客户登录
解决方案
解决方案
SiFive Data Center Innovation

Read about SiFive’s data center strategy

Leadership in Embedded Markets

Learn about our 4th generation Essential products

产品
核心 IP
SiFive 系统 IP
开发板
软件与工具
产品
Essential 系列
Intelligence 系列
Performance 系列
Automotive 系列
产品
产品
现已上市
旧版开发板
All Boards
BigSky Development Server

New platform enables RISC-V in the Datacenter

HiFive P550 Review

Level1Techs reviews the Premier P550

产品
软件与工具
Software Expansion

The RISC-V Software ecosystem is expanding rapidly

Premier P550 Boards

Learn about this premier development platform

资源与支持
资源与支持
RISC-V Software Expansion

The RISC-V Software ecosystem is expanding rapidly

Premier P550 Boards

Learn about this premier development platform

关于我们
关于我们
Cycuity Partnership

Collaborating with Cycuity and BAE Systems means better security for all

SiFive’s New Look

Read about our new visual identity

SiFive Accelerator Control Unit (ACU) Solutions

Offload host CPUs, optimize data movement, & future proof custom AI engines with a programmable RISC-V based control layer

Intelligent Orchestration for Custom AI

In addition to providing Edge to Datacenter inference solutions, SiFive Intelligence IP is designed to be the ideal Accelerator Control Unit (ACU), combining programmable RISC-V scalar and vector compute processing and control capabilities, with direct high speed, high bandwidth connectivity for optimized data movement, allowing designers to focus on their custom accelerator’s core data processing potentials.

An ACU reduces, and can eliminate or replace, the system level host processor’s low level orchestration burden and provides better support for evolving workloads, allowing designers to focus on their custom accelerator’s core data processing capabilities now and in the future.

The Evolution of Accelerator Control

Traditional custom accelerators rely on fixed state machines and shared system buses, leading to high host CPU overhead, memory bottlenecks, and rigid hardware limitations.

Traditional Approach has Challenges 

Legacy system architectures are designed around a host processor that manages a custom accelerator through a dedicated system bus, mapped control registers, state machine, and Direct Memory Access (DMA).

Challenges:

  • Fixed data manipulation & processing capabilities
  • Large data movement around common bus 
  • Long latency for communication & data sharing
  • Complex and fixed control logic
  • Nonstandard to program and hard to updated

Traditional Approach

Intelligent ACU Orchestration 

The SiFive Intelligence™ Accelerator Control Unit (ACU) solution shifts system management to a dedicated, programmable RISC-V front end controller. By taking over task scheduling, resource allocation, command dispatches, and data formatting, the ACU enables your host processor to focus on primary application execution or the ACU can act as the host processor too,  as the SiFive Intelligence series ACU solution is RTOS/Linux capable.

SiFive ACU Solution Benefits

SiFive Intelligence cores offer additional compute functionality with attached vector units and can offload or provide new functionality not planned when the custom accelerator was originally developed.

Key Advantages

  • Future-Proof Flexibility: Replace rigid state machines with a programmable RISC-V core. Deploy software and firmware patches to fix bugs or add functionality without costly hardware re-spins.
  • Complete Host Offloading: The ACU acts as an intelligent orchestration layer, allowing the data throughput of the accelerator engine to operate completely independent of the main processor.
  • Fallback Vector Assist: SiFive Intelligence cores feature vector units to handle nonmatrix operations like activation functions, Softmax, or entirely new algorithms your custom accelerator wasn’t originally designed for.
  • Open Ecosystem: Built on 100% RISC-V ISA compatibility, allowing you to leverage industry supported, open standard software stack to dramatically reduce time to market.

SiFive Intelligent Accelerator Approach

SiFive Benefits for Custom Accelerator Design

Classic architectures throttle performance by forcing massive data transfers and control commands over a single shared bus. SiFive IP eliminates this bottleneck with three complementary, direct connections to the custom accelerator:

  • VCIX (High-Bandwidth Data Movement): Directly interfaces to the SiFIve Intelligence core vector registers. It streams packed tensor and matrix datasets straight into the custom hardware accelerator at core clock frequency to keep heavy compute engines constantly fed.
  • SSCI (Low Latency Control): Directly interfaces to the SiFIve Intelligence scalar CPU execution pipeline. It keeps your engine responsive by dispatching custom scalar instructions and handling data-dependent decisions with single-cycle round trips.
  • CLP (Core Local Port): Bypasses the L2 memory to cache subsystem to provide the AI core with high bandwidth, low latency access to local or customer RAM.

