I Came, I Prompted, I Left Part 2: Building a GPU Driver From Scratch in One Month — Cody HoCody Ho

BlogProjectsAboutContactResumeGitHub

I Came, I Prompted, I Left Part 2: Building a GPU Driver From Scratch in One MonthPrevious blog post: https://codyho.dev/blog/hypervisor-macbook-neo/What We DidTL;DR: Niklas and I built a fully OpenGL ES 3.0 compliant GPU driver for the
M4 Mac Mini and MacBook Neo in about a month, a process which normally takes
years. Here is Chrome and Firefox running WebGL on the M4 Mac Mini with working
compositing:Most importantly, the driver is fast enough to run Minecraft at 200fps:Building this driver involved reverse engineering the AGX’s (Apple’s name for
the GPU) incredibly complicated firmware ABI and user-space components. This
was all done in a transparent, verifiably clean room manner using well
established techniques. The code is not yet ready for end users, but we are
looking to get it to end users as soon as possible.How We Did ItPreviously, I built a hypervisor to reverse engineer macOS. Now the goal became
to actually do something useful with it, and what better target than writing a
GPU driver. The GPU is effectively a requirement for any modern system,
otherwise everything needs to be CPU rendered which is orders of magnitude
slower and less power efficient. Our goal was to implement conformant OpenGL
(and soon, Vulkan) drivers for the M4 Mac Mini and MacBook Neo.Normally, building a GPU driver is an endeavor that takes years; our goal was
to do it in days. It turns out that days was overly optimistic, but weeks is
still a massive improvement. In those weeks we have:Reverse engineered the M4, A18 Pro, and (mostly) M5 user space using only
live probing, discovering hardware-supported features and instructions not
emitted by Apple’s driverBuilt a fully working user-space driver, including a new custom IR/shader
compiler, command stream builder, and many more componentsReverse engineered, from scratch, the full AGX firmware ABI using traces from
the hypervisor I previously builtImplemented a full Linux kernel driver for said firmware ABIThroughout this process, we have not looked at any Apple binaries, only
hardware traces (from our hypervisor) and shaders we built ourselves. For
user-space graphics RE, we were careful to treat any required Apple blobs as
opaque objects. We had a friend write documentation on these blobs 1 so we
could write a clean room implementation ourselves (which was mostly built by
just blindly trying stuff until it worked). We have published all of our
experiments so that anyone can verify the provenance of our work (see the twin
agx-re repos under Deliverables).This blog post is divided into two parts, user and kernel space. This mirrors
the split in all modern GPU drivers: the kernel is responsible for interfacing
with the firmware, allocating buffers, and managing scheduling, while the
actual contents of those buffers and what is being scheduled are opaque.
User space is responsible for actually understanding how the GPU works and
filling those buffers with stuff.Kernel SpaceOn Apple Silicon, the kernel driver does not interface directly with the
hardware. Instead, it talks to the GPU firmware running a custom RTOS called
RTKit. That means that the first step to a kernel driver is not talking to
hardware, it’s figuring out the firmware ABI.The firmware ABI was by far the most annoying part of this project, because
rather than doing the sane thing of coming up with a reasonable ABI with nice
interfaces, Apple essentially took a regular kernel driver, cut it in half, and
then put half of it in the AGX and called it firmware, with the other half of
the kernel driver communicating using shared structs in memory. Many of these
structs have firmware owned fields (which we must never modify and which we
must learn from reverse engineering) interleaved with host controlled fields.
