Crashes during inference

[metairieman5158f02df]

Infinite RS AI 2NVZ9-1289US-B9285K5093264GS2TX11PHZ9 Desktop with GeForce RTX 5090 32G GAMING TRIO OC with 64GB RAM
According to all the specifications for this model, I purchased this system mainly for using AI inference mainly using ComfyUI for various image/video inferences on a hobbyist level. So this post mainly deals with some instability I'm experiencing with the computer suddenly shutting down during some of those sessions, a complete reboot of the system, so I'm trying to troubleshoot the reason. The system--for those not familiar--has the Liquid-cooled processor feature but only air-cooled RTX5090 if temperature has anything to do with it. Possibly 64GB RAM not enough?

I don't really suspect hardware failure because when it's working, it works well. And I don't not push the system with oversized models (40GB!) usually using quantized models <24GB and less and using GGUF models on a lot of things. I try to be conservative but the crash to boot happens more than I would think with such a power system.

If anyone has experienced this, suggestions of added air-cooling schemes for the RTX5090 or any other input you could offer would be much appreciated.

I forgot to mention, the system is running the latest/greatest Win11, too. Thanks in advanced for any suggestions!

 

[citay]

First, I would check if the BIOS version is the latest one (currently 1AC4), and if not, update to that, https://us.msi.com/Desktop/Infinite-RS-AI-2nd/support?sku_id=96327

HWinfo can show you the version, on the System Summary screen.

Then, you need to find out which component is involved in the issue. I can tell you right off the bat, the RAM won't be it. If the RAM gets full, the PC will just become slow, it will never shut down out of the blue. Shutting down happens either due to overheating or due to some power-related issue. Although you later speak of rebooting, which is a bit less severe than it suddenly shutting down, but still should never happen of course.

Download OCCT, it's a stress-testing tool where you can check each component individually. Do a CPU test, then RAM, then GPU, and finally you can do the "Power" test, which creates full load for the CPU and GPU at the same time (rarely happens in real life, usually it's either/or). If any of the tests can trigger the same issue, depending on which one, we can be one step closer to finding the cause.

 

[metairieman5158f02df]

My apologies for not responding. I didn't have the notification turned on and didn't know I had a response! Thank you for the suggestion of OCCT. I am running that now (15 mins in) so I will report back when I find something out. I did do a BIOS upgrade just the other day but a "1.ADO" (3/05/2026) rather than your linked "1.AC4." (11/2025). I should stick with that later one, correct? Appreciate it!

 

[Svet]

Topic moved to correct area.

I did do a BIOS upgrade just the other day but a "1.ADO" (3/05/2026) rather than your linked "1.AC4." (11/2025). I should stick with that later one, correct? Appreciate it!

Where you find this "1.ADO" (3/05/2026)"? It's not listed on web page:

MSI USA

Welcome to the MSI USA website. MSI designs and creates Mainboard, AIO, Graphics card, Notebook, Netbook, Tablet PC, Consumer electronics, Communication, Barebone, Server, industrial computing, Multimedia, Clean Machine and Car Infotainment.

us.msi.com

 

[metairieman5158f02df]

That is weird. I remember cutting and pasting the model number from my Member page and I know that was there! But you are correct . . . it's not the one showing now! I have 1AC4 so I'll do that now. Must've goofed somewhere along the process.

On the OCCT issue, I did use that Saturday but just for some of the lighter tests to get used to the program, about the 1 -1.5 hour tests. What I did notice more than anything else is temperature. Voltages and such seemed okay though I'm not an expert.

The idling temperatures seemed to hang around the 50°C (or above). FYI, I bought this system around 11/26 in the winter months and wasn't having this issue. I'm in New Orleans, South Louisiana and it seems the issue of 80-90°F temps (I do have A/C, though!) plus airflow out of the cabinet may be the issue. I noticed you have a Suprim (which I would have gone with) but this prepackaged system had the air-cooled so I just went with it.

In any event, step by step I'm going to install the CORRECT BIOS(!), Then I'll go back to the OCCT testing and see what happens.

Admin Svet says this was moved to the "correct area." Is this correct now or do I need to now post in the other area? Please advise.

Again, much appreciate the tips and wisdom!

 

[Svet]

That is weird. I remember cutting and pasting the model number from my Member page and I know that was there! But you are correct . . . it's not the one showing now! I have 1AC4 so I'll do that now. Must've goofed somewhere along the process.

Yes, best reflash to 1.AC4, 1AD0 could be pulled out due to some issues.

