Open Source · Apache 2.0GitHub ->

RVDon · RISC-V Domain-specific Open Node

Standing on the shoulders of giants
Building our own acceleration silicon

Starting from the open-source Vortex RISC-V GPU, adding hardware-native instructions for the Pairformer module of AlphaFold3 / Protenix--triangle mask matrix multiplication and Flash Attention are no longer software simulations, but first-class citizens on silicon.

Vortex GPU+WGMMA Datapath+PF Extension=RVDon

WHY RVDon

Why We Must Do This

Not reinventing the wheel, but building our own telescope on the shoulders of open-source giants

Standing on Giants

Vortex is open source, RISC-V is open source, Protenix is open source. We use others' shoulders to build our own path--and then open source the path too.

Real Silicon, Not Slides

637 lines of VX_tcu_fa.sv, 3-stage pipeline, 16-entry LUT exp approximation--every line of RTL has passed Verilator simulation, 0/128 errors verified.

Open Source as Faith and Strategy

RTL is the blueprint, the chip is the product. Open-sourcing blueprints doesn't harm commercial interests--it eliminates trust costs, attracts FPGA validators, and builds academic influence.

ARCHITECTURE

Chip Architecture Stack

Three layers progressive--from general GPU base to domain-specific instructions

Vortex RISC-V GPGPU

Open-source RISC-V GPGPU Base

Multi-core GPU based on the Vortex open-source RISC-V GPGPU architecture--providing mature WGMMA datapath and thread scheduling infrastructure for RVDon.

L0 · GPU

TCU · WGMMA Pipeline

Tensor Core Unit · WGMMA

Tensor Compute Unit within Vortex Core, providing WGMMA (Warp Group Matrix Multiply-Accumulate)--RVDon PF extension reuses this datapath.

L1 · TCU

PF Extension ISA

Pairformer Extension ISA

PF_TMM / PF_TMM_INC / PF_FLASH_ATTN--natively supporting triangle mask matrix multiplication and Flash Attention in the TCU pipeline, eliminating redundant computation and software masking overhead.

L2 · PF ISA

PAIRFORMER EXTENSION ISA

PF Extension Instruction Set

Natively supporting Pairformer core computation in the TCU pipeline--replacing software masking, eliminating redundancy

Pairformer Core Computation Patterns

🔺

Triangle Multiplication

Triangular matrix multiplication + symmetry masking--half the computation is redundant on general GPUs, PF_TMM hardware triangle masking skips it directly

Triangle Attention

Triangular attention + causal masking + online Softmax--FA_SOFTMAX completes within the TCU pipeline, no global synchronization needed

funct3=3✅ RTL + SimX

PF_TMM

Outgoing triangle mask matrix multiplication (Triangle Multiplication Outgoing)

funct3=4✅ RTL + SimX

PF_TMM_INC

Incoming triangle mask matrix multiplication (Triangle Multiplication Incoming)

funct3=5✅ RTL 0/128

PF_FLASH_ATTN

Flash Attention (FA_MMA / FA_SOFTMAX / FA_UPDATE sub-operations)

funct3=-⏸ reserved

PF_SLOAD

Strided Load--reserved for future extension

PROGRESS

Development Progress

Every step passed simulation--0/128 errors means complete

Phase 1.0PF_TMM RTL0/128 errors
Phase 1.1VX_tcu_fa.sv 637 linesFA Online Softmax pipeline
Phase 2.1Pipeline timing alignmentfa_p_delay_pipe
Phase 2.3FP32 arithmetic fixfp32_sub + fp32_floor_int
Phase 2.4FA_SOFTMAX validation0/128 errors ✅
Phase 3FPGA prototype validationNext step

OPEN SOURCE

Open Source Strategy · Open Design · Free Use

All Apache 2.0--design, code, documentation, freely usable, modifiable, and commercializable by anyone

RTL Code + SimX Model

Apache 2.0

RVDon PF extension RTL (VX_tcu_fa.sv etc.) and SimX behavioral model (tcu_unit.cpp) on top of Vortex, open-sourced under Apache 2.0. Anyone can use, modify, and commercialize--but must retain attribution and patent grant terms.

PF Intrinsics + Test Framework

Apache 2.0

vx_pf.h header defining PF extension intrinsics (pf_tmm_sync / pf_tmm_inc_sync / fa_mma_sync etc.) and regression test framework, also released under Apache 2.0--complete toolchain open source.

Architecture Docs + Design Reports

Apache 2.0

ARCHITECTURE.md specification (PF extension ISA, register mapping, programming model) and docs/ design documents, phase reports all open source--letting the community understand the intent of every signal line.

Mandatory Attribution Clause

All derivative works (hardware, software, documentation) must include the following attribution in NOTICE / README / product documentation:

Based on DiVo Gen²AI RVDon (RISC-V Domain-specific Open Node) by 王觉菊 Jueju.Wang, https://github.com/sealionking/rvdon

RVDon, DiVo, Gen²AI are project self-names--the code is yours, the attribution is ours.

BUSINESS

Two-Layer Value Loop

Open source builds brand -> consulting earns revenue

Open Source Layer

RVDon PF extension RTL + SimX behavioral model + intrinsics + regression tests--hardware developers can freely use, modify, and build their own Pairformer accelerators

Community co-building, brand penetration

Consulting Layer

Chip design consulting--from architecture selection to PF extension RTL customization, helping clients optimize Pairformer / Attention acceleration on the RVDon base

Paid service, 1-on-1 customization

ECOSYSTEM

From Silicon to Targets

RVDon is not an isolated hardware project--it is the bottommost layer of the DiVo bio-computing full stack

RVDon PF Extension->Protenix Structure Prediction->pMHC 3D Structure->Neoantigen / CAR-T

Pairformer is the core module of AlphaFold3 / Protenix. RVDon's PF extension instructions accelerate triangle matrix multiplication and Flash Attention at the silicon level, directly boosting Protenix inference throughput--and Protenix is the engine for DiVo's neoantigen pipeline S5 (pMHC structure prediction) and CAR-T pipeline CT5 (structure validation). From hardware instructions to patient targets, this is a complete causal chain.

Build Your Device with RVDon

Take the open-source design to build your own hardware--or have us customize your domain-specific chip.