For an idea of how complicated the ABI is, this is what the shared memory
tree looks like on the M1/M2:Asahi Lina famously figured all of this out over grueling 12-hour days to build
the M1/M2 kernel driver, an amazing technical accomplishment. Unfortunately,
the A18 Pro firmware ABI (I started my RE work on the MacBook Neo and later
pivoted to the M4 Mac Mini) is significantly more complicated than the
already very complicated M1 firmware ABI:What the F@!#, Apple. Note how the A18 has:1.5x as many structstwice as many pointersa significantly more complicated process for submitting workThere are many other issues that add friction to the RE process 2. I did
have some documentation on the firmware ABI, but it was highly incomplete and
honestly was not very useful 3.My approach was simple and based on the approach used to successfully reverse
engineer the M1/M2 machines: watch what macOS did, replay it, then try to do it
ourselves, which is made possible by the hypervisor.When I described this approach to the LLM, it took replay extremely literally:
the first thing it did was wait for the first firmware visible event (these are
called “kicks”), then saved a copy of the entire GPU memory state. After a
reboot, it copied the saved memory state straight back into host memory,
performed the kick, and saw the output pages change. It would then try to
reconstruct these objects in code, following all the pointers and making sense
of the contents. Over successive experiments, Codex would reduce the number of
pages it copied until there was no more replayed state and everything was built
from source. 4 Amazingly, I noticed Codex had good taste regarding when it
should poke the hardware some more and when it should just run the hypervisor
and capture the state itself.There were three major issues, and all were caused by our inability to
get a clean capture of host work:The first issue was render work submitted after the GPU firmware started. We
could prestage work before the firmware started, start the GPU, and that work
would be completed as expected, but once the firmware started any work
submitted would just be ACKed and retired without actually doing anything. Once
the firmware has started, capturing state is much harder because everything
becomes dynamic and the firmware becomes a stateful object with state you can’t
easily replay.I had to step in at this point and examine Codex’s process. It turns out it was
trying to replay a capture very late in the AGX’s lifecycle, where there had
already been many previous events. When I told it to choose a capture far
earlier in the AGX’s lifecycle, the very first capture after firmware start,
Codex was able to almost immediately discover the issue (it was missing
a single byte descriptor). This took a few days.The second, and only major blocking, issue was compute. The AGX, broadly,
supports two kinds of work: compute and render. In the regular GUI path,
compute work is only scheduled after a significant amount of render work was
already executed. Thus, it took a long time to get a clean capture of a compute
workload, and when Codex finally did it was 336 MB and impossible to replay (it
tried, for a long time). It also tried to construct the objects itself by
looking at the capture, and spent over a week doing this, but was ultimately
unsuccessful. There was just too much nonsense to sift through. This was
exacerbated by issues on my side– after getting render working, I expected
that submitting compute work would be simpler (the firmware ABI for compute is
indeed simpler, so I was correct here), but lost my humility and thought it
would be a cakewalk that would only take a few hours. Thus, I didn’t scaffold
out the task properly for the LLM.The fix actually was given to me in another Codex session. In essence:Disable the GUI by booting into single-user mode; this means no render work
would be done.Install a LaunchDaemon to run at the earliest possible point, the moment
Metal (Apple’s proprietary graphics framework) became available.Run a tiny Metal program that we suppliedCapture and replay this tiny, pure compute trace.The trace was captured successfully. Within a few hours, Codex had
deconstructed it, and within a few days, Codex had compute working. As for why
the original compute codebase didn’t work… Codex has no idea. The working one
and the broken one look very similar.In hindsight, this should have been the strategy from the start– smallest
possible capture, run in single-user mode so as not to perturb results. I
learned from my mistakes here for the final issue:Partial renders ended up being one of the hardest things to figure out. They
occur when the Tiled Vertex Buffer (TVB) isn’t large enough to store the
current geometry (ie, there’s just too many triangles to draw). In these cases,
there are two options, and the driver needs to support both: either increase
the size of the TVB, or perform a partial render, ie, render part of
the geometry, then reload the buffer with the rest of the triangles, and finish
the partial render. These partial renders turned out to be very, very finicky,
even more so than the rest of the work because they essentially mean adding
save and resume to the GPU driver.The workflow I discovered earlier came in very handy here. Codex was able to
replay one partial render transaction, and then modified our Metal shader to
perform multiple partial renders (this is pretty easy by just hammering a
single tile with thousands of triangles until a partial render is triggered)
and then learned how to replay these. Once Codex had a successful replay, it
was only a matter of time until it learned how to build it ourselves.Building the Kernel DriverMoving from a Python prototype driver to a fully featured Linux driver took
three days, and one of those days was almost totally wasted because Codex, for
some reason I still do not understand, chose to tackle partial renders first
(by far the hardest task) instead of doing compute first (the easiest task).
