NVIDIA just moved another piece into place for its expanding CPU ambitions. Hours ago the company upstreamed basic support for its upcoming Rigel core into the LLVM and Clang codebase. The change targets the processor that will power the Rosa CPU, the successor to the recently detailed Vera platform. And it arrives remarkably early. The patch landed in time for next week’s branching of LLVM 23.1.
Developers can now compile with -mcpu=rigel. The addition registers the new Arm v9.2 core, enables its feature set for code generation, yet stops short of specialized tuning tables or cost models. That work will come later. Still, the move signals NVIDIA’s seriousness about open-source compiler support well before silicon ships. Early upstreaming like this smooths the path for software readiness across the data-center stack.
The Rigel core builds directly on lessons from Olympus, the heart of Vera. NVIDIA’s own blog lays out the priorities with blunt clarity. “Rigel is NVIDIA’s next-generation Arm v9.2 CPU core, delivering higher per-core performance than Olympus while keeping the same silicon footprint,” the post states (NVIDIA Blog). Key changes include improved instruction delivery, a larger L2 cache and more efficient memory handling. These tweaks target one metric above all others: single-threaded speed at scale.
Why obsess over single-thread performance when GPUs dominate AI training? The answer sits in the emerging world of agentic AI. Autonomous agents don’t just run massive matrix multiplies. They loop through reasoning steps, tool calls, verification and correction. Each iteration depends on fast CPU execution of the next command. “Agents count in nanoseconds,” the NVIDIA post explains. Data-center CPUs have long traded per-core speed for higher core counts and lower cost. That bargain no longer holds when latency in the control loop determines overall system throughput.
Vera already shows the payoff. Equipped with Olympus cores, it delivers 50 percent higher instructions per cycle than the earlier Grace CPU. Memory bandwidth hits 1.2 TB/s via LPDDR5X while drawing under 40 watts for the memory subsystem. Core-to-core bandwidth reaches 3.4 TB/s. Real customer results from Perplexity demonstrate the difference. Coding workflows run 1.5 times faster. Sandbox environment starts finish 1.9 times quicker. Large-scale SQL analytics complete three times faster. Streaming latency drops by as much as six times.
Those gains come from sustained per-core performance that reaches 1.8 times what comparable x86 processors achieve on the same agentic workloads. The numbers matter because they come from production deployments, not synthetic benchmarks. But NVIDIA clearly views Vera as only the first step. Rosa with Rigel will push the same philosophy further. Same die area. Higher performance per core. Continued focus on the control-plane demands of AI agents.
Compiler support forms a critical part of that plan. NVIDIA engineer Dhruv Chawla drove the upstreaming effort. The LLVM commit adds the basic target definition and detection logic (LLVM Commit). It mirrors a similar patch merged into GCC just days earlier after NVIDIA first publicly confirmed Rigel’s existence. Phoronix first reported the GCC work and followed hours ago with coverage of the Clang addition (Phoronix).
Both compiler projects now carry the -mcpu=rigel option. Neither yet includes the deep scheduling models or cache-cost heuristics that deliver peak performance. Those refinements typically arrive closer to product launch. Yet the early presence of the target name allows distribution builders, benchmarkers and early adopters to begin testing code paths. It also lets the broader open-source community start reviewing and extending the support.
Look at the timeline. Vera with Olympus cores has only recently begun appearing in customer deployments. NVIDIA confirmed Rosa and Rigel in the same blog post that touted Vera’s production wins. The CPU will launch alongside the Feynman GPU architecture, according to multiple reports placing it in the 2028-2029 timeframe. That gives the software stack years to mature. NVIDIA clearly intends to avoid the last-minute scramble that has sometimes plagued new CPU launches.
The company already ships optimized Clang builds for its Grace CPU. The Rigel work extends that investment. It also reflects a broader shift. Once content to let others maintain Arm server software, NVIDIA now drives patches across GCC, LLVM, the Linux kernel and more. Patches for Rosa and Rigel will continue flowing to the kernel and other components in coming months, according to Phoronix reporting.
Competitive pressure helps explain the urgency. AMD prepares its EPYC Venice processors while Intel readies Xeon cores codenamed Diamond Rapids. Both target the same AI-infrastructure market. Yet NVIDIA’s pitch rests on tight integration between CPU, GPU and networking. A fast control-plane CPU that keeps agents moving without stalls becomes a differentiator when the rest of the rack already carries the NVIDIA logo.
Single-threaded performance at scale turns into a feature when AI systems grow more autonomous. The industry once measured success in FLOPS. Now it also measures how quickly an agent can evaluate its last action and decide on the next. Rosa’s Rigel core aims to win that race while fitting inside the same power and area budget as its predecessor.
The LLVM change itself stays modest. A new entry in the target parser. Some feature bits enabled. Basic scheduling inherited from the generic Armv9 model. Future commits will add the microarchitectural details that let Clang generate truly optimal code. But the foundation now exists in the mainline project. Developers compiling for Arm servers can already experiment with the flag and watch how generated code evolves.
NVIDIA’s pace impresses. Confirmation of Rigel. GCC support. LLVM support. All within days. The pattern suggests a disciplined upstreaming strategy designed to have production-quality compiler support ready long before first customer shipments. That matters for a company whose data-center revenue now dwarfs its gaming business and whose customers demand immediate performance out of the box.
So the Rigel core moves from roadmap slide to compiler target. The Rosa CPU inches closer to reality. And the agentic AI systems that will run on them gain another building block in their software foundation. The real test will come when the silicon arrives and the full tuning lands. Until then, the early patches keep the pipeline full and the community engaged.
WebProNews is an iEntry Publication