Admin Svet says this was moved to the "correct area." Is this correct now

Yes, topic been moved and it's now in correct area.

or do I need to now post in the other area? Please advise.

Nope. keep posting here.

 

[metairieman5158f02df]

I don't mean to be a PITA but I'm getting crossed signals. I wanted to get this right but only got more confused! I looked up a YouTube video on installing the BIOS and ran across "Mike's Github BIOS Report (https://github.com/Mikesunboxing/PreBIOSCheck/tree/bcf16ef6572dc932119821bcae556ab72899cc7c) which was produced through PowerShell (admin) which produced the attached report. It also provided a search for the BIOS. This ended me up on the 1AD0 page! Somehow this was also where I ended up as mentioned in the beginning. But that was before I knew about Mike's Report! That is the my "Motherboard BIOS" attachment copy'n'pasted page. And note the very bottom sequential listing before 1AD0 is for 1AC1, which would even be a predecessor to 1AD4. It doesn't show on that page.

I guess what I'm trying to ask is "Is there a difference for a specific motherboard BIOS which would differ from a complete MSI pre-buiilt system BIOS?" I really don't recall how I exactly found the 1AD0, i.e., looked up the motherboard BIOS instead of the complete system BIOS . . . maybe? So, which is the best BIOS I should go with? I'm on 1AD0 now which said something about overclocking improvement--which I don't overclock anyway--or go with the pre-built system BIOS 1AC4 as indicated in our first encounter?

Again, I appreciate the wisdom of those more expert than myself. I've always assumed BIOS versions improved on the last version i.e., 1AD better than 1AC . . . but what do I know?! And finding out now that the motherboard has different BIOS than the complete system . . . that's where I need enlightenment if you don't mind.

Thank you for your consideration!

 

[citay]

How did you originally update the BIOS? Because what happened here is exactly this:

I really don't recall how I exactly found the 1AD0, i.e., looked up the motherboard BIOS instead of the complete system BIOS . . . maybe?

If you're on BIOS v1AD0, as the tool says, then you're on the retail board's latest BIOS version: https://www.msi.com/Motherboard/PRO-Z890-P-WIFI/support

For the pre-built PC as a whole, the latest version is the v1AC4 instead: https://us.msi.com/Desktop/Infinite-RS-AI-2nd/support?sku_id=96327

Now, usually, the existing BIOS version determines which one you can update to, because the filenames are different, and M-Flash (the updater in the BIOS) will only show filenames that match the already installed BIOS. All characters before the dot, and the first one after, have to match, otherwise the update file isn't shown.

Now let's look at the names as they appear in the BIOS:

Pre-built PC: E7E34IM90.1AC4
Retail board: E7E34IMSI.1AC1, E7E34IMSI.1AD0

With these differences, it should be impossible to "cross-flash" between the v1AC4 and the v1AD0 versions. You should've only been able to update to the v1AD0 version if your previous version was also some E7E34IMSI.1xxx. Only when you used the Forum Flash Tool, then it's possible to cross-flash.

So if you used M-Flash from within the BIOS, then for some reason your board already had a retail board BIOS version on there, not the pre-built BIOS version you'd normally expect.

Now, if the board model in this PC is really the PRO Z890-P WIFI (check the writing on the board), then you can stay on the v1AD0 version obviously, since it's the latest stable version for this board, and it has not been pulled or anything, it's legit. Then I just wonder why you have this version on there and not the version usually intended for the pre-built PC.

 

[metairieman5158f02df]

Hmmm. Interesting! Here I go ASSUMING again! Just for reference, I purchased this rig directly from MSI Store not some third-party vendor so I would suspect the pieces/parts are exactly as reported on the MSI website, thus the PRO Z890-P WIFI is legit. Everything was pristine and every reporting program I've run like HWINFO, CPU-Z, OCCT, etc., reports that what it is. So . . .

I'll reiterate from earlier. I DID flash the BIOS myself to the 1AD0 because I was checking updates on the MSI site. So of that I'm "guilty." I mentioned that early on. Maybe I shouldn't have but I did. In any event, your mention of the "system BIOS" "1AC4" versus the "MB BIOS" "1AD0" has me wondering if the following procedure would be worth the effort . . . starting back from scratch because I don't remember what the original BIOS was. The exert from the MSI System manual:

"Resetting BIOS●If you encounter certain issues with your computer, restoring the default BIOS settings might help. You can reset the BIOS settings using the following methods: Enter the BIOS setup and press F6 to load the optimized defaults.●●Use the Clear CMOS jumper on the motherboard to reset the BIOS.If your motherboard has a Clear CMOS button on the rear I/O panel, press it to reset the BIOS. ⚠ImportantEnsure the computer is powered off before clearing the CMOS data. For more details, refer to the Clear CMOS jumper/button section in the manual."

