AMD's Strix Halo Zen 5 APU tested in Geekbench AI benchmark — Ryzen AI Max 390 sample falls behind Ryzen 7 7840HS
The Strix Halo sample does not shine in this particular benchmark.
One of AMD's Ryzen AI Max 300 (Strix Halo) APU has gone through Geekbench AI, the AI-focused counterpart to Geekbench 6. Discovered by Everest on X, the Ryzen AI Max 390 (AMD Eng Sample: 100-000001421-50_Y) was tested, but only on the Zen 5 CPU cores, not the GPU or NPU.
As always, take the Geekbench result with a pinch of salt since we don't know the state of the silicon. According to the submission, the Ryzen AI Max 390 is inside HP's ZBook Ultra 14-inch G1a, which appears to be a mobile workstation device, so the cooling should be adequate.
Using all 12 CPU cores, the Ryzen AI Max 390 outputted a single precision score of 4,733 points, a half-precision score of 4,944 points, and a quantized score of 13,944 points. Testing was accomplished on Geekbench AI 1.1.0 using the OpenVINO framework.
Since the benchmark was relegated entirely to the CPU cores, AI performance was not as good as what the chip would be able to do on the GPU or NPU. For example, the chip's quantized score was inferior to that of Intel's entry-level Arc A310 graphics card, which outputted a score of 15,453 points. It was also worse than some Core Ultra 7 Meteor Lake integrated graphics chips, with an even higher score. A similar story also goes for the single-precision and half-precision scores.
The Strix Halo chip failed to impress when looking specifically at CPU scores in an apples-to-apples comparison. Out of the relatively few OpenVINO CPU scores I was able to find, the Ryzen AI Max chip outperformed AMD's previous generation eight-core Ryzen 7 7840HS Zen 4 mobile CPU, which outputted a single precision score of 5,099 points, a half-precision score of 5,118 points, and a quantized score of 14,680 points.
CPU: | Single Precision Score: | Half Precision Score: | Quantized Score: |
---|---|---|---|
Ryzen 7 7840HS | 5,099 | 5,118 | 14,680 |
Ryzen AI Max 390 | 4,733 | 4,944 | 13,944 |
Core Ultra 7 165H | 3,011 | 2,955 | 7,110 |
Core Ultra 7 165H (2nd listing) | 2,401 | 2,393 | 5,798 |
Ryzen 7 5800X3D | 5,428 | 5,522 | 8,115 |
Xeon W-3175X | 10,636 | 11,074 | 14,102 |
I also took the liberty of testing my personal Ryzen 7 5800X3D to provide a desktop reference. It also outperformed the Strix Halo APU (for the most part). The 5800X3D featured a single precision score of 5,428 points, a half-precision score of 5,522 points, and a quantized score of 8,115 points. The quantized results were the only area where the Strix Halo chip was at an advantage.
From what I could find, the only chips that the Ryzen AI Max 390 can consistently outperform are Intel's Meteor Lake notebook chips, such as the Core Ultra 7 165H. Two listings revealed single-precision and half-precision scores below 3000 points and quantized scores below 7200 for this CPU model.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Looking at CPU-only results for AI performance is not as useless as it might seem. When AI workloads can tax the entire chip (CPU, NPU, and GPU), having capable AI performance on the CPU can be beneficial. Regardless, we are disappointed the benchmarker did not test the NPU and monster-integrated GPU inside the Strix Halo chip, which is far more critical for AI-focused workloads.
The Ryzen AI Max 300 series, aka Strix Halo, is an upcoming APU lineup focused on high-performance computing for mobile devices. Strix Halo is slated to arrive in early 2025.
Aaron Klotz is a contributing writer for Tom’s Hardware, covering news related to computer hardware such as CPUs, and graphics cards.
-
abufrejoval To my understanding, Strix Halo isn't really an APU, as those have been strictly monolithic.Reply
Keeping with doing a nicely big range of products with a small number of constant parts (Lego), it's really just a new and vastly bigger IOD complemented by rather normal Zen 5 CCDs, perhaps even V-cache variants, eventually.
That hefty new IOD contains the console-grade iGPU, an improved LPDDR5 optimized RAM controller, as well as what might still be needed for PCIe, USB and other IOD stuff, but everything else should be pretty standard and somewhat constrained by the overall power limits of the entire package.
I'd hazard there won't be an AM5 socket variant, as attractive as that might be, because it could be too difficult to fit the expanded IOD. But I sure wouldn't mind it if they made that happen.
But back to the perhaps "disappointing" CPU numbers:
Strix Halo is all about doing the very biggest iGPU possible with what the very best but otherwise normal DRAM can deliver today. And the M1 has shown, that this can be rather a lot, if you play with speeds and the number of channels.
But even when going extra wide and fast on DRAM, that gets you at most into mid-range in terms of VRAM dGPU competitors. And at that point real-world gaming performance isn't limited by CPU power.
I'd think that Strix-point shouldn't really deliver leading edge CPU performance, because every extra Watt is better invested on the GPU side for its target market.
Now, with everything doing dynamic power management and Zen 5 CCDs in the package it might not do badly on CPU-only benchmarks as the final product, as long as the iGPU isn't competing, and if hot-spots and overall power limits aren't hitting any ceiling.
