A quietly significant change is coming to the Linux kernel’s virtualization layer. The upcoming Linux 7.1 release will include patches enabling KVM — the kernel-based virtual machine hypervisor — to properly manage AVX-512 state through a mechanism called Borrowed Mapped Memory, or BMM. For cloud providers and enterprise operators running dense virtual machine workloads on Intel and AMD hardware, this is the kind of under-the-hood improvement that translates directly into measurable gains in throughput and reduced overhead.
The technical details, first reported by Phoronix, center on how KVM handles the extended processor state associated with AVX-512 instructions. AVX-512, Intel’s wide vector instruction set that operates on 512-bit registers, has long been a source of tension in the Linux kernel community. The instructions are powerful — they can dramatically accelerate workloads in scientific computing, machine learning inference, database analytics, and cryptography. But they’re expensive to context-switch. Every time the hypervisor saves and restores the full AVX-512 register state during a VM exit or entry, it burns cycles and memory bandwidth that could be spent doing actual work.
That’s where BMM enters the picture.
Borrowed Mapped Memory is a technique that allows KVM to map guest state directly into the host’s address space in a way that avoids unnecessary copies. Instead of the hypervisor eagerly saving and restoring the full extended state — which for AVX-512 includes the ZMM registers, opmask registers, and associated XSAVE areas totaling potentially several kilobytes per vCPU — BMM lets the system defer or eliminate redundant state transfers. The guest’s AVX-512 state is “borrowed” in place, mapped rather than copied, and only materialized when actually needed.
This is not a new concept in operating systems design. Lazy state switching for floating-point and SIMD registers has existed in various forms for decades. But applying it cleanly to KVM’s AVX-512 handling, with the complexity of nested virtualization, live migration, and the XSAVE/XRSTOR instruction family, is a different matter entirely. The patches heading into Linux 7.1 represent months of careful engineering to get this right without introducing subtle correctness bugs — the kind that show up as silent data corruption in a guest’s vector registers during a poorly timed preemption.
The performance implications are substantial for anyone running virtualized workloads that touch AVX-512. Cloud providers like AWS, Google Cloud, and Azure all offer instance types backed by processors with AVX-512 support — Intel’s Ice Lake, Sapphire Rapids, and the recently shipping Granite Rapids chips all include the instruction set. AMD’s Zen 4 and Zen 5 architectures support AVX-512 as well, with AMD notably implementing the 512-bit operations using doubled-up 256-bit execution units rather than dedicated 512-bit hardware. In both cases, the KVM overhead of managing AVX-512 state has been a real cost for hypervisor operators.
How real? Consider the numbers. A full XSAVE of AVX-512 state can consume upward of 2.5 KB per vCPU. On a server running 200 or 300 virtual CPUs — not unusual in a modern cloud node — that’s hundreds of kilobytes being shuffled in and out of memory on every VM exit. VM exits happen frequently. Thousands per second per vCPU in some workloads. Multiply it out and you’re looking at gigabytes per second of memory bandwidth consumed purely by state management overhead. BMM’s approach of mapping rather than copying this state can cut that overhead dramatically.
The Linux 7.1 merge window is expected to open in the coming weeks, following the release of Linux 7.0. The KVM AVX-512 BMM patches are part of a broader set of KVM improvements queued for this cycle. As Phoronix noted, the patches have been through multiple rounds of review on the kernel mailing list, a process that reflects both the sensitivity of the code — touching the hypervisor’s register state management is inherently risky — and the high bar the KVM maintainers set for changes that affect correctness.
Sean Christopherson, one of the primary KVM maintainers at Google, has been central to the review process. His involvement signals the seriousness with which the KVM team treats modifications to state save/restore paths. Any bug here doesn’t just crash a process; it can corrupt a virtual machine’s computation silently, potentially across a live migration boundary where the state is serialized and deserialized on a completely different physical host.
The timing of this improvement aligns with a broader industry trend. AVX-512 adoption has been accelerating after a period of uncertainty. Intel famously disabled AVX-512 on its Alder Lake consumer processors, leading some to wonder whether the company was retreating from the instruction set. It wasn’t. Server-side, Intel has doubled down. Sapphire Rapids and Granite Rapids both feature full AVX-512 support with AMX (Advanced Matrix Extensions) layered on top. And AMD’s embrace of AVX-512 starting with Zen 4 removed any lingering doubt about the instruction set’s longevity in data center workloads.
For database vendors, this matters. Systems like ClickHouse, DuckDB, and even PostgreSQL with vectorized execution extensions can exploit AVX-512 for faster analytical queries. When those databases run inside virtual machines — as they increasingly do in managed cloud database services — the hypervisor’s ability to handle AVX-512 state efficiently becomes a direct factor in query latency and throughput.
The same is true for AI inference. Running quantized neural network models with AVX-512 VNNI (Vector Neural Network Instructions) is a cost-effective alternative to GPU inference for many production workloads. Companies deploying these models inside VMs or containers running on KVM-based hosts will see the benefits of reduced state management overhead immediately, even if they never know the specific kernel patch responsible.
There’s a subtlety here that deserves attention. BMM doesn’t just help with raw performance. It also helps with power consumption. Modern Intel processors implement a behavior where the use of AVX-512 instructions can cause the core to reduce its clock frequency — the so-called “AVX-512 frequency throttling” or license-based downclocking. While this throttling is triggered by the guest’s actual use of AVX-512 instructions and not by the hypervisor’s state management, reducing the time spent in state save/restore means the core spends less total time in the higher-power AVX-512 license state. The power savings are modest per core but meaningful across a fleet of thousands of servers.
Not everyone is celebrating AVX-512’s continued prominence, of course. Linus Torvalds himself has expressed skepticism about the instruction set in the past, famously hoping it would “die a painful death” due to concerns about its impact on CPU frequency scaling and die area. But the market has spoken. Workloads that benefit from wide SIMD are too valuable, and both Intel and AMD have committed to supporting 512-bit vectors going forward. The kernel, pragmatically, follows the hardware.
And so the KVM subsystem adapts. The BMM patches for AVX-512 in Linux 7.1 represent exactly the kind of incremental but impactful work that keeps Linux dominant in the server and cloud space. No flashy announcements. No marketing slides. Just careful engineering that shaves microseconds off VM transitions and megabytes off memory bandwidth consumption, compounding across millions of servers worldwide.
For KVM users evaluating the upgrade path to Linux 7.1, the AVX-512 BMM support is worth testing early, particularly in environments with heavy AVX-512 utilization. The patches should be available in release candidate kernels once the merge window closes, and early adopters in the cloud provider space are likely already running them in internal testing.
The broader lesson here isn’t about any single patch. It’s about the relentless optimization work happening in the Linux kernel’s virtualization stack — work that rarely makes headlines but quietly determines the economics of cloud computing. Every microsecond saved on a VM exit, every unnecessary memory copy eliminated, every watt of power conserved across a hyperscale fleet — these are the margins that separate profitable cloud operations from unprofitable ones. Linux 7.1’s KVM improvements for AVX-512 are a small piece of that puzzle. But they fit precisely where they need to.
WebProNews is an iEntry Publication