What would you suggest? I don't have a clue about "cross-flashing" prohibitions so IF MAYBE your 1AC4 BIOS was a special "system tuning" for the system with the RTX5090, that might be better for the system as designed. Of course, I have no clue what effect that BIOS would produce but the fact that it exists with no updates showing from MSI Center since November would reverting to 1AC4 be the better choice? Since you have a better understanding, I'd defer to your advice.

My first post about this system was really about the system crashing ONLY during AI inference sessions (ComfyUI mainly) and there was/is no system crashing with normal usage like browsing, office progs, email and other non-intensive graphics, just with heavier RTX5090 functions. I previously had a much smaller AMD system with the RTX4060 Ti which never crashed. I simply had a financial windfall fall of last year and decided to get the best system I could afford to play around with AI a little more efficiently, bigger models, video . . . keep up with the trends of the day. That's what my goal is in this thread to get things working the best they can. Nothing more. And the bugaboo lately is I never use the BIGGEST models. I tend to the "lower VRAM" models so I DON'T tax the system resources. But . . .

Thanks again for your quick replies!

 

[citay]

Well, the question is, did you flash the BIOS using M-Flash in the BIOS? Then, if you now have E7E34IMSI.1AD0 on there (as shown in the BIOS on the top right), your board already had a retail board's BIOS on there before.

 

[RemusM]

My first post about this system was really about the system crashing ONLY during AI inference sessions (ComfyUI mainly) and there was/is no system crashing with normal usage like browsing, office progs, email and other non-intensive graphics, just with heavier RTX5090 functions. I previously had a much smaller AMD system with the RTX4060 Ti which never crashed.

I'm not saying that it's your case for sure, but these days the market is full of buggy pieces of software.
Now talking about these "AI inference" sessions, the results heavily depend on your app, drivers and system.
In many cases the authors of that specific app leave some "interesting" notes, such as:

You must install the app to the main C:\ drive. Installing it to a secondary drive will cause instability.

That's a sign how immature their app is.

P.S.
GPU processing was always a challenge (from any point of view).

 

[metairieman5158f02df]

Well, the question is, did you flash the BIOS using M-Flash in the BIOS? Then, if you now have E7E34IMSI.1AD0 on there (as shown in the BIOS on the top right), your board already had a retail board's BIOS on there before.

I did use M-Flash. So would there be any advantage trying the system version you presented (1AC4) or just leave well enough alone?

 

[metairieman5158f02df]

I'm not saying that it's your case for sure, but these days the market is full of buggy pieces of software.
Now talking about these "AI inference" sessions, the results heavily depend on your app, drivers and system.
In many cases the authors of that specific app leave some "interesting" notes, such as:

That's a sign how immature their app is.

P.S.
GPU processing was always a challenge (from any point of view).

I know what you mean! I've been playing with ComfyUI for a couple of years now and their node system depends on a node's developer keeping it current. So CUI is the base product which uses other peoples' concocted nodes. So Win11 generally plays along well with open source CUI. BUT . . . I use a "portable" version of CUI which basically means the Python engine is contained within the Comfy system along with everything else dependent from it. Windows can have one Python version and CUI works with its own! Basically, I could take the drive out of this computer, move it to another and everything would work (as long as the VRAM is compatible, that is). I have CUI isolated on a separate drive, the models on another drive all not dependent of the OS C: It's been working for a couple of years that way. I like to leave my OS and AI separate in case I have to reinstall Windows or CUI & her models and not have to reinstall BOTH.

Thanks for your input, RemusM!

 

[citay]

I did use M-Flash. So would there be any advantage trying the system version you presented (1AC4) or just leave well enough alone?

If you have the E7E34IMSI.1AD0 on there now, and you used M-Flash, then you cannot flash the E7E34IM90.1AC4 of the pre-built system (or any future updates for it), you can only update to the newer versions of the retail board. Unless you jump through some hoops with the Forum Flash Tool to cross-flash back to the pre-built BIOS versions.

