The AMD Ryzen AI 9 HX 370 Review: Unleashing Zen 5 and RDNA 3.5 Into Notebooks
by Gavin Bonshor on July 28, 2024 9:00 AM ESTASUS Zenbook S 16: Power Consumption & Performance Modes
A big part of any laptop's performance is its TDP – how much energy the chassis can consistently dissipate – meaning that two laptops can have wildly different performance, despite using the same chip, just by virtue of their cooling capabilities. So measuring the peak and sustained power consumed by a laptop is particularly important, both to get an idea of where it falls on the ultrabook-to-DTR spectrum, but also how a design compares to other laptops. Especially with an architecture launch like this one, we need to know if a performance lead comes from architecture and efficiency, or just running a lot of power through a chip.
The AMD Ryzen AI 9 HX 370 has a rather large configurable TDP range overall, with the chip capable of being set to anything between 15 and 54 W. This gives Strix Point a wide berth to fit in everything from ultrabooks to larger gaming notebooks, but it also means that the chip's performance, particularly in heavily multi-threaded and mixed CPU/GPU workloads is going to be defined by its TDP.
Despite its 16-inch frame, In the ASUS Zenbook S 16 is configured rather conservatively. ASUS has taken what's nominally a 28W chip and dialed it down to 17W for it's out-of-the-box experience. Dubbed "Standard Mode" the out-of-the-box experience is a laptop that's highly efficient and highly quiet, but does leave some performance on the table for the sake of acoustics.
ASUS also gives the user quite a bit of latitude to pick performance modes here, with the Zenbook offering 4 different modes altogether. Besides the standard mode, there's the ultra quiet "Whisper mode", the performant "Performance Mode" with a 28W TDP, and the no-holds-barred "Full-speed Mode".
For most of the Zenbook S 16 you'll see today – ours included – you'll find these laptops are running in the 28 Watt Performance Mode. This comes by request of AMD, who is looking to show off what the chip can do at its standard TDP, rather than ASUS's dialed-down TDP. And while we're not beholden to this (or any other) request from AMD, from a pragmatic standpoint it's a lot easier to find 28W laptops than it is 17W laptops. So for our first round of testing, we have dialed up our laptop to its 28W mode.
In our peak power test, the Ryzen AI 9 HX 370 ramped up and peaked at 33 W. We typically see a higher peak power value than the TDP before things settle down during an intensive workload, although the Intel Core Ultra 7 155H within the ASUS Zenbook 14 OLED did manage to peak at 64 W for a very brief moment, which is way out of spec.
Looking at the power of the Ryzen AI 9 HX 370 when using Maxon's latest Cinebench 2024 (multi-threaded test) benchmark, which is quite an intensive rendering workload, we can see how it compares to the Ryzen 9 7940HS when dialed down to 35W, as well as the Intel Core Ultra 7 155H which is running at 28W within the ASUS Zenbook 14 OLED. Aside from the very large spike up to 45W on the Core Ultra 7 155H, we see a pretty consistent level of power usage from the CPU package (the cores).
Focusing on the Ryzen AI 9 HX 370, we can see it quickly ramped up and achieved a very consistent load level, bouncing between 32 and 33 W throughout the Cinebench 2024 MT benchmark loop. As the benchmarking loop carried on, power consumption slowly dropped to a steady-state TDP of 28W, matching the configured TDP of the laptop. The slow and gradual drop in power towards the end was very smooth. The same can be said about the Ryzen 9 7940HS. In contrast, the Intel Core Ultra 7 155H produces some slightly noisier data, bouncing around between 26.5 W and 29 W throughout the test.
Facebook
Twitter
RSS
72 Comments
View All Comments
Khanan - Tuesday, July 30, 2024 - link
It’s practically maxed out, these are high end mobile parts. However there will be a APU later with an extra big GPU called Strix Halo, the only difference will be that it will have a mid range desktop level GPU with over 32 CUs - comparable to console APUs.mczak - Tuesday, July 30, 2024 - link
You can find annotated die shots for Strix like here:https://www.techpowerup.com/325035/amd-strix-point...
The die area spent for the NPU is quite significant, someone estimated around ~14mm².
