Split View: Rust 1.97이 Volta 이전 GPU를 잘라냈다 — nvptx64 베이스라인 상향의 내막

Rust 1.97 Cut Off Pre-Volta GPUs — Inside the nvptx64 Baseline Bump

• Introduction — The Line That Didn't Make the Release Notes Headline
• What Is the nvptx64 Target
• What Changed — the Exact Numbers
• Why It Was Cut — Three Concrete Defects
• Why CUDA 11, Specifically, as the Cutoff
• Who Breaks, and What to Do About It
• But Does Anyone Actually Use This — the 2026 Rust-on-GPU Landscape
• The Honest Tradeoff — and When Not to Use This
• Closing
• References

Introduction — The Line That Didn't Make the Release Notes Headline

Rust 1.97.0 shipped on July 9, 2026. The release notes carried three headline items — promoting v0 symbol mangling to the stable default, Cargo's build.warnings setting, and surfacing linker output that had been hidden until now. I already covered these three in a post from a few days ago.

But the same release's Platform Support section had one more line that didn't make the headlines.

nvptx64-nvidia-cuda: drop support for old architectures and old ISAs   (PR #152443)

Here's the interesting part. All three headline items were backward-compatible changes. Code still builds the same even with the symbol-mangling change, the warnings setting is opt-in, and the linker output is simply more visible now. This one line, by contrast, is the only change in 1.97 that actually breaks something. It's just that very few people get broken.

And behind this change lies a story that has suddenly stopped being quiet in 2026 — writing GPU kernels in Rust.

What Is the nvptx64 Target

nvptx64-nvidia-cuda is the target rustc uses to emit code for NVIDIA GPUs. What comes out isn't machine code but PTX — NVIDIA's virtual ISA, which the CUDA driver JIT-compiles into the actual GPU's SASS at runtime. rustc generates this PTX through LLVM's NVPTX backend.

According to the rustc platform support docs, this target's characteristics are:

• It's a Tier 2 target, with a single maintainer, @kjetilkjeka.
• It's a no_std target. The default linker is llvm-bitcode-linker, and you build with crate-type cdylib to get PTX.
• Kernel functions are declared with extern "ptx-kernel".
• The build example in the docs uses nightly.

$ RUSTFLAGS='-Ctarget-cpu=sm_89' cargo +nightly rustc \
    --target=nvptx64-nvidia-cuda -Zbuild-std=core --crate-type=cdylib

What Tier 2 guarantees is spelled out precisely in the target tier policy — "Rust's continuous integration checks that tier 2 targets will always build, but they may or may not pass tests." In other words, build is guaranteed, but passing tests is not. That's the difference from Tier 1, which guarantees both "builds + passes tests." Keep this sentence in mind — it's the key to why defects piled up, as explained below.

What Changed — the Exact Numbers

On May 1, 2026, target maintainer Kjetil Kjeka posted an advance-notice post. In the post's own words, the new minimum requirements are:

"The new minimum supported versions will be: PTX ISA 7.0 (requires a CUDA 11 driver or newer) SM 7.0 (GPUs with compute capability below 7.0 are no longer supported)"

Mapped onto the rustc docs' table, the size of the change is clear at a glance.

SM 2.0 to 7.0. This isn't a gradual adjustment — it's an entire generation swept away in one step. In the advance-notice post's words, "Until now, Rust has supported emitting PTX for a wide range of GPU architectures and PTX ISA versions," and that "wide range" has now narrowed to Volta and later.

Mapping compute capability numbers to architecture names: 5.x is Maxwell, 6.x is Pascal, 7.0 is Volta. So what got cut is all of Maxwell and Pascal — per NVIDIA's legacy GPU list, things like the GTX 1080 (6.1), Tesla P100 (6.0), and Jetson TX2 (6.2). Tesla V100 and TITAN V, at 7.0, are the first generation to survive.

For reference, NVIDIA's current GPU list already starts at 7.5, with everything below moved to the legacy page. The 7.0 line Rust drew is actually one notch below the line NVIDIA itself calls "current."

Why It Was Cut — Three Concrete Defects

This is where the real substance of the change lives. The advance-notice post summarizes the reasoning like this:

"In practice, several defects existed that could cause valid Rust code to trigger compiler crashes or miscompilations."

This means valid Rust code was triggering compiler crashes or incorrect compilation. A miscompilation quietly emits wrong code, which is worse than a crash.

The advance-notice post doesn't spell out what the "several defects" were, but compiler team MCP #965, which pushed the change through, states them concretely. It's a proposal opened on February 1, 2026 and marked major-change-accepted, citing three issues as grounds.

1. rust-lang/rust#147672 — below PTX ISA 7.0, there's an LLVM-side constraint on debug symbol generation.
2. rust-lang/rust#150515 — atomic ordering isn't sufficiently supported on architectures older than SM 7.0.
3. rust-lang/rust#141468 — the PTX ISA versioning scheme doesn't mesh with Rust's target-feature mechanism.