Once I told it to do compute first, everything went smoothly.At all high level, the entire process was, basically:Rewrite the existing drm-shim in Rust following the exact same pattern;
this gives us a synchronous Rust driver.Rewrite the frontend to be asynchronous; the actual GPU submission remains
synchronous.Refactor the GPU submissions to be asynchronous and, instead of polling,
listen for firmware events and associate work with a fenceImplement some low-hanging optimizations, such as batched work submission.This is all pretty routine engineering work that LLMs are definitely capable
of.The only notable thing I found is that Codex aggressively used the hypervisor
to debug why its code didn’t work, including capturing the full address space
and comparing it to known good samples. This sort of systematic debugging is
why Codex is by far my favorite coding agent.User SpaceThe A18 Pro user space is very different from the M1/M2; it has new descriptor
formats, a new ISA, and a bunch of other new things. Aside from being a tile
based deferred renderer designed to run Metal, it’s just a different GPU.The good news is user-space RE has a very well defined process. Simply write a
small Metal program, compile it, run it, see what changed, then take it apart
and start fiddling with the bits until we understand what all of them do. If
you’re thinking this sounds like the sort of boring, repetitive, rote work that
LLMs are very good at, you would be correct.The RE work occurred in two phases. For the first phase, I had Claude look at
every possible Metal program it could find and trying to build a disassembler,
assembler, and understand the format of all the other descriptors/command
streams/etc required for the GPU driver. I had Claude enumerate everything,
including stuff Linux can’t use (like tessellation) for completeness. This was
successful, but just because Claude could disassemble and then reassemble
programs doesn’t mean it knew how to build one itself. When trying to close
this gap, ie, understand every instruction enough to actually be able to
compile our own arbitrary programs, Claude did a horrible job and made
basically zero progress.At this point, Niklas finished his drm-shim for the M4 Mac Mini and joined me
for the second phase of user-space RE. We had two different approaches to
actually finishing the user-space driver:My approach was to prioritize hardware RE, and focus on just figuring out how
the hardware and all the instructions worked. Then I would write a spec and let
the LLM implement it, hopefully ending with full OpenGL and Vulkan compliance.
This means that most of my LLM’s time was spent writing experiments on
hardware, not actually implementing Mesa code. The idea was that once I
understood the hardware, everything else followed.Niklas took a different approach, that I’d describe as “Mesa first”.
Essentially, he tried to build out Mesa first and would only do RE in order to
build out some functionality. His time was split between building and testing
Mesa, and performing RE.It turns out that Niklas made significantly faster progress than I did,
because my agent would spend a lot of time on minor, inconsequential tasks in
the name of completeness. By contrast, his agent was grounded by the need to
actually build Mesa, so it used time and resources a lot more effectively. He
ended up moving so much faster than me that ended up just trying to support his
work by investigating any behavior he didn’t yet understand.This was one big limitation of Codex I noticed. The best word I can think of to
describe it is “pedantic”– it is extremely thorough all the time, which can be
a major benefit in some scenarios, but other times it gets stuck in the weeds
on some random tangent to the detriment of the overall goal.During our RE, we found behavior that was supported by the hardware but not
supported by Metal; this was found by directly messing with the bits of the
different instructions and extrapolating what might exist based off what we
know did exist, just like Alyssa Rosenzweig did when REing the M1/M2. This
included:A native single-instruction 64-bit addAnisotropy to 128x (Metal caps at 16x)A new mode of the matrix unit7-bit immediate support for uniform_movMesa DevelopmentThere are a few things that massively work in our favor when building out
user-space graphics. Most notably, the Khronos compatibility test suite (CTS)
is already an exhaustive corpus of tests our driver must pass. In other words,
the hardest and most sensitive part of working with LLMs, giving them good
