Can AMD’s Ryzen AI Chips Take on Intel and Nvidia?

A significant change in technology is usually preceded by a certain type of quiet confidence. Something more subdued, not an executive keynote with prepared gasps or a press release bluster. A benchmark figure that defies logic. A specification sheet that requires two readings. That’s the general sentiment surrounding AMD’s most recent Ryzen AI chips, and it’s something to be aware of.

The chip conversation had a fairly predictable script for years. Raw CPU performance was owned by Intel. AI and graphics acceleration were owned by Nvidia. AMD was competitive and occasionally impressive, but it hardly ever established the parameters of the discussion. In 2025, that dynamic is clearly under stress, with the Ryzen AI Max+ series bearing the brunt of that strain.

AMD is placing its trust in the flagship Ryzen AI Max+ 395. It has a 50 TOPS XDNA 2 neural processing unit, up to 16 Zen 5 performance cores, and—this is the part that really ends the discussion—40 RDNA 3.5 Compute Units connected to a unified memory architecture that can pull up to 96GB at 256GB/s bandwidth. There is no typo in the last figure. In the Windows ecosystem, no integrated graphics chip has even approached that figure. In short, AMD has effectively packed workstation-class GPU memory onto a mobile processor and asked if you wanted it in a thin laptop.

However, the comparison that caused the most controversy wasn’t directed at Intel. It was up against Nvidia’s GeForce RTX 4090, which has long been regarded as the unchallenged champion of desktop high-performance AI workloads. According to AMD’s data, the Max+ 395 uses 87% less power while providing up to 2.2 times faster AI performance on a large language model with 70 billion parameters, such as Llama 3.3. It all boils down to memory. There is 24GB of VRAM in a discrete RTX 4090. 96GB of unified memory is available to the Ryzen AI Max+ 395. Large models cause memory to become the bottleneck long before raw compute does. AMD has discreetly developed a chip that avoids the entire issue.

It’s important to be clear that these benchmarks were chosen by AMD and published before extensive independent testing. The chip hasn’t undergone the kind of independent, months-long examination that garners genuine trust, the sample size is small, and the test conditions are carefully chosen. Even so, it is difficult to completely reject the directional claim.

The Core Ultra series is Intel’s most significant architectural rethink in many years. Built specifically around Windows 11’s Copilot+ features, the hybrid design, which combines high-performance P-cores, effective E-cores, and a dedicated NPU, is truly thoughtful engineering. Real-time transcription, AI-assisted image editing, and recall are all handled by Intel’s NPU without causing the main processor to overheat or deplete its battery. After using a Core Ultra 9 laptop for a few weeks, it feels much more comfortable to multitask than earlier models. The tasks that the majority of professionals actually perform throughout the day don’t require much effort. However, Intel’s integrated Arc graphics still lag behind AMD’s RDNA 3.5 in raw GPU-adjacent AI tasks, such as running local models or managing complex creative rendering.

A similar intriguing tale is told by AMD’s more affordable Ryzen AI 7 350, which is positioned beneath the Max+ family. It reduces the number of cores and drops to a 15W base TDP while matching the AI 9 365 in NPU performance and clock speeds. According to AMD, it outperforms Intel’s Core Ultra 7 258V by about 30% in productivity benchmarks and has a battery life that is reportedly longer than 24 hours when playing videos. Even a partial result here would represent a significant shift in what a mid-range laptop chip can deliver, though it’s possible that those numbers are somewhat cut in real-world testing.

In general, AMD is opting to compete on a different level than it has in the past. Instead of attempting to undercut Nvidia on raw GPU throughput in discrete cards or trading CPU benchmark punches with Intel, it is creating chips where memory bandwidth and AI efficiency converge. It’s a wise move, particularly as local AI inference moves from being a theoretical task to a practical one. Not too long ago, it was novel to run a large language model locally on a laptop. It’s becoming less and less.

It is truly unclear if this results in market momentum. Intel has strong ties to both enterprise procurement teams and PC manufacturers. Years of developer loyalty are ingrained in Nvidia’s software ecosystem, especially CUDA, and they won’t just vanish because a benchmark doesn’t look good. And let’s face it, Apple continues to be the standard. In Cinebench multi-threaded performance, the Ryzen AI Max+ 365 loses to the M4 Pro by roughly 3%. This is a close result, but Apple’s chips come in carefully designed machines where hardware and software are tuned together in ways AMD doesn’t fully control.

However, there’s a sense that the conversation’s equilibrium is changing. The once-settled chip war between AMD for value, Intel for productivity, and Nvidia for artificial intelligence is becoming less predictable. The breadth of AMD’s AI lineup was startling when looking through the hardware announcements at this year’s CES: Max+ workstation chips, mainstream Ryzen AI 300 variants for regular laptops, and 11 new Ryzen 200 series entries meant to make light AI workloads affordable for consumers on a tight budget. It’s not a catch-up strategy. It reads more like someone who thinks they’ve found their footing.

When independent reviewers have months to work with these chips inside real shipping products from companies like Asus, Gigabyte, MSI, and others, that will be the true test. In everyday situations, benchmark theatrics tend to become more subdued. However, AMD’s architectural decisions, especially the unified memory strategy, aren’t tricks. They show a sincere awareness of the future of AI computing. That is not insignificant.