But honestly, most people with pre-built PCs want to achieve exactly what you have now: Get away from the pre-built PC versions, because they are much less frequently updated and often lag behind a lot, and get on the retail board's BIOS versions, to get the latest fixes much sooner. So, I don't know why you have the retail board's BIOS version now, but it's not a negative. Yes, the pre-built PC BIOS can have some other defaults here and there, but there's not much benefit of that, the defaults are usually not ideal anyway.

Your BIOS is now almost half a year ahead of the pre-built PC's versions, and it will only continue like this. Sometimes the retail board's BIOS versions can be years ahead, because they just stopped releasing new versions for the pre-built PCs.

 

[metairieman5158f02df]

Well, great! That is refreshing to hear! I did just what "suspected" (LOL!) before I opened my email tonight and saw your reply. [I prayed before] I used the Flash BIOS Button with 1AC4 twice and it stuck on 1AD0! So here I stand. I'll follow the MB updates instead. I appreciate your guided patience with that part!

Since my setup has been for years in a rather enclosed "hutch" type wooden computer desk, noticing the temperatures around 50°C idle, I decided it was time relocate the box in a more open, free flowing air area. It's been jammed in between with my Truenas box so the air flow has not been what it should which is probably affecting my original post before the BIOS took the lead. The system has the water-cooled CPU system which keeps that tamed but only air cooling for the GPU. More fan(s) maybe now? I've been researching . . .

Thanks for all the insights. I've learned more about BIOS now! I have another issue I'm curious about so I'll start another thread about the "Silent<>Gaming" switch. If you've got insights on that, please comment.

 

[Svet]

I guess what I'm trying to ask is "Is there a difference for a specific motherboard BIOS which would differ from a complete MSI pre-buiilt system BIOS?"

Depends, but for your case with those units yes it is aside from the cosmetics from the OEM.
Normally they have tweaked different power limits and other fine tuned things.

Infinite RS AI 2NVZ9-1289US-B9285K5093264GS2TX11PHZ9

Can you check mainboard S/N too? (you can skip the last 3 digits from it or PM me the S/N)

Because if you really have:

Infinite RS AI 2NVZ9-1289US-B9285K5093264GS2TX11PHZ9

Then your current E7E34IMSI.1AD0 is the wrong BIOS,
which can explain the shutdown issues.

 

[metairieman5158f02df]

System s/n: MSBZ01P4A0107703 If you're asking for the board s/n, I'll have to get into the case for that. I was planning on moving the case to a more ventilated area tomorrow to see what effect that has so I can get the serial for that tomorrow. Thanks.

 

[Svet]

Yes, board S/N, not system S/N.

 

[metairieman5158f02df]

Yes, board S/N, not system S/N.

s/n: 601-7E34-010B2408023787. For clarification, that's the tag on the main board power connector if that's correct.

Okay, so after relocating the computer to a less constricted area (this box weighs 50lbs!), I checked temps with HWINFO and here's the screenshot. I wasn't expecting GPU fans at 0 and took a look and they were not spinning at all. I would think at 130°F they would be. I went into MSI Center and switched on "Cooler Boost" and they did kick on for the 100% for 30 secs. The Zero Frozr is clicked on and shows On: Tranquility Off: Cooling Enhanced Mode. All of this was done virtually at idle with no load other than this email. You didn't request this info but since I had it . . .

Never mind that paragraph! I had the switch understanding bassackwards. I found a video explaining it on MSI Youtube about ZeroFrozr. I had it switched on and that may have been the problem all along. They are spinning now and the temps (at idle) are in the 90's. But if you do get something out of the serial number. Let me know.
Thanks!

 

[citay]

Can you set the temperatures to Celsius, perhaps? In idle, the fans are not meant to spin, since a 5090 is totally ok at 40-50°C in idle, see here for example, https://www.techpowerup.com/review/msi-geforce-rtx-5090-suprim/39.html

All the 5090 cards have a 0dB-zero-fan-mode in idle, the fans only come on when the temperature exceeds a certain threshold, which only happens when the GPU is under some load. You want the fans to stay off in idle to have no noise and no dust-build up, and it's fine, the GPU doesn't particularly care if it's at 30°C or 50°C.

You do have a rather high idle power draw even for a 5090, normally it should be around "just" 25-30W (still quite high for a GPU), not 45-50W like on your screenshot. So some power-saving mechanisms are not working to the full degree here, which might go some ways to explaining higher idle temperatures.

Make sure the power plan is set to "Balanced" (the only proper one), then you can try setting PCIe Link State Management to Maximum Power Savings in the Balanced power plan: