RISC-V · Compute Early access

C0

A compact, Linux-capable RV32IMAC core running at 650+ MHz in TSMC 28HPC+.

CLINTPLICICCMUARTC0 coreRV32IMACPMPDCCMC0 · RV32IMAC63,894 cells, placed and routedCLINT · Core-local interruptorTimer and software interruptsPLIC · Platform-level interrupt controller31 external interrupt sourcesICCM · Instruction closely coupled memoryFetch only, over AXI4 or AHBUART · Universal asynchronous receiver-transmitterOne-word console portCore · RV32IMACOne hart, in order, up to eight pipeline stagesPMP · Physical memory protection16 entries: OFF, TOR, NA4 and NAPOTDCCM · Data closely coupled memoryLoads and stores, over AXI4 or AHB
CoreMark/MHz
3.43 CoreMark/MHz Official sources, rv32imac
DMIPS/MHz
1.76 DMIPS/MHz Official sources, rv32imac
MHz
650+ MHz TSMC 28HPC+ Highest-performance configuration. Leaner ones run at 1+ GHz.
mm²
0.027 mm² Core only, TSMC 28HPC+
Fully compliant
uClinux Fully compliant Linux 6.12, no MMU

Architecture

An in-order core with a configurable pipeline of up to eight stages. Closely coupled instruction and data memories sit on either side over AXI4 or AHB, with the standard RISC-V timer and interrupt controllers alongside.

C0 block diagram: the core contains fetch, decode, execute, load/store, register file, CSRs, PMP and interrupt control. The ICCM feeds fetch and load/store reaches the DCCM, both over AXI4 or AHB, and the CLINT, PLIC and UART sit on the device map.C0 · RV32IMAC · ONE HART · IN ORDERICCMfetch onlyDCCMload and storeFetchDecoderegister readExecuteLoad / storeRegister file32 × 32-bitCSRsM and U modePMP16 entriesInterruptsdirect · vectoredAXI4 / AHBAXI4 / AHBdevice mapCLINTtimer · software interruptsPLIC31 external sourcesUARTone-word port

Swipe sideways to see the full diagram.

Specifications

ISA
Full RV32IMAC: base integer, multiply and divide, atomics and compressed instructions. Zicsr and Zifencei included.
Pipeline
One hart, in order, single issue. Depth is configurable, from a very short pipeline up to eight stages.
Multiply
Configurable, from a small iterative multiplier up to a four-stage pipelined one.
Privilege
Machine and user mode.
Protection
16 PMP entries, with OFF, TOR, NA4 and NAPOT matching.
Interrupts
Machine software, timer and external. Direct and vectored modes, and wfi.
Timer
Standard RISC-V CLINT, with the machine timer and software interrupts.
External interrupts
PLIC with one machine-mode context and 31 sources.
Interfaces
AXI4 or AHB (AMBA) on the instruction and data memory ports.

Performance

3.43

CoreMark/MHz

1.76

DMIPS/MHz

Measured on C0 running on an AMD Alveo U280 FPGA, and in simulation, with the official Dhrystone and CoreMark sources compiled for rv32imac.

Silicon

We take C0 from RTL all the way to a routed layout.

650+ MHz is C0 in its highest-performance configuration, which is also its slowest clock. Leaner configurations, with deeper pipelines up to eight stages, run at 1+ GHz.

C0 standard-cell placement, filling the die.
C0 placed and routed: 63,894 cells.
Flow
RTL to GDS, synthesized, placed and routed
Frequency
650+ MHz in TSMC 28HPC+, typical corner
Area
0.027 mm² core-only in TSMC 28HPC+, 9-track

Runs uClinux

uClinux booting on C0, on an AMD Alveo U280.

Real time, recorded from the serial console.

Example SoC

Tiny-Vedas

Our open-source SoC for edge vision, built around C0. A Zve32x vector co-processor and an 8×8 matrix engine sit beside the core, sized for the cameras on drones and other small devices. It runs YOLOv3-Tiny, and we’ve tested it on an AMD Alveo U280.

Hover over a block to see what it does.

Read the Tiny-Vedas paper
ICCMC0RV32IMACVECTORCO-PROCESSORXDAIDCCMPCIeGEMM 8×8HBMUARTTiny-Vedas · Edge vision SoCBuilt around C0, runs YOLOv3-TinyICCM · Instruction memoryHolds the compiled modelC0 · RV32IMACRuns the model and drives the acceleratorsVector co-processor · Zve32xActivations and poolingXDAI · Accelerator interconnectAXI fabric to memories and acceleratorsDCCM · Data memoryWorking tiles for each layerPCIe · Host linkLoads code, weights and framesGEMM engine · int88×8 matrix multiply for the convolutionsHBM · High-bandwidth memoryWeights and spill spaceUART · ConsoleConsole and end-of-test

Roadmap

What’s coming to C0.

  • Bit manipulation
  • Single- and double-precision floating point (F and D)
  • RISC-V debug module

Supervisor mode and multiple harts will come with the larger cores in the range.

Early access

Early-access customers get C0 first and a say in what lands next on the roadmap.

About the numbers

  • CoreMark and Dhrystone were run on C0 on an AMD Alveo U280 FPGA and in simulation, using the official sources compiled for rv32imac.
  • Frequency and area are for TSMC 28HPC+, typical corner, with a 9-track library. The area covers the core on its own, without the PMP or memories.

Early access

Get C0 first

Tell us a little about your project. We reply within 24 business hours, and an NDA is available.