Tuesday, December 10th 2024
AMD Ryzen AI MAX+ PRO 395 "Strix Halo" APU Spotted in Geekbench Leak
CES 2025 is less than a month away and leaks about AMD Strix Halo APUs are starting to emerge. Today we have confirmation via a leaked Geekbench Vulcan test that AMD will launch the Ryzen AI MAX+ PRO 395 w/ Radeon 8060S "Strix Halo" APU. Information reveals that this APU is equipped with 16 Zen 5 cores with 32 threads, a 3 GHz base frequency (4.4 GHz max), and a boost up to 5.1 GHz. It sports a 32 MB L3 cache per CCD for a total of 64 MB since it uses a dual CCD chip design. The TDP should be between 55-130 W. Moreover, the "PRO" in the product naming suggests that AMD could release non-PRO models at a later date.
The integrated Radeon 8060S iGPU adopts the RDNA 3.5 architecture with 40 computing units and was tested using an AMD reference board design codenamed AMD MAPLE-STXH and 64 GB of memory scoring 67,004 points in the Geekbench Vulkan test. This initial result is lower than AMD Radeon RX 7600 RDNA 3 discrete entry-level products (despite having more cores 40CU vs 32CU), and higher than NVIDIA GeForce RTX 3050. However, since the benchmark was run on an evaluation platform without optimized drivers and more likely using an early test sample product, we can expect the actual performance of the Radeon 8060S iGPU to be higher.
Source:
Videocardz
The integrated Radeon 8060S iGPU adopts the RDNA 3.5 architecture with 40 computing units and was tested using an AMD reference board design codenamed AMD MAPLE-STXH and 64 GB of memory scoring 67,004 points in the Geekbench Vulkan test. This initial result is lower than AMD Radeon RX 7600 RDNA 3 discrete entry-level products (despite having more cores 40CU vs 32CU), and higher than NVIDIA GeForce RTX 3050. However, since the benchmark was run on an evaluation platform without optimized drivers and more likely using an early test sample product, we can expect the actual performance of the Radeon 8060S iGPU to be higher.
29 Comments on AMD Ryzen AI MAX+ PRO 395 "Strix Halo" APU Spotted in Geekbench Leak
We currently have the Apple M series, Intel Core Ultra (Lunar Lake) series and the AMD Ryzen AI (Strix Point and Strix Halo) series. They all compete at overlapping power, CPU and GPU levels. Eventually, it would be nice to find a good way to benchmark and review these SoC products. But with a wide difference in cooling, form factors and ISAs, conventional benchmarking is extremely difficult.
Max AI to the max on your laptop!
130w though, interesting.
A fully monolithic device (like strix point) with a bigger memory bus would be a really nice middle-ground to have.
Edit: And speak of the devil, we have a leak of a RTX5000 mobile GPU in a Dell.
Dell Pro Max 18/16 Plus laptop leak confirms Core Ultra 200HX CPU, RTX 5000 GPU and CAMM2 memory - VideoCardz.com
That's 255W total CPU+GPU in a < 1 in form factor and THREE!!! fans working in parallel. Half of that power (~130W) would easily fit in a mini PC.I wish this was going to be a standalone, socketable part but I highly, highly doubt we will see any of these outside of OEM parts.
Intel could've gone this way with LNL, but I guess not doing so makes sense since it's meant to be a low power chip (competitor to the base Mx chip).
Strix Halo kinda can compete with a M4 Max CPU-perf wise, but it's a 100W+ chip vs a ~70W one, with the latter having a 384~512-bit memory bus vs 256-bit from AMD (which is the same value for the new Pro chips). The GPU might be beefier, but it'll have way less bandwidth, and I doubt it'll be any cheap.
I guess ideally we would have something like what we already have with strix point, but with a 192 or 256-bit memory bus, and maybe a little bigger GPU (like 20~24CUs). Would still be able to fit within a 50~70W envelope, and would be a really interesting option for the $1~2k range of premium lightweight laptops with a long battery life.I don't think this chip will have a desktop variant at all, way too big and different memory subsystem to the current offerings.
When its the correct type (Intel, Ngreedia) then its not a problem.:DSee above...:roll:
This particular halo APU would be great in the form of a SOC like the Apple Silicons, with memory integrated in the same chip.
Yes, it will be beyond expensive, but given how Ngreedia got away with their halo offerings prices, who knows.
Or, maybe using quad memory channels, instead of dual.
An APU is a monolithic chip. This has several chiplets.
en.wikipedia.org/wiki/AMD_APUYup ^^^
Since it is using DDR5, which is dual channel per dimm (or something like that, bit rusty in the subject) thats why its showing as such.
But I dont know if this can be compared to a proper quad channel system, like whats used with ThreadRippers, for example.
It also uses something akin to "quad channel", since it's a 256-bit bus, double of what you have on your consumer desktop (128-bit at dual-channel), that's the most interesting part of this product IMO.DDR5 muddied the terms a bit.
Before DDR5, there was kind of a convention to call a 64-bit bus a "channel", since that was the norm for what we had in a single stick/mobo slot. 2 sticks in different channels? 128-bit, or the so called "dual-channel".
Now with DDR5 each DIMM/channel has 2x "sub-channels" of 32-bit each, but your total bus size is still 128-bit when using 2 sticks, so nothing changed.
Threadrippers, as an example, have a bus width of 256-bit (4x64-bit) for the non-Pro variants, and 512-bit (8x-64-bit) for the pro ones.
Strix halo has a 256-bit bus, so it'd be similar to what we call a quad-channel system.