Open GPUs: status in silicon
AbstractAn open driver can make a closed graphics chip work without revealing its design. Publishing the design opens another route, but a full GPU still has to come back from the fab working.
Your computer can run an open operating system and an open instruction set while the chip drawing its screen remains closed. No open GPU design was found in a phone, laptop or desktop sold in shops. The five published designs examined here have reached different stages. None has been reported working as a full GPU in silicon.
Documentation is not the design. Broadcom published a VideoCore IV guide, but Mesa’s V3D documentation says no public V3D 3.x or 4.x specification was released; access requires an NDA. An open driver does not make the hardware public.
Without a licence, a design is a document
tiny-gpu runs example addition and multiplication kernels in simulation. Layout files exist, but no attached shuttle submission was found and the Tiny Tapeout pages do not list it. Its repository still has no licence file.
A builder made a Tiny Tapeout adapter and clean layout, then requested the OSI-approved licence needed for submission. The owner allowed that submission to proceed, without adding a licence. The next builder still has no written grant to use or improve the design elsewhere.
The source is visible. Permission has not been published.
Open designs already run, on FPGAs
Vortex supplies a RISC-V GPGPU and OpenCL, Vulkan and HIP software under Apache 2.0, running on Altera and Xilinx FPGA cards. Its paper reports ASIC synthesis and layout of one core in an educational cell library, not a returned chip. An ASIC port also needs clock-tree, reset, power-management and memory work.
Ventus supplies a RISC-V GPGPU with OpenCL on large Xilinx FPGAs under Mulan PSL 2. Its conference synthesis report is not a tape-out. No fabrication submission was found for either.
Both licences include contributor royalty-free patent grants, terminated for a licensee suing over patents in the work. Neither grants contributors’ trade names. Apache permits naming the work’s origin; Mulan does not address it. No logo or conformance programme was found for either project. Under the three tests, free access and building do not remove the need to check the remaining terms.
The FPGA works. The ASIC still needs making.
Funding did not deliver a working GPU
Libre-SOC reports more than EUR 550,000 in NLnet grants for a hybrid CPU, VPU and GPU based on OpenPOWER with its own vector extension. Its source uses LGPL 3 or later.
Its submitted test chip contained a fixed-point OpenPOWER subset, not the GPU or the planned vector extension. A later project report still placed the dies with the packaging company. No subsequent working-chip report was found. A trademark claim over the Simple-V extension name was asserted on the mailing list, relevant to compatibility wording.
Funding, submission and a tested GPU are separate results.
A small design submitted to a shuttle
Borg combines tile-based graphics with a RISC-V core. Most hardware files carry CERN-OHL-S-2.0 markings, with other licences elsewhere in the tree. On an FPGA it renders vkcube through its Vulkan driver (Mesa-list report).
Its first chip entered Tiny Tapeout IHP 26a on IHP’s open 130 nm process. The shuttle page forecasts chips in February 2027; the project poster gave September 2026. Neither is a working-chip report. The submitted design runs a small processor’s 16 × 16 triangle animation, not the FPGA’s vkcube. Returned-chip testing remains pending.
CERN-OHL-S-2.0 supplies a royalty-free patent licence from each licensor. Distributors must provide complete source to recipients or tell them where to obtain it. The licence is available. The submitted GPU still has to return.
The route to silicon is already open
Two smaller rasterisers returned on Tiny Tapeout 7. One repository reports testing on the chip, with data errors at 50 MHz; the other shows no silicon test. This establishes observed silicon at a smaller scale, not operation of the full GPUs above.
Mesa lists no open GPU design among its supported systems. Vortex and Borg maintain Mesa forks; Borg’s author reported no concrete upstreaming plans. A fork can keep working if someone maintains and rebases it. Fabricating the chip does not provide that person.
| Design | Licence | Furthest stage |
|---|---|---|
| tiny-gpu | None | Simulation |
| Vortex | Apache 2.0 | FPGA; synthesis and layout of one ASIC core, no fabrication |
| Ventus | Mulan PSL 2 | FPGA; synthesis only for an ASIC |
| Libre-SOC | LGPL 3 or later | 180 nm test chip without the GPU; no working chip found |
| Borg | CERN-OHL-S-2.0 | Submitted to Tiny Tapeout IHP 26a; chips expected 2027 |
The last closed part
A GPU is difficult to build. The evidence still makes the next work specific: a clear licence, fabrication and bring-up, followed by durable drivers. The owner needs all of them. An open CPU leaves the display and local-model work dependent on the GPU’s maker if its design cannot be inspected or its software maintained independently.
Agenxy proposes clear licences for published designs, grants that fund returned-chip testing as well as design, and drivers brought into the main tree. Closed-GPU makers should publish programming specifications without NDAs; Broadcom, AMD and NVIDIA have published information for some cores, though no public design for those cores was found.
The owner should be able to inspect and repair the part that makes the computer visible. Small open rasterisers have reached a fab and one has been tested. The full GPU needs the rest of that path, not a different name for a milestone.
Sources
The VideoCore IV guide is from 2013. tiny-gpu has 11 instructions, 8-bit data and 256 memory rows; SkyWater 130 nm layout files appeared in April 2024. The licence issue is from that month; the adapter/layout from July 2026 and owner consent from September 2026. Its Tiny Tapeout 7 adapter remained on the to-do list.
Vortex is from Georgia Tech; release 3.0 in June 2026 developed synthesis flows. Its2021 paper explicitly places the ASIC port outside its scope. Ventus is from Tsinghua DSP Lab; its synthesis slides are from 2024. Libre-SOC submitted the 180 nm, 130,000-gate fixed-point v3.0B test chip in July 2021; the packaging report is January 2022; the Simple-V claim March 2024.
Borg’s IHP 26a shuttle closed23 March 2026. Tiny Tapeout 7 used SkyWater 130 nm and returned in November 2024; the tested design reports about 16,000 gates/cells.
Sources were consulted on 16 and 21 October 2026.