+match and the intention to demonstrate the FU740 on a PC, the FU740 may well be based on new octa-core capable U7-series variants. The only multi-core U7 core — the U74-MC — offers four identical mainstream cores along with the S7 companion core, so there is nothing to mix and match.
Block diagram for a hypothetical U84-based SoC that incorporates U74 and S2 cores using SiFive Mix+Match
(click image to enlarge)
SiFive may intend to launch two new octa-core capable U7 core variants with different clock speeds. In introducing the U8-series, SiFive posted a block diagram for a hypothetical octa-core SoC that uses mix+match to combine four, up to 2.6GHz U84 cores with four U74-MC cores and an S2 companion core.
SiFive Intelligence AI SoC design with vector extensions
SiFive will also demonstrate an upcoming SiFive Intelligence AI SoC design that combines “scalable vector processing with a Linux-capable superscalar multi-core processor.” The design “features a complete implementation of the latest RISC-V Vector (RVV) extension,” says SiFive.
SiFive benchmarks showing U84 performance compared to U74 and U54
(click image to enlarge)
Although SiFive makes no mention of the U8-series, superscalar and vector processing were first announced with the U8. At the time, SiFive announced that an upcoming U87 core that would follow the initial U84 core would add the RVV extensions for vector processing. Earlier this month, SiFive joined with the Barcelona Supercomputing Center to release an API based on RVV for vector processing applications ranging from “AI to signal processing and scientific research.”
Last month, SiFive received $60 million in additional funding from SK Hynix and Aramco. The company also announced an OpenFive business unit for enabling “customizable, silicon-focused solutions with differentiated-IP.”
ARM licensees weigh Nvidia claims it will play fair with licensing
Nvidia’s blockbuster announcement yesterday of its agreement to purchase Arm from SoftBank for $40 billion may be welcome by America-firsters who would rather see the world’s leading CPU IP licensing firm folded into a US rather than a Japan-based company. Yet Arm licensees in the U.S. and around the world must now worry that Arm IP is controlled by a chip-making rival rather than by a software and investment company. Those concerns could lead to increasing investment in RISC-V designs.
Jetson Xavier NX
The merger faces intense regulatory review and may take up to 18 months to be completed. As part of the deal, SoftBank will hold an under 10 percent stake in Nvidia. Known for its AI-enabled GPUs an automotive technology, Nvidia also sells a line of Jetson compute modules that combine embedded versions of its GPUs with Arm-licensed CPU cores. The high-end Jetson AGX Xavier targets applications ranging from robotics to self-driving cars.
Nvidia says it plans to offer some of its GPU technology to Arm for licensing. Arm also offers its own mobile Mali line of GPUs.
As noted in an ArsTechnica analysis of the proposed acquisition, which is one of the biggest in history, Nvidia claims it will retain Arm’s vendor neutral licensing model and will maintain and further invest in Arm’s Cambridge UK facilities. Yet the story also notes that Nvidia has made a lot of enemies over the years with rival chipmakers and customers, including Apple.
The story mentions Nvidia’s fractious relationship with Linux, which is by far the most dominant OS on Arm, and not only via the Linux-based Android. Nvidia GPUs are part of Linux-based servers in addition to the Jetson connection. For years, Nvidia was infamous in the Linux community for shipping its GPUs as proprietary binary driver “blobs” and otherwise making life tough for Linux developers.
In 2012, Nvidia received a public scolding from Torvalds as being “one of the worst trouble spots we’ve had with hardware manufacturers.” He went on to give a middle finger salute to the company, with a four-letter oath to match. Since 2014, when Nvidia released some open source driver code for the Tegra K1 SoC, its Linux support has begun to improve.
Yesterday, Reuters reported on a backlash to the acquisition brewing among Arm licensees. Chinese vendors such as Huawei are particularly concerned, despite Nvidia’s assurances that because it is remaining in the UK, Arm will retain independence from US trade policy. Yet, the story also mentions South Korea’s Samsung and the US-based Qualcomm as other vendors that are worried about potential unfair advantages given to Nvidia, especially in the automotive market.
Both reports, as well as others, have suggested that if Nvidia’s acquisition is approved, chipmakers will accelerate their development of open RISC-V designs. Arm has responded to the RISC-V threat by opening up its proprietary architecture somewhat with its Custom Instructions extension and Arm Flexible Access licensing program. Yet Arm still offers a mostly proprietary licensing scheme.
MIPS goes to China
Wave Computing, which acquired the MIPS portfolio from Arm’s mobile GPU rival Imagination Technologies, which was later acquired by China-backed Canyon Ridge, responded to the RISC-V threat by releasing some open source MIPS ISA. Yet, the company recently sold its MIPS portfolio through a complex series of deals that eventually ended up at China-based CIP United. One of CIP’s customers is Huawei.
SiFive will announced its FU740 SoC and SiFive Intelligence AI SoC at The Linley Group’s Linley Fall Processor Conference 2020, which will be held with virtual presentations during Oct. 20-22 and Oct. 27-29. Other new products will be introduced at the event from Brainchip, CEVA, Synopsys, and Tenstorrent.
Note: Circuit Cellar did an interview with Krste Asanovic, SiFive Chief Architect and RISC-V architecture co-developer, back in July 2017