tests to ground them, is already done for us.Additionally, Mesa already has great abstractions that make our lives
significantly easier. This is what the modern OpenGL stack on Linux looks like:All we have to do is translate between Gallium, Mesa’s internal API, and AGX
hardware semantics. One of the biggest parts of this process is translating
from NIR, Mesa’s internal IR that’s quite similar to LLVM IR, to the AGX’s
proprietary ISA. As a bonus, this compiler can be reused for a future Vulkan
driver.Niklas was able to slowly iterate through OpenGL features, REing the user space
as he went, until he finally achieved full OpenGL ES 3.0 compliance (the
unsupported tests are optional extensions):Throughout the process, we benefited from the existing M1/M2 work: while the
exact hardware semantics differ, the overall shape remains similar and thus
many of the right tradeoffs/decisions were already made for us. Throughout my
time building this driver, it became clear that Alyssa Rosenzweig and the
others who built the M1/M2 driver are utter wizards– hats off to them!DeliverablesMesa: https://github.com/niklassheth/mesaLinux Kernel Driver: https://github.com/GravityLinux/linux/gravity-m4User-Space RE Documentation (horrible pile of LLM slop, but functional): Cody NiklasRemaining WorkVulkan 1.4, OpenGL 4.6, OpenGL ES 3.2, OpenCL 3.1, Direct3D 12 (via Proton),
and ray tracing are all in scope. We want our driver to be as good as the best
graphics drivers in the world.Additionally, Niklas and I want to upstream all of this, but there are some
significant obstacles. We used the unmodified Asahi UAPI, so there are no
policy issues with Mesa upstreaming, but it needs far more testing, human
review, and to be refactored into a reviewable PR. We also expect significant
skepticism given that this is likely the first ever fully LLM-written GPU
driver, and that our code will be held to a higher standard than
human-written code. We are ready for these challenges, but they are primarily
human and nontechnical, which LLMs cannot help with.The Linux kernel driver will be an even bigger problem, since the M1/M2 driver
is not yet upstream, and practically we are not in a position to change this.
We think the best approach here is just to wait for that driver to be
upstreamed, and then upstream our driver after M1/M2 is upstream (after all the
required refactoring + review + decomposition + whatever). This may be a while
unfortunately.When Can I Use It?Patience young grasshopper, we’re looking forward to getting this code into
your hands soon enough, and the wait may be much less than you may expect.
After all, the apple doesn’t fall far from the tree.Join UsIf you’re interested in being a part of this, we invite you to join our
Discord Server. Feel free to come by to
discuss ideas, chat, or just hang out!Addendum: Really Sam?I used Codex with GPT-5.6 Sol (later GPT-6 Astra when it came out) for the
kernel RE task. One of the biggest issues I hit was the overly aggressive
cybersecurity restrictions (I am not enrolled in trusted access).Ninety-nine percent of the time a simple /goal resume or “keep going” was
enough to have the LLM continue (which also shows the restrictions were overly
aggressive), but they still broke my unattended workflow. I coded the fastest,
dumbest possible solution: a daemon that takes a screenshot every minute, diffs
it against the last screenshot, and if identical (because Codex stopped making
progress) types /goal resume. I accidentally left it on in some group chats:That said, GPT-6 Astra and GPT-5.6 Sol are absolutely insane and by far the
best performers at firmware ABI RE, so this hack was more than worth it.Addendum: M4 vs A18 Pro vs M5As far as I can tell, the user-space implementations of the M4 and A18 Pro are
effectively identical, with the only difference I can find being one value is
slightly larger on the M4, consistent with it having more cores. The firmware
ABIs of the two differ significantly: the second RTKit coprocessor on the A18
Pro makes everything more complicated. In this way, the A18 Pro is more akin to
the M5 in terms of firmware than it is to M4. The M5’s user space has some
similarities to M4, with the ISA being mostly a superset, but some parts are
completely different, such as its texture descriptors. The M5’s user space has
been partially reverse engineered; its firmware ABI is fully reverse
engineered; a prototype drm-shim has been built and thoroughly tested; and I
don’t think it would take long to promote the prototype to a full Rust driver.
