Latest

Technology

RISC-V gains 11% performance in Linux 7.4

Linux 7.4 moves a RISC-V optimization that gains up to 11% bandwidth to runtime detection, and it already works in distro kernels.

RISC-V gains 11% performance in Linux 7.4
RISC-V gains 11% performance in Linux 7.4

The architecture changes for Linux 7.4 were merged on October 30, and among them is an optimization that is nothing spectacular and very practical. The kernel stops software-aligning memory copies on RISC-V processors that support fast unaligned accesses, and it decides this at runtime instead of requiring a specific build. In the tests that accompanied the original patch, the improvement reached up to 11% bandwidth.

The figure is not new. Tenstorrent had already achieved it in Linux 7.3 with an equivalent optimization, but there the support had to be compiled into the kernel. That made it useless for any distribution that wanted a single binary capable of booting on boards from different manufacturers, which is exactly what the normal RISC-V user needs in 2026.

The step 7.4 takes is to move that decision to boot time. The kernel checks what the hardware in front of it can do and chooses the fast path only when the board supports it, so that a single image works for a range of machines. It is a discreet engineering change that decides whether RISC-V stays in the labs or enters the machines people buy.

A simple change boosts performance

The 7.3 code relied on a configuration option that declared, at compile time, that the processor had efficient unaligned accesses. With that declaration, the kernel skipped the software realignment of user copies, which is pure work the processor can save, and the result was that bandwidth gain of up to 11% measured by Tenstorrent.

The problem appeared when distributing the binary. A kernel compiled that way could fail or perform poorly on boards without that capability, and distributions cannot afford one image per model. The patch series sent by Andy Chiu from Tenstorrent in September 2026 solved the dilemma by moving the check to boot time, with the same logic the kernel already uses for other optional capabilities.

The consequence is that the optimization stops being a promise and becomes something the user receives when updating. There is nothing to recompile, no image variant to choose and no need to know which silicon revision the board carries. In an architecture as fragmented as RISC-V, that normality is worth more than the percentage.

Why does the distro kernel matter?

The RISC-V ecosystem lives a paradox. Boards and processors multiply, prices fall and there are distributions that publish stable images, but optimized software still arrives in the form of patches that must be compiled by hand. Every feature that depends on a compile option adds a variant and pushes the user experience further from what x86 or Arm offers.

The generic kernel is the piece that breaks that dynamic. If a distribution can compile a single image and serve it to its entire fleet, the cost of maintaining RISC-V plummets and manufacturers' interest in sending patches rises. That virtuous circle is what has sustained Arm for a decade and what RISC-V needs to stop being a curiosity at trade fairs.

Runtime detection also fits the way the kernel already handles other capabilities. It is not an exception, but rather applying the same pattern used for cryptography instructions or vector extensions, and that consistency is what makes maintainers accept the change without debating it for months.

Tenstorrent's 11% and what it measures

The figure should be put in context so as not to oversell. The 11% corresponds to bandwidth improvements in memory copy tests carried out by Tenstorrent, not to an increase in the performance of any program. In a real web server workload the effect is diluted; in a task that moves large blocks of data, it shows.

That profile fits the use these machines are being given. The RISC-V equipment reaching the market is sold for edge computing, storage and network tasks, where memory copies are a central part of the work. On a board with Linux and a network workload, removing the software realignment frees cycles that were previously spent moving bytes from one place to another.

Change Version Who benefits
Software realignment jump 7.3, custom-compiled Those who compile their own kernel
Runtime detection 7.4, active by default Any user of a distro image
Measured gain Up to 11% bandwidth Workloads with large memory copies

The other reading of the data is that RISC-V is already competing in the boring territory. For years news about the architecture was about experimental support and boards for enthusiasts. Now it is about bandwidth in memory copies and distribution kernels, which is where it is decided whether a platform is good for work or just for showing off.

RISC-V stops being an experiment

The context of recent weeks reinforces that idea. NVIDIA has moved CUDA toward RISC-V and has changed the map of open hardware, and projects like DC-ROMA bring Ubuntu with local AI to boards that fit in your hand. When the world's largest proprietary software ecosystem adapts its platform, the question about the architecture stops being whether it will arrive and becomes when it will be boring.

That "boring" is the goal, and it is not an insult. A boring platform is one that boots without tricks, updates without recompiling and runs the software people use without asking about the architecture. Linux 7.4 provides a small piece of that normality, and the missing ones are of the same type: mature drivers, single images and tools that do not require being an expert.

The user who buys a RISC-V board this year is still a pioneer, and will probably have to wrestle with a driver or a bootloader. The difference is that each merge window takes one fight off the list, and this one removes the need to compile your own kernel to gain an 11% improvement that should already come as standard.

The open architecture is now competing on the boring stuff

The news won't move prices or change a purchasing decision this week. What it does do is confirm that the groundwork on RISC-V is heading in the right direction, with patches that solve distribution problems rather than lab demonstrations. That is the kind of progress that generates no headlines and decides which architecture is in use five years from now.

If you have a RISC-V board with kernel support, the update to 7.4 will give you the improvement without touching anything. If you are waiting to buy, the signal is that software is no longer the reason to wait and hardware becomes the only variable. And if you work on a distribution, the message is that the generic kernel increasingly covers more cases without the need for variants.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
From Huang to Botín, from Zara to Byron: corporate responsibility in global capitalism

Keep reading