Starting from the third one reveals the nature of the problem. Rust's -C target-feature is a set model — "this feature is present or absent." But PTX ISA versions aren't a set, they're a linear version number. The two models are mismatched from the start, and the wider the supported range, the bigger the combinatorial explosion this mismatch produces.

But the core issue is the second one. Carrying the MCP's rationale over directly: below SM 7.0, atomic ordering isn't sufficiently supported — SM 7.0 and above get official guarantees, but anything before that is best-effort. Why is this fatal for Rust? Because in Rust's memory model, the acquire/release semantics of core::sync::atomic aren't optional. What C++ could shrug off as "just don't use atomics on that GPU" becomes, in a language that promises ordering guarantees at the language level, a case of the language breaking its own promise. That's exactly what "produces miscompilation" means. (For reference, Volta is also the generation that introduced independent thread scheduling within a warp — it's not a coincidence that SM 7.0 became the cutoff.)

In other words, this isn't "cleaning up old hardware" — it's removing a promise that couldn't be kept. The advance-notice post reads the same way — "Raising the baseline addresses these issues and enables more complete support for the remaining supported hardware." They chose narrow-but-intact support over wide-but-broken support.

And this connects precisely to the Tier 2 definition we saw earlier. On a target where CI enforces the build but not passing tests, defects in old architectures that nobody was actually running hardware against had quietly piled up for years.

Why CUDA 11, Specifically, as the Cutoff

Here's the sentence the MCP cites as grounds for the version choice — "CUDA versions older than 12 are end-of-life, and PTX ISA 7.0 is supported starting with CUDA 11.0."

Read closely, the two clauses point at different lines. Following the EOL logic alone, the cutoff should be CUDA 12, but what was actually set is CUDA 11 (PTX ISA 7.0). In other words, they set it one notch more conservatively than the EOL logic would allow. It looks like a choice that raised the bar just enough to eliminate the defects, without cutting more than necessary.

The MCP also cited NVIDIA's own current state as grounds — CUDA 13 supports SM 7.5–12.1, and CUDA 12.9 supports SM 5.0–12.1. This is an interesting detail. The SM 7.0 cutoff Rust settled on is actually more generous than the line (7.5) the latest CUDA 13 has already cut at.

Who Breaks, and What to Do About It

The advance-notice post guides it like this:

"If you do not specify -C target-cpu, the new default will be sm_70, and your build should continue to work (but will no longer be compatible with pre-Volta GPUs)."

"If you currently specify an older -C target-cpu (for example, sm_60), you will need to either: remove that flag and let it default to sm_70, or update it to sm_70 or a newer architecture."

Summarized:

Don't pass -C target-cpu        -> default becomes sm_70. Build passes.
                                    but the output won't run on pre-Volta GPUs.
Pass -C target-cpu=sm_60        -> must drop the flag or raise it to sm_70+.
CUDA 10 or older driver         -> not supported.
Maxwell / Pascal GPU            -> not supported. No opt-out.

Worth noting: there is no opt-out. The advance-notice post has no "flag to restore the old behavior" either. If you need to emit PTX for Pascal, the only way is to pin a toolchain at 1.96 or older — which also means living with the defects.

The advance-notice post's assessment of the blast radius is matter-of-fact.

"In this case, the most recent affected GPU architectures date back to 2017 and are no longer actively supported by NVIDIA."

"We therefore expect the overall impact of this change to be limited."

Keep in mind this is the author's own estimate, but the reasoning holds up — the most recent generation among the affected GPUs dates to 2017, and NVIDIA itself no longer actively supports it.

But Does Anyone Actually Use This — the 2026 Rust-on-GPU Landscape

Time to be honest. Tier 2, one maintainer, nightly required, no_std. Not many people use this target raw. So why does this change matter?

Because in 2026, the scene around handling GPUs from Rust suddenly got crowded. Looking at the ecosystem map the Rust CUDA project put together, the approaches split like this.

• cuda-oxide — an rustc codegen backend targeting NVIDIA PTX/SASS. Its design center is "bringing CUDA to Rust" (writing kernels, device intrinsics, the SIMT execution model).
• rust-cuda — goes rustc → NVVM IR → PTX via rustc_codegen_nvvm. Its design center is "bringing Rust to the GPU" — Rust conveniences like async/.await on the device.
• Rust-GPU — graphics-oriented, targeting SPIR-V. Vulkan/Metal/DirectX. Centered on cross-vendor portability.
• CubeCL — a built-in DSL plus a JIT runtime. CUDA/ROCm/WGPU from one kernel. It's a proc-macro approach, not a rustc backend.
• std::offload — a nightly language feature. Implicitly offloads CPU loops to an accelerator via the LLVM offload runtime.
• cudarc — safe CUDA driver bindings. The kernel is written elsewhere; only the host side is Rust.
• wgpu — a Rust implementation of the WebGPU API. A different layer of the stack entirely.

