I guess I’ll stay on 7840U with BIOS 3.05, hopefully it’ll last for a decade
So, I have been using the HX 370 for nearly two days now (arrived earlier). The heat output is at least +12 degrees Celsius compared to the 7640u and 7840u. The machine’s (mostly idle) baseline temp is ~52+ degrees compared to ~40+ degrees for the 7040 series.
Weirdly, the fan doesn’t spin up as much and so it’s quieter normally. But, when ramped up, this mainboard is definitely noisier than the 7040. Another interesting thing I noted is that the 7840u’s temp would ramp up quite sharply from around 40 to ~60 degrees when browsing a heavy site like twitter.com. HX 370 barely touches 58-59 degrees for twitter.com, for example, even though its general baseline is 52-55 degrees.
This machine feels significantly more powerful and handles snappier. But, the heat output and fan noise (when ramped up) is not ideal. This is expected to improve over time, as many of you have noted.
I’ll post another update or two when I have something interesting to add.
2 Likes
That may be a side-effect of the higher idle power rn. Do you have numbers for that? There is quite a difference of being 10 degrees hotter idling at 4w or at 8w. The thermal solution in the new one is supposed to be more powerful so I’d not expect it to be significantly hotter at the same power levels. Then again amd may also just have moved the sensors/changed how they expose them.
Do you know what power limits it is using?
Well the fan noise when ramped is probably not going to change but how often it’s doing it may be.
From btop, the power draw ranges from 4 to 18 Watts for garden variety productivity tasks (power-draw shoots up when initially loading a heavy website or multiple websites at the same time when restoring a browser session). At idle, where I have a bunch of browser tabs open and maybe a terminal or two open, I see power draw (as reported by btop) hovering around 4W and temperature sitting at 52C.
When I wake the system from sleep, the temp is usually b/w 37-41C. After using the system for a few minutes, it gradually ramps up to ~52C and stays there, never falling below 50C.
When not connected to a power outlet, power draw drops to 3W and idle temp sits at ~50C.
One thing I’d like to note is that, based on my experience, compared to the 7040 series, it takes a lot to push the AI 9 300 series beyond 60C. But the temp at idle is abysmal as it stands rn.
Let me know if there’s a better way of collecting and reporting power draw, or if you’d want me to run an experiment. I’m on Bluefin so totally open to installing tools and running tests in distrobox containers.
Can you check the powerlimits unsing ryzenadj or something? I am particularly interested in what they set P3T to on this one.
Thanks to my 256g FW flash card, I ran this experiment on the latest version of Fedora Workstation. I followed the instructions at GitHub - FlyGoat/RyzenAdj: Adjust power management settings for Ryzen APUs to:
- build & load
ryzen_smu - install new key using
mbok - build & install
ryzenadj
The module seems loaded:
$ lsmod | grep ryzen_smu
ryzen_smu 45056 0
But, there are errors in the log:
$ sudo dmesg | grep ryzen_smu
...
ryzen_smu: SMU v11.93.4.0
ryzen_smu: Unknown PM table version: 0x005D0009
ryzen_smu: Failed to probe the PM table -- disabling feature (249)
...
AI suggests that the Ryzen AI 9 HX 370 is using a newer SMU firmware (v11.93.4.0) with a Power Management (PM) table version 0x005D0009 that the current ryzen_smu module doesn’t recognize.
So, this test didn’t directly yield what we were looking for. But, I did note that the high-temp issue was not present on the latest version of Fedora which is sporting kernel 6.15.6-200.fc42.x86_64 vs. the 6.14.11-200.fc41.x86_64 on my Bluefin-dx GTS.
So, it seems that kernel 6.15 has the high idle temp issue more or less addressed.
Does ryzenadj -i or --dump-table still work? dump-table should work with the unknown table version I think cause it also worked on the 7040U with the new agesa before ryzenadj was updated.
That is good to see, the high temps were likely caused by excessive idle power which they may have improved in newer kernels.
No, it doesn’t work 
.