As for Phoenix, I've never seen a die shot of it (only Phoenix2 which doesn't feature the NPU). According to AMD, the NPU in Phoenix though was made up of 5x4 tiles, whereas the one in Strix is 8x4 tiles (with twice the MAC throughput per tile, and other improvments), as can be seen in the slides from AMD: https://www.anandtech.com/show/21469/amd-details-r... - so maybe it was like half the size? But that's just a crude guess.
zamroni - Tuesday, July 30, 2024 - link
the npu is overkill.i'd rather have bigger igpu or more cpu caches.
gpu can do anything that npu can anyway
Khanan - Tuesday, July 30, 2024 - link
It’s not, as per MS it needs to be at least 40 TOPS fast and AMD made sure to be faster than the competitors, it’s a marketing win. And copilot doesn’t run on CPU or GPU as that costs too much energy on the go, and in general it won’t use those, even if you’re on desktop and have no battery constraints.jmke - Tuesday, July 30, 2024 - link
> . The first to market in the consumer space with a dedicated AI engine in the notebook space was AMD with their Zen 4-based Phoenix Point or the Ryzen 7040 seriesRyzen 7040 released March 2023
Macbook M1 with dedicated ML chip released in 2020, that one already has dedicated ML chip with support for a slew of models, including text and image (stable diff); ....
jeromec - Tuesday, July 30, 2024 - link
Did I miss the battery test?And why no comparison with the Nvidia-powered ARM Windows PCs? Was it an AMD request?
Also, which model of MacBook Pro is included in the benchmarks? Apple's naming convention is currently very confusing and there are 14" MacBook Pros with M3, 2 flavors of M3 Pro and 2 flavors of M3 Max.
Also, I do not see the point of comparing, weight-wise, this Asus with a 16" MacBook Pro with M3 Max, which is in a different league, performance and price-wise.
lmcd - Tuesday, July 30, 2024 - link
Nvidia-powered ARM Windows PCs?I'm so sorry to inform you that the last Nvidia ARM Windows PC was the Surface RT 2, which launched with a Tegra 4. We have not seen anything since.
Hopefully that will change, but seriously: what are you talking about?
nandnandnand - Tuesday, July 30, 2024 - link
There's rumors of a MediaTek/Nvidia chip coming within a couple years, but jeromec meant the Qualcomm Snapdragon X Elite.patel21 - Tuesday, July 30, 2024 - link
Minor Correction:>> Core Ultra 7 155H, which is a 6P+8E/22T chip
It should be:
Core Ultra 7 155H, which is a 6P+8E/20T chip
emn13 - Wednesday, July 31, 2024 - link
The argument to use LLVM for SPEC based on cross-platform ease is not a good one. Compilers don't perform equally competitively across different platforms. A compiler that works well on, say, apple silicon need not be optimal nor typical elsewhere. Fortunately, LLVM is fairly broadly liked, so hopefully this isn't a huge impact, but we don't know. Could be this alone pretty invalidates the spec2017 resultsAdditionally, particularly for *new* platforms the exact version of the compiler and whatever tuning it has had can make a difference; it's easy to test with a compiler that's older than the platform and thus tunes poorly. As it happens the LLVM version listed - 10.0.0 is simply ancient; over 4 years old - the current version is 18. A bit of web searching suggests that at least until fairly recently no znver5 tuning was merged into even dev LLVM; GCC happened to have tuning merged in march (not sure if this is typical between those two compilers).
Finally, the text implies you disabled avx512. I'm not sure why; even on zen4 with its half-rate execution path this can have considerable performance advantages; it pretty plausible it will on zen5 too (even on models with half-rate execution rates such as mobile chips).
All in all, I'm not sure what the point of the spec result is in this form. On the one hand, the vast majority of software run is compiled targetting old processors, so that's realistic. On the other hand, you already have benchmarks for that, and kind of the whole allure of a compiled-from-scratch benchmark like SPEC is that you can see what the platform is capable of natively, not when running legacy software. It's not too different from how it's interesting to see ARM chips running both emulated x86 code, but also see native results even if that's not 100% representative.
Still I guess an interesting result, but those caveats are large and worth mentioning IMHO. It's very likely a more reasonably tuned run could be significantly different for all of these chips, particularly newer ones that deviate most from the platform described by -march=x86-64 -mtune=core-avx2 -mfma -mavx -mavx2 - is that haswell? It's ancient anyhow.