The biggest news among these is cuda-oxide. It's a Rust→CUDA compiler put out directly by NVIDIA Labs, and in the repo's own description it's "an experimental Rust-to-CUDA compiler that lets you write (SIMT) GPU kernels in safe(ish), idiomatic Rust." It's neither a DSL nor bindings — it compiles standard Rust straight to PTX. The pipeline is Rust → MIR → Pliron IR → LLVM IR → PTX, and it's licensed Apache-2.0.

A GPU company shipping its own Rust compiler for its own platform — that's the signal of 2026. And for these tools to actually work, the PTX generation underneath them has to be correct. The baseline bump is exactly the work of firming up that floor.

Evidence that the floor was shaky is in the Rust CUDA project's own documentation. It explains its own reason for existing this way — that until now "the only viable option was using the LLVM PTX backend," and that it "generates invalid PTX for many common Rust operations." That's a strong enough distrust of the default path that they had to dig a separate one. 1.97's cleanup narrows that default path in exchange for making it trustworthy.

The Honest Tradeoff — and When Not to Use This

The change itself is a good trade. For old architectures nobody actually verifies against real hardware anymore, it's better to narrow the supported range and get it right on the remaining hardware than to promise atomic guarantees you can't keep and quietly emit wrong code. It chose correctness over compatibility, and since correctness was actually broken, that choice is justified.

That doesn't mean you should go write GPU kernels in Rust right now, though. Written honestly:

• The nvptx64 target is Tier 2 — only the build is guaranteed, not passing tests. Even the build example in the docs uses nightly.
• It's no_std, and has target-specific constraints of its own. For example, the compiler rejects a static initializer that forms a cycle (static A: Foo = Foo(&A); is an error).
• Even NVIDIA's own cuda-oxide describes itself this way — "in an early stage (alpha) and under active development: you should expect bugs, incomplete features, and API breakage." On top of that, it's pinned to a specific nightly (nightly-2026-04-03) and requires CUDA 12.x or newer and Linux.

So for most cases, the answer is this. If you just want to use a GPU from Rust, write the kernel in CUDA C++ and attach only the host side with a safe driver binding like cudarc. This combination runs in production today. Writing the kernel itself in Rust is still an option reserved for research, experimentation, or cases where a single-source codebase gets a large benefit from sharing types between host and device.

If Maxwell or Pascal is still in active service — the call is one of two things. Pin the toolchain at 1.96 or older (living with the defects), or don't write kernels in Rust. In practice, the latter was already the answer in most cases.

One thing worth adding: the fact that cuda-oxide requires CUDA 12.x or newer to begin with says a lot about the nature of this change. The ecosystem was already ahead of CUDA 11, and the compiler's baseline is closer to catching up late to where it already was.

Closing

Rust 1.97's nvptx64 baseline bump is a one-line change in the release notes. What it actually does: going from SM 2.0/PTX 3.2 to SM 7.0/PTX 7.0, removing Maxwell and Pascal, and in exchange making atomic ordering and debug symbols work correctly on the remaining hardware. There's no opt-out, and by the author's own judgment, the impact is limited.

The reason this quiet cleanup is worth reading at length is the backstory. Defects had piled up for years on a Tier 2 target nobody was watching, and in 2026, real tools — including a compiler NVIDIA itself shipped — started building on top of that target, giving people a reason to look at the floor. Choosing to narrow the supported range and make it correct, rather than widen it, looks like the decision of people who actually intend to build something on top of it.

Still, the conclusion is matter-of-fact. Rust-on-GPU has moved from "impossible" to "starting to be managed," but it is not "production-ready." Take the documentation that says alpha at face value. And for most people, the only thing to do with 1.97 is still rustup update stable.

References

• Raising the baseline for the nvptx64-nvidia-cuda target (Kjetil Kjeka, 2026-05-01)
• Announcing Rust 1.97.0 (2026-07-09)
• Compiler team MCP #965 — proposal to drop support below SM 7.0 / PTX ISA 7.0
• rustc platform support docs — nvptx64-nvidia-cuda
• rustc target tier policy
• NVlabs/cuda-oxide — NVIDIA Labs' experimental Rust→CUDA compiler
• The Rust + GPU Ecosystem — comparing projects by approach
• The Rust CUDA Guide — rustc_codegen_nvvm
• CUDA GPU Compute Capability — NVIDIA (current) / legacy list
• Volta Compatibility Guide — NVIDIA (compute capability 7.0, independent thread scheduling)
• Rust 1.97.0 Dropped — and I Was Already Building With That Exact cargo (related post)