AMD and Intel are working together on new x86 AI Compute Extensions, known as ACE, to improve how future CPUs handle AI workloads. The latest ACE specification focuses on faster matrix multiplication and support for low precision data formats, both of which are important for machine learning, neural networks, and large language models.
The work is part of the wider x86 Ecosystem Advisory Group, which AMD and Intel formed to strengthen the x86 platform and make future CPU features more consistent across architectures. The group previously outlined several key technologies, including FRED, AVX10, ChkTag, and ACE.
ACE is aimed at one of the biggest challenges facing x86 CPUs right now. AI workloads increasingly rely on matrix math, but traditional SIMD extensions can have limits in scalability and compute density. AMD and Intel want ACE to close that gap by adding new hardware and instruction support for matrix operations directly into future x86 designs.
Why ACE matters for future x86 CPUs
Matrix multiplication is a core part of AI computing. It is used heavily in neural networks, inference, training, and many machine learning tasks. GPUs and dedicated accelerators are strong in this area because they are built to process large amounts of matrix math efficiently.
CPUs can already handle some of this work through technologies such as AVX and AVX10, but they are not always as efficient when workloads grow larger. ACE is designed to improve that by giving x86 CPUs new tools for matrix processing.
| Feature | What ACE adds |
|---|---|
| Matrix acceleration | Faster processing for AI related math |
| Tile registers | New register state for dense matrix operations |
| AVX integration | Works alongside existing vector instructions |
| Low precision formats | Supports smaller data types used in AI |
| Scalability | Designed for future x86 architectures |
| Energy efficiency | Aims to improve performance without wasting power |
This does not mean CPUs will replace GPUs in large AI data centers. It means future x86 chips may become much better at handling AI tasks locally and efficiently.
ACE builds on AVX instead of replacing it
One important part of ACE is that it works with AVX rather than replacing it. The specification adds ACE register state, including tile and block scale registers, while also allowing data to move between ACE registers and AVX registers.
That means future CPUs could combine the flexibility of AVX with denser matrix processing through ACE. AVX remains useful for broad vector workloads, while ACE adds more specialized support for AI style computation.
This kind of integration matters because software developers need predictable and scalable CPU features. If AMD and Intel support the same core extensions, developers can build AI and machine learning software with more confidence across x86 systems.
Low precision formats are a major part of the plan
Modern AI workloads often use lower precision data formats to improve speed and reduce memory usage. Instead of relying only on FP32, AI models may use BF16, FP16, FP8, FP6, FP4, INT8, and other compact formats.
ACE includes support for several of these formats and conversion operations through the AVX10 framework. That is important because smaller formats can allow more data to be processed faster while using less bandwidth and power.
| Format type | Why it matters |
|---|---|
| FP32 | Standard high precision floating point |
| BF16 and FP16 | Common AI and training formats |
| FP8 | Useful for efficient AI inference and training |
| FP6 and FP4 | Lower precision formats for future AI efficiency |
| INT8 and INT32 | Widely used for inference and accumulation |
| MX formats | Support for microscaling style AI data handling |
The inclusion of these formats shows that AMD and Intel are not only thinking about traditional CPU workloads. They are preparing x86 for AI models that depend on lower precision math.
AMD and Intel are trying to protect x86’s role in AI
The AI boom has put pressure on every chip architecture. GPUs dominate large scale AI training and inference, while Arm and custom accelerators are growing in importance across data centers, laptops, phones, and edge devices.
For x86, the challenge is clear. It needs to stay relevant in a market where AI performance is becoming a major selling point. ACE is one way AMD and Intel are trying to make sure future x86 CPUs are not left behind.
This effort also gives the two rivals a reason to cooperate. AMD and Intel still compete heavily in processors, but both benefit if developers continue to see x86 as a strong and modern platform for AI software.
ACE could help client PCs and servers
ACE may matter across both consumer and enterprise systems. On laptops and desktops, better AI instructions could improve local AI features, image tools, productivity apps, voice processing, and small model inference.
In servers, ACE could help CPUs handle AI support tasks more efficiently, even when GPUs or accelerators are doing the heaviest work. CPUs still manage data movement, preprocessing, scheduling, and many mixed workloads. Better matrix support could improve those areas.
The exact benefits will depend on how AMD and Intel implement ACE in future chips and how quickly software developers adopt the extensions.
This is part of a larger x86 modernization push
ACE is not the only change coming to x86. AMD and Intel are also working on other extensions such as AVX10 and APX, which are meant to improve performance, compatibility, and future scalability.
Together, these changes show that x86 is being updated for a new era. AI workloads, lower precision math, faster memory, and hybrid CPU designs are changing what processors need to do.
ACE is especially important because AI is no longer limited to cloud data centers. It is moving into PCs, workstations, business software, creative tools, and developer workflows.
Future CPUs will decide how important ACE becomes
The ACE specification is an important step, but its real impact will depend on future CPU launches. AMD and Intel still need to build ACE compliant architectures, ship them at scale, and work with software partners to make sure applications use the new instructions properly.
If that happens, x86 CPUs could become much more capable for AI workloads without relying entirely on separate accelerators.
For now, ACE shows that AMD and Intel understand the problem. AI computing is changing quickly, and x86 needs stronger native support for matrix math and low precision formats. With ACE, both companies are laying the groundwork for future CPUs that can handle more AI work directly and more efficiently.
Discussion (0)
Be the first to comment.