@Adrian_Joachim Works for me:
# ryzenadj -i
no compatible ryzen_smu kernel module found, fallback to /dev/mem
CPU Family: Strix Point
SMU BIOS Interface Version: 21
Version: v0.17.0
PM Table Version: 5d0009
| Name | Value | Parameter |
|---------------------|-----------|--------------------|
| STAPM LIMIT | 25.000 | stapm-limit |
| STAPM VALUE | 7.018 | |
| PPT LIMIT FAST | 30.000 | fast-limit |
| PPT VALUE FAST | 5.925 | |
| PPT LIMIT SLOW | 30.000 | slow-limit |
| PPT VALUE SLOW | 10.246 | |
| StapmTimeConst | nan | stapm-time |
| SlowPPTTimeConst | nan | slow-time |
| PPT LIMIT APU | nan | apu-slow-limit |
| PPT VALUE APU | nan | |
| TDC LIMIT VDD | nan | vrm-current |
| TDC VALUE VDD | nan | |
| TDC LIMIT SOC | nan | vrmsoc-current |
| TDC VALUE SOC | nan | |
| EDC LIMIT VDD | nan | vrmmax-current |
| EDC VALUE VDD | nan | |
| EDC LIMIT SOC | nan | vrmsocmax-current |
| EDC VALUE SOC | nan | |
| THM LIMIT CORE | nan | tctl-temp |
| THM VALUE CORE | nan | |
| STT LIMIT APU | nan | apu-skin-temp |
| STT VALUE APU | nan | |
| STT LIMIT dGPU | nan | dgpu-skin-temp |
| STT VALUE dGPU | nan | |
| CCLK Boost SETPOINT | nan | power-saving / |
| CCLK BUSY VALUE | nan | max-performance |
I did have a couple of hangs when altering power settings though.
@ehsanj you may have trusted boot enabled; linux detects that and disallows direct tampering with memory.
I wonder if repasting AI 300 with PTX helps with the noise. I’ve got a significant improvement after repasting my old 7040. I think I’ve read somewhere that FW now uses PTX everywhere, not just for 16” machines.
One data point I just accidentally discovered … Replacing the WD SN850X in my AI 300 with a WD SN7100 dropped my idle temps at least 8 degrees. The numbers indicate a bigger drop, but my eyes are kind of bugging out, so that’s all I’ll admit right now.
Part of that might be installing a brand new Bluefin? But, in theory, this new OS should be the same as the old one, down to the kernel stepping. I hope it isn’t THAT different. If you haven’t tried already, maybe try booting the Bluefin ISO over USB?
(The SN850X is going into the Desktop that just arrived.)
I have a 7840U and no experience with the AI series, but I would caution you to wait based upon my intuition and a few early reports on this forum. It’s not clear how well supported the TPU is by software, and current reports suggest that at best it’s “uneven”. I would love a local AI assistant, and I’m convinced that’s coming soon, but even Alpaca/OLLAMA aren’t working on a regular CPU for me right now. I’ve seen a number of posts detailing now AMD can’t figure out if ROCM is going to work with the TPU or how, and it seems increasingly likely that new chip generations will be out before some or all of those questions get answered.
That’s a bigger issue because the AI series is very much a sidegrade in terms of pure CPU performance due to the R5c and R5 cores, which are allotted as 8x R5c and 4x R5 cores. Most laptops aren’t terribly CPU bound as it is, but the ONE HUGE AREA WHERE THEY ARE IS LCOAL AI/LLM computation. It would be pretty disheartening to get a R9AI370 to replace your 7840/7640 only to find that the TPU doesn’t work while the mix of compact/full fat cores is no faster IN AI WORKLOADS than the chip you just replaced for $700-900.
I would very strongly urge folks to wait until you’ve got confirmation from whatever source you trust that your software stack and hardware will work with AMD’s TPU before splashing out. I’m as excited as anyone and I want to buy this, but I’ve gotten a BIT wiser in my decades of errors.
1 Like
Remember the good old days of delidding Core 2 Quads to install LiquidMetal? Good times.
A quick update that with the most recent Bluefin GTS update (Kernel 6.16.7-100.fc41.x86_64), the heat/noise issue with my HX 370 is fully resolved. Idle temps are b/w 38c and 42c and wattage hovers b/w 3w and 4w (values as reported by btop).
On the AI front, with ROCm 6.4.4+, AMD has expanded (preview) support for PyTorch on Ryzen APUs. I can fully utilize the iGPU for local AI workloads using LM Studio, getting 20-23 tok/sec (impressive!) using the openai/gpt-oss-20b model (this is through Vulkan as Bluefin GTS ships with ROCm 6.2).
Once full ROCm support lands, the perf should improve even more. At this point, I’m thoroughly impressed with the power & efficiency of the Ryzen AI HX 370 APU!
For more details, please review this comment (and follow through to the end for a boatload of valuable AI-related info with respect to the HX 370 APU).
3 Likes