Post reply

Name
Email
Subject
Message icon

[quote author=No DLSS-like ML upscaling link=msg=744471 date=1773230544]
[b]AAA gaming[/b]
Know that Strix Halo' has a RDNA3.5 iGPU and as such does not support hardware-based, machine learning (ML), DLSS-like, upscaling. Only RDNA4 does that and it's only available in dedicated desktop GPU cards.

[b]AI/LLM token generation (tg)[/b]
But Strix Halo' RAM is connected to a 256-bit memory bus (equivalent to quad-channel (4 * 64-bit per RAM slot) in a desktop PC/SODIMM laptop PC) and as such has double (actually more, because it's 8000 MT/s memory, too (vs 5600 to 6200 MT/s on a desktop)) the memory bandwidth of normal desktop- and mini-PCs out there as they are all 128-bit (2*64-bit, aka dual-channel), which is necessary for the AI/LLM token generation (tg) speed.
[b]AI/LLM prompt processing (pp)[/b]
Strix Halo' 8060S iGPU is equivalent to a RTX 4060 desktop GPU:
3dmark.com/search:
Radeon 8060S: Average score: 10022
RTX 4060 (desktop): Average score: 10398,
so the pp speed will be equivalent to that of the RTX 4060 desktop.

Man, if only this had 192 GB RAM, because then it could fit the very usable 3-bit quants of huggingface.co/unsloth/Qwen3.5-397B-A17B-GGUF or other big LLMs.
[/quote]

Other options

Return to this topic
Don't use smileys

Verification:

Please leave this box empty:

Shortcuts: ALT+S post or ALT+P preview

Topic summary

Posted by No DLSS-like ML upscaling

- Today at 13:02:24

AAA gaming
Know that Strix Halo' has a RDNA3.5 iGPU and as such does not support hardware-based, machine learning (ML), DLSS-like, upscaling. Only RDNA4 does that and it's only available in dedicated desktop GPU cards.

AI/LLM token generation (tg)
But Strix Halo' RAM is connected to a 256-bit memory bus (equivalent to quad-channel (4 * 64-bit per RAM slot) in a desktop PC/SODIMM laptop PC) and as such has double (actually more, because it's 8000 MT/s memory, too (vs 5600 to 6200 MT/s on a desktop)) the memory bandwidth of normal desktop- and mini-PCs out there as they are all 128-bit (2*64-bit, aka dual-channel), which is necessary for the AI/LLM token generation (tg) speed.
AI/LLM prompt processing (pp)
Strix Halo' 8060S iGPU is equivalent to a RTX 4060 desktop GPU:
3dmark.com/search:
Radeon 8060S: Average score: 10022
RTX 4060 (desktop): Average score: 10398,
so the pp speed will be equivalent to that of the RTX 4060 desktop.

Man, if only this had 192 GB RAM, because then it could fit the very usable 3-bit quants of huggingface.co/unsloth/Qwen3.5-397B-A17B-GGUF or other big LLMs.

Posted by Sey

- Today at 09:38:17

Why didn't you compare it to the results of the Bosgame M5, which was your first test of a mini PC equipped with a Strix Halo processor? In my opinion, the Bosgame M5 is still better in almost every way.

Posted by Redaktion

- Today at 09:00:33

The Acemagic M1A Pro+ combines powerful high-end hardware with an eye-catching case: Equipped with the AMD Ryzen AI Max+ 395, 128 GB RAM and a powerful Radeon graphics unit, the mini PC offers plenty of performance for gaming and professional workloads. In this test, we show how the system performs in benchmarks, games and under continuous load - and whether it is one of the most powerful mini PCs in its class.

https://www.notebookcheck.net/AMD-Ryzen-AI-Max-395-with-128-GB-RAM-in-a-cube-Is-the-Acemagic-M1A-Pro-the-new-performance-king-among-mini-PCs.1246642.0.html