AMD Rebrands Next-Gen Mobile CPUs ‘Ryzen AI’, Claims 50 TOPS NPUs

Just six months ago, AMD (and Intel) touted its “AI PC” laptop processors with the company’s second-generation Neural Processing Module for low-power AI acceleration—and they already were last year. Now it announces new mobile processors with a theoretically much better third-generation NPU (XDNA 2, with up to 50 TOPS), embedded in an updated processor using the new generation Zen 5 process and an upgraded integrated GPU (RDNA 3.5). Rebranded Ryzen AI, AMD is launching with several 300-series parts ready for Copilot Plus.

Previously codenamed Strix Point, the Ryzen AI 300 series is the successor to the Ryzen 8000 generation; since the older chips don’t have the NPU bandwidth of the 300s — they’re pretty limited — they’ll keep their current naming.

Laptop launch partners for the new processor are the same Copilot Plus PClike the ones Nvidia had for its RTX AI PCs earlier in the day, mostly from Asus.

What’s in a name?

Arriving in laptops from July, Ryzen AI chips retain the same basic naming conventions, only with AI enabled and modifiers like HX now used to denote market tier rather than power consumption class.

AMD launched with two flagships, the Ryzen AI 9 HX 370 and the Ryzen AI 9 365, which differ in the number of CPU and GPU cores.

The new Zen 5 processors actually appear to be split into 4nm Zen 5 cores and 3nm Zen 5c cores; AMD doesn’t provide any details about the architecture, so it’s unclear why the processor mixes the two, although this isn’t unprecedented for AMD (it also mixes on its graphics cards).

This could be a way to match the performance/efficiency cores of the Intel and Apple chips for additional power savings or thermal management. Because the new chips are designed to serve a wide range of laptop types, with power consumption from 15 watts to 54 watts (small ultralights to larger thin and light), they can be structured to allow manufacturers more granularity in defining the total power of the system. As you’d expect, there are uneven core counts, frequencies, and cache sizes, which should result in at least the typical generation-by-generation performance improvements.

The integrated Radeon 890M and 880M GPUs aren’t a huge leap over the last generation, but the 890M sets a new high for the number of compute units. The previous top-of-the-line was the 780M with 12 CPUs, and now there’s the 890M, which increases the maximum number of GPU CUs to 16.

Watch this: Watch everything announced from the AMD Keynote at Computex 2024

12:30 p.m

NPU Hopping

If you’re wondering why Qualcomm and Microsoft made theirs Launch Copilot Plus two weeks before Computex, my theory is that it would otherwise have been overshadowed by AMD (and maybe intel?) — because, among other reductionist talking points, 50 TOPS is more than 45 TOPS. TOPS, a measure of integer math performance on the NPU, really only matters to Microsoft because that’s its primary criteria for being considered a Copilot Plus computer. Copilot uses only integer math to perform tasks such as summarizing, plotting, quick and dirty background removal in video conferencing, tracking everything you do, and more.

But integer math is quite limited if it has to perform calculations with very large or very small numbers, such as those needed to generate AI for image and video output. They use floating point calculations (which refers to how numbers are manipulated in computer registers). Most existing applications that have implemented AI tools or features—creative, gaming, and otherwise—already use 16-bit floating point (called “single precision”) and, to a lesser extent, 32-bit FP (double precision).

That’s all to say that there’s another aspect of consumer CPU NPUs where Qualcomm has just been overtaken by AMD: Floating-point blocking. Essentially, it’s a way to compress 16-bit floating-point values ​​so that they can be manipulated potentially as fast as the 8-bit integers used by the NPU, theoretically without losing 16-bit precision.

AMD also debuted new Ryzen 9000 X-series desktop processors using the Zen 5 architecture. They still use the AM5 socket, so you can upgrade old systems with them, but if you replace your motherboard with one of the new models featuring the X870/X870 chipsets, you’ll gain USB 4.0, PCIe 5 and higher speeds for EXPO at AMD (profiles used for one-click overclocking).