All News | Boards | Chips | Devices | Software | Archive | About | Contact | Subscribe
Follow LinuxGizmos:
Twitter Facebook Pinterest RSS feed
*   get email updates   *

SiFive Intelligence platform tapped by Tenstorrent and Renesas

Apr 22, 2021 — by Eric Brown 1,165 views

[Updated: 11AM] — SiFive announced that AI chip startup Tenstorrent will license a new 64-bit SiFive Intelligence X280 CPU for its Tensix cores and that Renesas will adopt SiFive Intelligence for its automotive processors.

Tomorrow at the Linley Spring Processor Conference, RISC-V IP designer SiFive will unveil the X280, the first of the SiFive Intelligence family of processors announced last September. Meanwhile, SiFive has announced two early adopters of SiFive Intelligence. AI chip and software startup Tenstorrent will license the SiFive Intelligence X280 as a platform for its Tensix NPU. Renesas says it will license SiFive Intelligence for its automotive processors but did not specify the X280. (See farther below for more on the Tenstorrent and Renesas news.)

The news follows other recent partnership announcements including SiFive’s participation in the new Intel Foundry Services for fabricating silicon for startups. There is also a deal with the DARPA in which the agency will license SiFive’s RISC-V IP for its DARPA Toolbox Initiative. In addition, SiFive’s new silicon manufacturing business unit OpenFive announced it has taped out a SoC with TSMC’s 5nm fabrication technology (see farther below.)


SiFive Intelligence

The SiFive Intelligence platform is based on new SiFive RISC-V Vector processors with AI ISA extensions. The platform combines scalable vector processing with a Linux-capable superscalar multi-core processor and features the latest RISC-V Vector (RVV) RV64GCV extension. The RV64GCV ISA, Which is also being adopted by other RISC-V processors such as the Allwinner/Alibaba XuanTie C906, enables more AI processing on the CPU, reducing the need to continually offload processes to an NPU or GPU.

SiFive Intelligence is not a neural processor itself but features a “differentiated” software toolchain for developing “scalable solutions for AI and ML applications.” A SiFive rep referred to the platform as offering a “software-first” approach, which “makes it easier for software to communicate with hardware, a common hurdle for developers working to stay ahead of the AI innovation curve.”

The first SiFive Intelligence core was the VIU75 design announced last December. Based on the same, Linux-ready Cortex-A55 like U7 core architecture found on the FU740 and some other SiFive chip designs, the VIU75 offers an 8-stage dual-issue in-order pipeline with a decoupled vector unit based on RV64GCV ISA.

The FU740 is used on the company’s HiFive Unmatched dev kit. There is also an upcoming U8-series, which may be the foundation for the X280.

VIU75-MC block diagram
(click image to enlarge)

The VIU75 core is available in a VIU75-MC SoC design with up to 4x of the vector-enabled VIU75 Intelligence cores and an optional SiFive S7 MCU. Applications include vision and machine learning, natural language processing, SLAM Processors, and sensor fusion. The Register published an interesting report on the VIU75 back in December.

Tenstorrent licenses X280

The SiFive Intelligence X280 processor could appear first on an AI training and inference SoC featuring Tenstorrent’s Tensix AI cores. Tensix is aimed initially at the datacenter, but with versions planned for “near edge” and “micro edge” devices.

Tenstorrent Grayskull platform (left) and its software architecture
(click images to enlarge)

Tensix “accommodates the exponential growth in the size of machine learning models while offering best-in-class performance,” says SiFive. “The combination of Tenstorrent’s Tensix processing cores alongside the SiFive Intelligence X280 processor reflects the growing trend of heterogeneous multi-core compute to stay ahead of the requirements of next-generation AI workloads.”

Tensix has a programmable architecture that supports fine-grain conditional execution, dynamic sparsity handling, and “an unprecedented ability to scale via tight integration of computation and networking,” says Tenstorrent. Other features include compiler based flexible parallelization and fine-grain adaptation of compute graph at runtime. Tenstorrent claims that Tensix enables AI systems that “go beyond pattern recognition and into cause-and-effect learning.”

Tensix is initially being deployed in a Grayskull reference platform built around a custom NOC (Network on Chip) aimed at the datacenter. Grayskull incorporates an unspecified number of Tensix cores to provide up to 368-TOPS AI performance at 65W. The platform is touted for its “unprecedented multi-cast flexibility and low software overhead data transfer.”

Each Tensix core has a matrix compute engine, a SIMD engine, and a C++ programmable, multi-threaded front-end. Grayskull adds to this with 120MB on-chip SRAM, 8x LPDDR4 channels, and 16x PCIe Gen4 lanes.

Stated Tenstorrent President and CTO Jim Keller: “The Tenstorrent architecture addresses the growing demands that come with data-written code as part of Software 2.0.” Keller joined the Toronto based company in January after holding an SVP a position at Intel Silicon Engineering Group. He was formerly at Tesla and AMD.

Renesas taps SiFive Intelligence for car SoCs

Yesterday, SiFive and Renesas announced a strategic partnership to jointly develop next-generation, high-end RISC-V solutions for automotive applications. As part of the partnership, SiFive is licensing the use of its RISC-V core IP portfolio to Renesas including its SiFive Intelligence platform.

R-Car Starter Kit

Renesas, which is known to LinuxGizmos readers primarily for its Arm-based IoT SoCs such as the RZ/G2 and new, lower-end RZ/G2L, manufactures a variety of automotive processors. These include SoCs, MCUs, and analog and power products for ADAS, Autonomous Driving (AD), Electric Vehicles (EV), and Connected Gateway (CoGW) products.

SiFive Intelligence will likely be used for a successor to Renesas’ top-of-the-line R-Car H3 SoC platform. The Linux-driven H3 SoC combines 4x 1.5GHz Cortex-A57 cores, 4x 1.2GHz Cortex-A53 cores, and a PowerVR GX6650 GPU, and features ISO 26262 safety compliance.

OpenFive tapes out 5nm SoC

On April 13, SiFive’s new OpenFive silicon manufacturing business unit announced the successful tape-out of an upcoming SoC using TSMC’s 5nm N5 process. The 32-bit SoC, which is built around SiFive’s high-end, Cortex-M7 like E76 MCU, is aimed at High Performance Computing (HPC)/AI, networking, and storage solutions.

The 5nm SoC features an up to 7.2Gbps ready OpenFive High Bandwidth Memory (HBM3) IP subsystem, as well as D2D I/Os. Samples are expected to be available by the end of the quarter.

DARPA to license SiFive IP

On Mar. 31, SiFive and U.S. Defense Advanced Research Projects Agency (DARPA) announced they had entered into an open licensing agreement for using SiFive’s 64- and 32-bit RISC-V IP in future DARPA Toolbox projects.

DARPA Toolboxis an agency-wide effort that offers open licensing opportunities with commercial technology vendors to the researchers behind DARPA-funded programs. Accepted participants receive access to commercial vendors’ technologies and tools via pre-negotiated, low-cost, non-production access frameworks and simplified legal terms. Commercial vendors, meanwhile, are said to benefit from the opportunity “to leverage the agency’s forward-looking research and a chance to develop new revenue streams based on achievements developed with their technologies.”

Further information

SiFive Intelligence is available now, says SiFive. More information should emerge soon at SiFive’s website.

(advertise here)

Print Friendly, PDF & Email

Please comment here...