But these early numbers might indicate better how much CPU power will be left when most of it goes towards the GPU on the IOD.
And that might just be plenty enough. -
Notton It says right there in the sku, "engineering sample"Reply
They underperform and do not reflect the final product. -
Amdlova For tdp reasons the cpu will be very limited... they will place all on the GPU. I Don't expect A cpu monster but a meh... Here the 35w cpu struggles with some titles need to increasingly the tdp/w target to 54w to keep with gpu. =]Reply
Will be a muit purpose cpu+IA trainer -
usertests
No, it should be considered an APU, or even a "mega APU".abufrejoval said:To my understanding, Strix Halo isn't really an APU, as those have been strictly monolithic.
Chiplets were eventually going to come to AMD APUs. Intel has moved their mainstream and efficiency mobile parts (Meteor Lake, Lunar Lake, Arrow Lake) to chiplets. Although in Lunar Lake, CPU/GPU end up on the same "tile".
It was rumored that Strix Point could have used chiplets but was moved back to a monolithic design. But it's possible that we will see chiplet-based Zen 6 mainstream APUs.
AMD itself said it doesn't consider the Ryzen 7000/9000 desktop CPUs to be APUs despite the presence of iGPUs, simply because the graphics is so weak. The definition of APU is arbitrary. So I am calling Strix Halo a "mega APU". It has current-gen desktop-class graphics and a doubled memory bus, which would seem to qualify.
With 16-core Zen 5 over two chiplets and relatively high TDPs (>100W), it should be able to impress when needed, and should be at least comparable to "Fire Range" (Zen 5 upgrade of Dragon Range).abufrejoval said:I'd think that Strix-point shouldn't really deliver leading edge CPU performance, because every extra Watt is better invested on the GPU side for its target market.
But this is an engineering sample benchmark. Nothing to see here. -
abufrejoval
Yeah, nomenclature needs to evolve and it's getting complicated...usertests said:No, it should be considered an APU, or even a "mega APU".
Chiplets were eventually going to come to AMD APUs. Intel has moved their mainstream and efficiency mobile parts (Meteor Lake, Lunar Lake, Arrow Lake) to chiplets. Although in Lunar Lake, CPU/GPU end up on the same "tile".
I'd have said that a Lunar Lake approach, where chiplets are connected via "silicon class interconnects" vs. organic die carriers, leave an APU moniker intact, because power requirements, latencies etc. are not too different to staying on chip.
An organic die carrier is very much like a mini-mainboard so you need to put in amplifiers, extra synchronization gates and perhaps even protocols and transformations to make that work. To me that's no longer an APU.
But then APU as a term doesn't really add any value or useful discrimination to SoC and thus may disappear soon.
How we'll name and distinguish all those different interconnects and packaging will be a complete mess, I'm sure, but the technology amazing. -
Xajel abufrejoval said:To my understanding, Strix Halo isn't really an APU, as those have been strictly monolithic.
It's an APU, maybe a Super APU also.
The reason AMD didn't consider Ryzen 7000 series an APU isn't because it's chiplet and not monolithic, instead because its iGPU is considered very basic and low-performer for the whole package to be considered an APU.
The iGPU was designed only for basic video output (and a low power video engine). In theory, Ryzen 7000 systems could disable the dGPU altogether for any desktop workload, including browsing, office work & video viewing, but the whole platform was not designed for such functionality (it needs MUX switch and internal video routing between the dGPU and the iGPU/integerated display output, just like how laptops works).
I would love for AMD, Intel & NVIDIA to work on this for one to two monitors as it will allow much better idle and standby power consumption. -
DS426 I agree, Strix Point Halo is still an APU, just a mega, super, etc. APU as others have said. The main difference starts with what the CPU product is targeted and marketed for, not necessarily specific technical aspects. APU's target mobile and USFF/NUC while the CPU moniker remains intact for desktop-class products. Sure, in a sense, Ryzen 7000 and 9000 are APU's, but they aren't formally said to be such because they aren't intended to run without discrete GPU's. Then there's the 8700G and siblings which kind of straddle the middle, though still APU's.Reply
Also echoing others, AMD APU's were to eventually adopt a chiplet architecture -- it just hasn't been necessary and appropriate until now as the latest in node efficiency is needed for these to operate within their TDP envelope, especially considering the strength of the iGPU. -
subspruce
The MI300A is a chiplet APU for servers.DS426 said:I agree, Strix Point Halo is still an APU, just a mega, super, etc. APU as others have said. The main difference starts with what the CPU product is targeted and marketed for, not necessarily specific technical aspects. APU's target mobile and USFF/NUC while the CPU moniker remains intact for desktop-class products. Sure, in a sense, Ryzen 7000 and 9000 are APU's, but they aren't formally said to be such because they aren't intended to run without discrete GPU's. Then there's the 8700G and siblings which kind of straddle the middle, though still APU's.
Also echoing others, AMD APU's were to eventually adopt a chiplet architecture -- it just hasn't been necessary and appropriate until now as the latest in node efficiency is needed for these to operate within their TDP envelope, especially considering the strength of the iGPU.