Intel’s new Jay shader compiler has cleared Vulkan CTS on Xe2 and Xe3 hardware, and is awkwardly close to being useful.
According to Phoronix Chipzilla has taken a step forward with its open sauce graphics work. The Jay shader compiler, built inside the Mesa project, now passes the Vulkan Conformance Test Suite on Xe2 and the latest Xe3 hardware.
That is not bad going for a compiler that only appeared in public code a few months ago. Intel compiler engineer Alyssa Rosenzweig joined Chipzilla in 2025. She arrived after work on Valve’s graphics stack and after leading the Asahi Linux driver for the Fruity Cargo Cult Apple Silicon.
Rosenzweig was handed the job of building a new compiler to replace the ageing backend used by the Iris OpenGL and ANV Vulkan drivers. The result was Jay, a modern static single assignment design. Similar ideas can be found in ACO for AMD, NAK for Nvidia and AGX for Job’s Mob.
Jay landed in Mesa in April 2026. At first it cleared OpenGL ES 3.0 and OpenCL 3.0 tests, while Vulkan support was still sulking in the corner.
No passes were registered on the full CTS at the time. Four months later, a batch of 63 patches merged into Mesa 26.3 development code this month. Those changes added features, applied vector register-transfer optimisations, and squashed bugs. Phoronix reported that Vulkan CTS conformance now holds on Xe2 and Xe3 platforms.
“Excitingly this MR cleans out the last of the ‘hack, do not upstream’ patches I’ve been carrying, so upstream should be essentially passing Vulkan CTS on Xe2 and Xe3 once this MR goes in,” Rosenzweig said.
That is less than a year after Rosenzweig arrived at Chipzilla, which makes the progress look rather sharp.
Earlier benchmarks hinted that Jay had some proper legs. One demanding conformance test, math_bruteforce sin, showed a fairly brutal gap between the old and new compilers. The older BRW compiler produced 12,980 instructions with 578 spills and 1,144 fills. Runtime hit 19.91 seconds. Jay generated 6,768 instructions, 361 spills and 396 fills. It finished in 7.00 seconds.
Those numbers came from the initial announcement covered by Phoronix in April 2026. Fewer instructions, less register pressure and faster compilation all point to gains once the thing matures.
Still, this does not mean Jay is ready to be shoved into production and forgotten. It remains a work in progress. Jay supports Intel graphics from Skylake onward, covering plenty of integrated and discrete parts. The code still carries experimental labels.
Developers are still refining register allocation for Intel’s oddball regioning rules. Spilling follows established SSA techniques, but it has to bend around the hardware’s quirks.
Rosenzweig has promised more detail at an upcoming XDC conference. The wider context matters. Intel’s Linux graphics stack has spent years iterating on the i915 and Xe kernel drivers, while user-space bits leaned on a compiler tuned for older architectures.
Jay is a clean-sheet effort written in about 14,000 lines of C. Its SSA form simplifies analysis without encumbering everything with old baggage.
The Colombet register allocator, drawn from recent NIR backends, handles the complexity without heroic bodges. The single test case does not represent every workload, but it shows real headroom.
Recent chatter on X has been cautiously upbeat. One post noted that Jay moved from zero Vulkan CTS passes to full conformance in roughly four months. Hardware watchers tracking Lunar Lake and Panther Lake systems have been keeping a close eye on the Mesa updates. No new articles from the past 48 hours added fresh benchmarks.
Searches turned up references to ongoing oneAPI and Level Zero compute work, but those follow different compiler paths. Jay remains a graphics shader story inside Mesa. Mesa 26.3 is scheduled for November. By then Jay could handle more tests and perhaps appear in more real applications.
The longer target is making it the default for Intel’s open drivers. That would close the door on the legacy BRW backend and give OpenGL, Vulkan and compute interfaces a modern shared compiler. The move matters beyond Chipzilla. Other vendors have shown what focused compiler teams can do in the open.
Valve’s ACO, Google’s work on Turnip and Freedreno, and the Asahi effort all raised expectations. Intel now has its own from-scratch design. The pace of patches, the removal of hack flags and the CTS numbers suggest the team has found a productive groove.
Production deployment will need more than conformance. Stability across thousands of applications, laptop power use and competitive game frame rates are still waiting. For now, Xe2 and Xe3 hardware get the first benefit. Future architectures will inherit the foundation.
Intel’s choice to upstream early and iterate in public mirrors successful patterns elsewhere in Mesa. It invites review, catches bugs faster and keeps the work aligned with the wider graphics crowd. Rosenzweig’s reverse engineering and driver background bring a useful edge. Her earlier work on Apple GPUs showed what a small, team could do.
That experience seems to translate neatly to Intel’s more conventional but still deeply peculiar architecture.
