Zen 5 APU specs: 32MB MALL cache and 70 TOPS NPU Strix Halo confirmed

Granite Ridge (desktop), Fire Range (laptops)

Let’s start with chiplet processors. There, the basic form is known and essentially nothing changes. The slide that the leaker and computer seller GoldenPigUpgrade released to the world on the BiliBili platform, however, contains one discrepancy, precisely in relation to chiplet processors. It states that these processors contain small integrated graphics with 2 CU (128 stream-processors) of the RDNA 3 architecture and further that the central chiplet (IOD) is taken from the current generation (Raphael). There is clearly an error in one or the other information, because the central chiplet from the processors Raphael it has RDNA generation 2 (not 3) graphics. But that’s just a formality, from a performance point of view it doesn’t matter (there would be a greater benefit up to RDNA 3.5).

Strix Point APU

Leaving aside the well-known fact that it is a 4nm monolith with 4 cores Zen 5 and 8 cores Zen 5c equipped with 24MB L3 cache, we learn that it is expected to have a clock frequency of 5.1GHz, which is the same value as the current APU Hawk Point. So single-core performance will only shift IPC, multi-core a combination of IPC and a higher number of cores. The integrated graphics will run at around 2.6-2.7 GHz. The integrated AI accelerator (NPU) will offer a performance of over 40 TOPS (according to another leak around 50 TOPS).

The first wave of laptops will be exclusively equipped with LPDDR5x memories. From AMD’s point of view, this is understandable, the company clearly does not want any of the manufacturers to bury the first wave of reviews by installing slow DDR5 modules, or even arriving with their single-channel configuration. A version with DDR5 will therefore appear later. The most powerful model will be called the Ryzen AI 9 HX 170. Although the company could have stuck with the existing nomenclature for the last year, but then it would have run out of numbers, so it has apparently decided to take advantage of the existing AI mania, amplified by Microsoft’s Windows AI PC certification, and will proceed with the change already this year.

CPU and APU specifications with Zen 5 cores (GoldenPigUpgrade)

APU Strix Halo (Sarlak)

A large chiplet APU, which will be especially interesting due to the presence of integrated graphics (many times more powerful than the current ones), is coming at the beginning of next year. Two processor chiplets with eight cores Zen 5 (i.e. 16 cores in total) are supposed to be different from those intended for separate processors in an unspecified way. The presence of 32MB Infinity Cache / MALL is confirmed, and the AI ​​accelerator is supposed to reach over 70 TOPS performance, which is quite a lot (the slide that appeared in late April talked about “up to 60 TOPS”).

Kraken

Smaller Zen 5 APU, which will be a performance replacement for the existing ones Phoenix / Hawk Pointis called here Crazy (with “and”) instead Kraken whose Kraken, however, this entry has already appeared in the past. Combination 4× Zen 5 + 4x Zen 5c with 16MB L3 cache matches previous leaks. The only difference is the performance of the AI ​​accelerator, which should exceed 40 TOPS, while according to older data it should have been roughly half. It’s possible that AMD changed the target with Microsoft’s requirements in mind. After all, according to the leaker, Bondrewd had the original Kraken look even more different, because due to the release scheduled for 2025, RDNA 4 graphics were originally planned (in the end it will be RDNA 3.5, just like the others Zen 5 APU).

Release of desktop models Zen 5 is expected sometime during the summer holidays, APU Strix Point around September and Strix Halo + Kraken in January at CES 2025. At least we should learn more about this year’s products in early June at Computex.