Intelligent Orchestration for Custom Accelerators 

ACU Rule #1 Keep the accelerator fed

ACU Rule #2 Keep control responsive

When paired with SiFive Intelligence X100, X200, or X300 series cores, the SSCI, VCIX, and CLP interfaces transform the custom accelerator engine to an Intelligent Orchestrated Engine that decouples the custom accelerator's data throughput and execution from the main host processor.

How the Intelligence series orchestration works:

  1. Move or stage the data to the Custom Accelerator Engine. The ACU can access system memory through its own memory path for bulk data movement. When data needs to pass through the ACU, its vector unit can stage or transfer it over VCIX. The ACU can also use CLP to reach local accelerator memory.
  2. Launch the operation. Once the required data is ready, the ACU scalar core sends an SSCI custom instruction to configure the accelerator and trigger the workload.
  3. Keep useful work moving. While the custom accelerator engine handles its primary operation, the ACU can perform additional vector work such as data formatting, activation functions, Softmax, or other algorithms not built into the custom engine and coordinate subsequent work.

SiFive Intelligent Orchestration

Size the Right ACU for Your Workload

Sizing the right SiFive Intelligence core for an Accelerator Control Unit (ACU) solution depends on balancing three primary factors:

  • Data Movement Bandwidth
  • Control Complexity
  • Auxiliary/Fallback Compute Needs

The vector length (VLEN) and datapath width (DLEN) of the selected core determine both the data movement bandwidth available over co-processor interfaces and the onboard computational capability of the ACU.

  • X100 Series (Control-Heavy & Compact): Ideal for accelerators needing minimal-latency instruction dispatches over SSCI without wasting die area. Runs efficiently on Tightly Coupled Memory (TCM/SRAM) with an RTOS.
  • X200 & X300 Series (Data-Heavy & Fallback Compute): Ideal for continuous, high-bandwidth input streaming. Wider VCIX datapaths (256-bit, 512-bit or 1024-bit VLEN) double or quadruple data transfer capacity. Features a full MMU and RVA23.1 profile support for Linux environments and robust vector units to handle heavy fallback workloads.

Ready to streamline your hardware integration, accelerate your design process and maximize system efficiency? Contact the SiFive sales team today to schedule an architecture review and explore how our Intelligence Series IP can accelerate your next datacenter, IoT, Automotive, Robotics or other Physical AI technology breakthrough.

Contact Sales Read Blog

SiFive Intelligence Family

X100 Series
Intelligence
  • 32-bit or 64-bit CPU
  • 128-bit Vector Length
  • SSCI and VCIX
View X100 Series
X200 Series
Intelligence
  • 512-bit Vector length
  • Single Vector ALU
  • SSCI and VCIX (1024-bit)
View X200 Series
X300 Series
Intelligence
  • 1024-bit Vector Length
  • Single/Dual Vector ALU
  • SSCI and VCIX (2048-bit)
View X300 Series

SiFive ACU Solutions: Learn More

SiFive Intelligence IP - Use Case: Accelerator Control Unit (ACU)
SiFive Intelligence IP - Use Case: Accelerator Control Unit (ACU)
SiFive 2nd Generation Intelligence Technology Explainer
SiFive 2nd Generation Intelligence Technology Explainer

Discover Customer Results

Case Study: Accelerator Control Unit Use Case

Tier 1 US Semiconductor Company Licensed SiFive X100 for Industrial Edge AI Applications.

Case Study: Google Datacenter

SiFive Intelligence X280 Gen 1 as AI Compute Host

Next Steps for Custom Accelerator Innovation

Optimizing custom accelerator control is essential for modern SoC success.

Ready to maximize your system's efficiency?

请阅读我们的《隐私政策.》，了解您的数据如何受到保护。