- AMD’s rumored “Sound Wave” chip is claimed to be an Arm-based APU
- It is believed to characteristic “RDNA 3.5+” CUs which are “enhanced” for machine studying
- MLID’s sources allege the chip will characteristic 16MB Infinity Cache
Alleged new info has been uncovered concerning the rumored AMD “Sound Wave” APU, claiming will probably be extra in direction of machine studying, with “enhanced RDNA 3.5” CUs.
The information comes from dependable {hardware} leaker Moore’s Legislation is Useless in a brand new YouTube video, with a protracted listing of up to date specs cited from “Late March 2025”. It is a substantial replace on the earlier rumor that we lined again in March 2024.
Beforehand, it was believed that AMD’s Sound Wave processor can be a Zen 6 cell chipset. Nevertheless, it is now claimed by MILD that the processor might be an Arm-based APU, much like the likes of Qualcomm’s Snapdragon X lineup as a substitute.
Particularly, MLID’s sources declare that AMD’s Sound Wave APU might be constructed on a TSMC 3nm node, focusing on 5-10W type components, with two efficiency cores, 4 effectivity cores (six complete), and 4MB of L3 cache. Crucially, it is alleged that there might be 4 RDNA 3.5 CUs with “improved ML Efficiency” which has been given the title of “RDNA 3.5+” unofficially, to indicate a distinction from the present RDNA 3.5 CUs obtainable on chips.
Its capabilities for AI computing look like a essential focus, as may be evidenced by the supposed “4th Era AI Engine” backed by 16GB LPDDR5X-9600 RAM which is alleged to be for the “Customary” programs. MLID doesn’t know if the newer RDNA 3.5 CUs will enable for FSR 4. Nevertheless, gaming seems to be a secondary concern for this silicon.
An fascinating improvement for the Sound Wave APU is that it’s going to allegedly characteristic 16MB of Reminiscence Entry at Final Stage (MALL) cache (identified to AMD customers as Infinity Cache) which is uncommon for a 5-10W APU. It is curious as a result of “Strix Level” doesn’t have this, which is reportedly “documented” by MLID.
MLID later factors out how the AI engine, CPU, and GPU will all work on the identical controller, so having the additional bandwidth of 16MB Infinity Cache shared between them may very well be useful, even on such a modest chip.
AI computing with out the facility draw
Whereas it is nonetheless early days for Sound Wave, all of the rumors we have heard from the chipset sound as if it is primarily being geared round AI workloads with out the (often) excessive energy attracts of rival {hardware}. 16MB of Infinity Cache does not sound too dramatic an inclusion at first, nevertheless, that is earlier than bearing in mind that the chip is barely rated between 5-10W TDP.
Moreover, the RDNA 3.5+ CUs may very well be an enormous driving think about AI workloads, relying on how optimized they are surely for machine studying. With a 2026 launch date claimed, it is unlikely we’ll obtain official phrase from AMD any time quickly, however we now comprehend it may very well be an optimized Arm-based APU relatively than the chip that ushers in Zen 6 as we beforehand thought.