For two decades, the question of which processor could power an NVIDIA GPU had two possible answers. The third has just entered the list of requirements, and not as a laboratory experiment. Frans Sijstermans, NVIDIA's vice president of Multimedia Architecture, announced in July 2025, during the fifth RISC-V summit in Shanghai, that CUDA will come to RISC-V.
Fifteen months later, the company published the fine print. In the tutorial it gave at Hot Chips 2026, held from August 23 to 25 in Palo Alto, its engineers listed what a RISC-V processor must meet to host CUDA. The reason the announcement is credible lies not in the new acronyms but in an earlier one: RVA23, the profile ratified in October 2024 that sets a stable hardware target. Sijstermans put it bluntly: "We would not have considered this without RVA23."
The practical difference is measured in compatibility. Today CUDA host support covers 64-bit x86 processors and Arm ones; with RISC-V there would be three architectures and, for the first time, one with an open license. If that door truly opens, the x86 monopoly as host cracks and Linux gains a new ecosystem to serve. What is missing is just as concrete: NVIDIA has given no timeline, there are no performance figures and the silicon that already runs CUDA is sold in limited quantities.
From an announcement in Shanghai to a list in Palo Alto
RISC-V International described that announcement as a strategic technology statement rather than a product launch. The distinction matters: that day there was no board, no signed driver, no support date, only the intention to port the stack.
At Hot Chips 2026, NVIDIA presented for the first time a concrete set of requirements for the host and named SiFive as a partner to have a working reference. The demonstration came on August 24, 2026, at an event parallel to the conference. SiFive presented BigSky SF-2U870, a 2U server with 32 P870-D cores at 2.0 GHz that already runs CUDA with an NVIDIA GPU as the primary node. The company says the machine is available from that day, albeit in limited quantities and with demand ahead of supply.
This is where the story stops being a slide. A RISC-V processor booting an operating system and talking to a graphics card does not mean it runs CUDA, because CUDA does not live in the GPU but in a stack of drivers and compilers that needs a predictable memory model.
RVA23 is the key to everything
A profile is the least glamorous part of an architecture and yet the one that decides whether a program compiled in one place boots in another. RVA23 sets the mandatory extensions for 64-bit application processors, so that a binary compiled for the profile works on any core that meets it. The consortium ratified it on October 21, 2024, two years ago, with a v1.0 document dated the 17th of that same month.
Without that common floor, each manufacturer could implement a different subset and force recompilation per chip. For a stack like CUDA, which pushes vector extensions and needs the operating system to discover the machine in a standard way, that amounts to a per-customer adaptation. Sijstermans said RVA23 gave them not only a stable target but the peace of mind they had long been waiting for.
The profile also moves other players. Canonical has moved all its Ubuntu development for RISC-V to RVA23, according to the consortium's own text, a decision that reduces the configurations a distribution must keep standing.
What the specification really demands
NVIDIA did not stop at the profile. Its list starts at RVA23 and continues with the server specifications that RISC-V International publishes separately, because the goal is not a single-board computer but a data center machine. Each row of the table is a condition the silicon must meet before the stack agrees to boot.
| Requirement |
What it demands |
Why NVIDIA asks for it |
| RVA23 profile |
Mandatory ISA extensions |
One binary works on any core that meets it |
| Server SoC and Server Platform |
RAS, dedicated security processor, defined boot |
Server behavior without per-manufacturer inventions |
| ACPI and the BRS specification |
Hardware discovery and power management |
The software team got stuck at first because of its absence |
| Vector extensions |
Predication to avoid branches |
Tighter loops and more efficient generated code |
| PCIe coherency |
A guarantee, not a recommendation |
Avoid cache invalidations on every DMA transfer |
| PCIe peer-to-peer |
Direct communication between devices |
No detour through host memory |
Coherency is the row that best explains the character of the announcement. The RISC-V server specification recommends implementing cache coherency on the PCIe link, and NVIDIA asks for a guarantee. If the CPU leaves modified data in its cache and CUDA launches a copy by DMA, the transfer engine may read an old version of memory, and without hardware coherency that invalidation would be done by software on every operation.
A contract that also reaches NVLink Fusion
Anyone who wants to plug their CPU into NVIDIA's interconnect fabric inherits the full customs check. NVLink Fusion, the program with which third parties integrate NVLink intellectual property into their own silicon and which the company presented at Computex in May 2025, adds support for DOCA and NCCL to the CUDA requirements. The template it cites is GB10, where a MediaTek CPU die joins the GPU via NVLink C2C.
The partner NVIDIA cited in the talk for that RISC-V variant is SiFive. A customer with its own silicon can place its CPU next to an NVIDIA GPU without going through the company's catalog, as long as it meets the server level.
A new pulse
The presentation had visible gaps. NVIDIA did not go through the full list, which by its own account runs to about two pages, and acknowledged that it targets a performance level rather than a minimum configuration, because designing for the lowest common denominator would force it to ship inefficient code. Nor are there availability dates for CUDA on RISC-V or latency figures.
The underlying effect is a market one. For years, the processor governing a GPU was a decision between two vendors, and that gave Intel, AMD and Arm bargaining power that did not depend on performance alone. With a third option, Arm loses its exclusivity as an alternative and the buyer regains room to design its own machine.
Software also counts in that pulse. NVIDIA has spent months moving work into the Linux kernel on its own turf, with tools such as boro, its Rust CLI for managing the kernel, and the arrival of its chips for Arm machines has pushed distributions to cover that architecture, as was seen with the aarch64 support that reached Omarchy with the N1X. A RISC-V host would be the third front of that strategy.
The new bottleneck is the factory
BigSky, with its 32 cores at 2.0 GHz, is a development machine, not a server bought from a catalog. Nor has NVIDIA said which of the boards sold today could meet the contract, and the technical analysis makes it clear: most existing RISC-V hardware does not, and few consumer machines will in the short term. ACPI is the hardest point on the list, because it requires firmware and validation that the board manufacturer takes on itself.
The comparison with Arm helps calibrate the wait. The British architecture has had ACPI in the standard for years and even so real support has been uneven, so a specification ratified in 2025 does not translate into compliant boards the following year.
Who will build the first server NVIDIA accepts?
The Shanghai announcement has already fulfilled its most important function: turning a hallway idea into a specification that any designer can read and certify against. That NVIDIA publishes it before having general silicon is a way of saying that the bottleneck is not set by it, but by whoever is capable of building a compliant board.
What remains to be seen is whether the contract is an entry threshold or a fence. If NVIDIA maintains the hardware coherency requirement without exceptions, the installed base of RISC-V will remain out of play for years. If it accepts slow paths for boards that do not comply, the door opens to cheaper models and to enthusiasts who have long been asking whether their graphics card can hang off an open processor.