APU Strix Halo has 32MB MALL cache, 60 TOPS AI processor and USB4 DP2.1 UHBR20

APU Strix Halo has 32MB MALL cache, 60 TOPS AI processor and USB4 DP2.1 UHBR20
APU Strix Halo has 32MB MALL cache, 60 TOPS AI processor and USB4 DP2.1 UHBR20
--

The first one will be released at the end of this summer Zen 5 APU Strix Point and it will be cheaper at the beginning of next year Kraken and more expensive Strix Hello (Sarlacc).

32MB MALL cache

The positive news is that the implementation of a faster AI processor (Microsoft’s requirement for AI PCs), which in other respects negatively affected this year’s mobile products from AMD and Intel, has Strix Halo she probably didn’t touch it. Recall that the APU Strix Point it was originally supposed to be equipped with a 16MB MALL cache (=SLC), but due to the implementation of a more powerful (larger) AI accelerator, something had to give way, and that was precisely this cache. When Strix Halo (Sarlacc), however, the MALL cache was preserved, in a capacity of 32MB.

Internal AMD slide posted by HKEPC website

MALL (Memory Attached Last Level) Cache or SLC (System Level Cache) is the last level cache, which in the case of APU is shared by both CPU and GPU and other possible circuits (e.g. integrated accelerators). In 3D load, it will probably serve mainly the graphics core (Infinity Cache style), in processor load, it will probably behave like a slightly slower V-cache with half the capacity.

60 TOPS NPU / AI accelerator

The AI ​​accelerator is known to be 4x larger than the APU Phoenix / Hawk Pointbut the unknown was the clock frequency, which (as the difference between Phoenix and Hawk Point) can move in a very wide range. According to the leaked slide, AMD is planning 60 TOPS performance, which is quite a lot.

SoC year AI Accelerator Performance (NPU)
AMD Phoenix 2023 10 TOPS
Apple M3 2023 18 TOPS
Intel Meteor Lake 2024 10 TOPS
AMD Hawk Point 2024 16 TOPS
Qualcomm Snapdragon X 2024 45 TOPS
AMD Strix Point 2024 50 TOPS
Intel Arrow Lake 2024 13 TOPS
Intel Lunar Lake 2024 45 TOPS
AMD Strix Halo 2025 60 TOPS

In essence, it can be said that AMD will push NPU performance to six times in a span of less than two years.

USB4 DisplayPort 2.1 UHBR20

AMD introduced DisplayPort 2.1 on Radeon RX 7000. The specification of the second generation of this interface enables the support of three transfer rates: UHBR10, which with a transfer rate of 40 Gb/s is less than a quarter faster than DisplayPort 1.3 / 1.4, and UHBR13.5 with 54 Gb/s (2/3 faster than DisplayPort 1.3 / 1.4) and finally UHBR20 with 80 Gb/s (approx. 2.5x faster than DisplayPort 1.3 / 1.4). Current RX 7000 Radeons support up to UHBR13.5, thanks to which (as the only one on the market) they can handle, for example, a 240Hz refresh rate on the Samsung Odyssey Neo G9 display.

Strix Halo will go a step further, as it supports DisplayPort 2.1 in UHBR20 when using USB4 Alt-mode, i.e. the maximum possible speed of this interface. In practice, this means the possibility of connecting, for example, a 16K display in full-color resolution with HDR, or, for example, a 3x 10K display in full-color resolution with HDR at 60 Hz, or a 3x 5K display in full-color resolution with HDR at 120 Hz, and the like.


It looks like AMD Strix Halo equipped richly in all respects. With 60TOPS NPU, support for the fastest USB version, the fastest DisplayPort version, the fastest HDMI version and a large MALL cache, it is hard to find what more fundamental weakness this solution could have. Let us recall that the processor part will consist of two chiplets with a total of 16 architecture cores Zen 5 and graphics with 2560 RDNA 3.5 stream-processors will bring at least 3x higher FPS than the current solution on Phoenix / Hawk Point. Well, actually one weakness presents itself: It won’t exactly be cheap.


The article is in Czech

Tags: APU Strix Halo #32MB MALL cache TOPS processor USB4 DP2 .1 UHBR20

-

NEXT Radeon RX 8900 XTX could have been a card with 50% more Shader Engine count