My primary targets remain the M4 Mac Mini and MacBook Neo.FootnotesMetal Helper ProgramsI’m not sure what’s with Claude’s header about this not being clean room, I
think it got confused about the whole “don’t look at Apple binaries”
instruction in agx-re. The doc itself clearly does not contain any
tainted information. ↩︎A brief list:If we ever crashed the firmware, the only recovery is a full rebootIf anything was wrong with the work handshake, no error would be
reported, the work would be ACKed and then never completed.Some problems resulted in the output changing as expected, but also
arbitrary corruption on pages that weren’t supposed to be touched. ↩︎https://github.com/mischa85/apple-gpu-firmware-abi ↩︎If we found a page containing a shader, we threw this page away and
substituted our own shader as an additional copyright measure. ↩︎© 2026 Cody Ho

The project involved building a fully OpenGL ES 3.0 compliant GPU driver for the M4 Mac Mini and MacBook Neo in approximately one month, a process typically requiring years, which was achieved by reverse engineering the AGX firmware ABI and user-space components. This effort resulted in a working driver fast enough to run applications like Minecraft at 200fps, and it was accomplished through a transparent, verifiable reverse engineering methodology using techniques derived from prior hypervisor work. The development was structured into kernel space and user space components, mirroring the architecture of modern GPU drivers.

The kernel driver’s primary challenge lay in understanding the firmware ABI, which Apple designed in a manner that complicated standard kernel driver development, especially in the A18 Pro firmware where the ABI involved significantly more structures and pointers than the M1/M2 systems. The process necessitated reverse engineering the entire AGX firmware ABI from scratch using traces obtained via the previously developed hypervisor. The kernel driver is responsible for interfacing with the firmware, allocating buffers, and managing scheduling, but it does not interact directly with the hardware, instead communicating with the GPU firmware running a custom Real-Time Operating System called RTKit. The complexity of the ABI stems from Apple partitioning the standard kernel driver, where portions reside in the firmware and others communicate via shared memory structures containing interleaved firmware-owned and host-controlled fields.

In the user space, the focus was on understanding how the GPU operates and filling the memory buffers. This area required reverse engineering command streams and shader formats. The approach involved using LLMs to systematically analyze possible Metal programs to construct a disassembler and assembler to understand the required descriptors and command streams. While this phase successfully cataloged the necessary formats, the agents encountered difficulty in fully compiling arbitrary programs, indicating that understanding the instruction set required deeper, more granular analysis than the LLMs could provide alone.

The process involved several technical obstacles related to state capture and replay. When attempting to replay GPU operations, issues arose concerning work submitted after the firmware initiated, where subsequent work was often simply acknowledged and retired. Furthermore, capturing state for compute workloads proved extremely difficult, as the AGX supports both compute and render tasks, and capturing a clean, replayable state for compute requests was a major blocker. The solution to this involved redesigning the capture strategy: disabling the GUI, booting into single-user mode, and capturing a minimal compute trace. This allowed the LLM to successfully deconstruct and reconstruct the compute workload. Another difficulty arose with partial renders, which demanded the driver support saving and resuming states, necessitating modifications to the workflow to trigger and replay these transactions.

The development of the kernel driver itself involved refactoring existing code, starting with rewriting the drm-shim in Rust and then refactoring the overall structure to separate synchronous driver execution from asynchronous GPU submissions. This required refactoring GPU submissions to listen for firmware events and associate work with fences. The overall success of the kernel driver depended on systematic debugging, where the LLM aggressively used the hypervisor to compare address spaces and capture known good samples, demonstrating a highly effective, albeit pedantic, debugging capability.

Regarding the user-space development, a difference in strategy was observed: one approach prioritized hardware reverse engineering to establish a specification before implementation, while the other focused on building Mesa first and using reverse engineering to support functionality. The former approach was found to be less efficient, as prioritizing an exhaustive search for completeness often led the agent to spend excessive time on less critical tasks. The success in achieving full OpenGL ES 3.0 compliance was facilitated by leveraging existing work from the M1/M2 development, which provided a similar structural foundation. The process relied heavily on translating between Mesa’s internal Intermediate Representation, similar to LLVM IR, and the AGX’s proprietary Instruction Set Architecture, a task that proved feasible through careful system-level translation.