Running a 7900XTX on a 750W PSU
Hello everyone,
I'm looking to upgrade my current setup with a 7900XTX and the manufacturer website recommends 850W at a minimum.
As the title says, I currently only possess a 750W PSU (Corsair RM750x (2021)).
Rest of the system (the parts that draw power at least):
Mainboard: Asus TUF GAMING X570-PLUS
CPU: AMD Ryzen 9 3900X
RAM: 4x8GB DDR4
Storage: 1x Samsung 990 Pro 2 TB, 1x Samsung 970 Evo 500 GB, 2x older Samsung SATA SSDs
Fans: 3x Case fans, 1x Noctua CPU cooler (forgot the exact name)
PSU: Corsair RM750x (2021)
Putting everything into PCPartpicker makes it spit out a max load of 608W, but considering that the 3900X can spike to draw up to ~145W (instead of the 105W TDP) and the 7900XTX can spike up to ~530W (instead of the 355W PCPartpicker assumes) I'd have peak loads just short of 850W...
My question now is twofold:
- How reliable is the Power Usage Limit I set in LACT enforced? If I can rely on it to keep the GPU to 355W I should (in theory) be fine, right?
- How bad is it to trigger the overdraw protection in the PSU? Obviously my PC would shut down immediately (with all that entails), but would I risk damage to components?
Thanks for any help :)
PS: I'm set on the 7900XTX, as I want the 24GB VRAM. It's quite literally the cheapest option (new, in my region) that is also a usable gaming card.
PPS: Yes, I'm going to upgrade the PSU soon. Just not now if I can avoid it. (They aren't exactly free...)
I suppose theoretically low voltage could damage something, but I doubt that it's very likely.
I'd probably favor getting the PSU, because otherwise, if you hit any stability issue, you're going to be thinking "is it the PSU"? But I wouldn't be too worried about component damage.
Note that if you're wanting local AI compute capacity --- the likely reason to want a lot of VRAM attached to a GPU other than gaming --- and you can live with less memory bandwidth and a substantial price bump isn't an issue (which it sounds like it is, but just throwing this out there), you might also consider a unified memory computer, like a Framework Desktop, as that can take a lot more memory than a consumer GPU. I have a 24 GB XT 7900 XTX and a 128 GB Framework Desktop, and I've found myself using the latter rather than the former, especially for LLM stuff.
If you're willing to wait until 2028, my expectation is also that memory prices will be better and thus bang-for-buck on local AI compute will also be better. Also give the hardware guys time to iterate on hardware; probably improve performance.
I was looking into a Framework Desktop, but even adding a 150€ PSU on top of the 7900XTX right now (before I have resold my old GPU) would be stretching things. It's definitely on my wishlist, but not realistic with my current income.
And I'm definitely too impatient to wait until 2028, especially since it isn't a guarantee...
Thanks for your input though :) appreciate it.
Yeah, I don't know if I'd get it myself with where RAM and NVMe prices are presently. But that sort of thing is presently about the cheapest way to get a ton of memory connected to a GPU.
Yeah, I know. And prices aren't likely gonna promptly crash to the money-losing 2024/2025 rates, either, since there's pent-up demand. But, just saying that my guess is that there's going to be substantial near-term depreciation on any AI hardware purchased right now.
Sure. I hope you do enjoy it if you get it (and it's certainly been more-than fine for gaming, in my experience).
I don't know about LACT cause i haven't used it, but if you have a good quality psu (and it looks like you have) it should be fine. I was in a similar situation when i went from a vega 64 to a 6950xt. I had an RM650x but 750w was recommened for the 6950xt. I just tried using it without upgrading the psu and for gaming it seemed to work fine. It wasn't until i started messing around with genAI that the psu would turn off my system. I'm assuming the AI stuff hit a power spike that gaming workloads did not. I eventually upgraded to a 1000w psu, but i was able to get by for a bit with the 650w.
You should be ok, even if you touch 750W every now and then. Corsair PSUs are generally decent and can likely hold over 750W for short periods. What you are sacrificing is PSU efficiency, but if it's temporary, whatevers. Don't overclock the CPU and under volt the GPU and don't run heavy and continuous compute loads for long periods of time.
While the power management is likely a hair different with my flavor of 7900XTX, it has always behaved well under heavy load and benchmarking and it rarely spikes to max power unless all the OC knobs are pushed to max.
Unfortunately, I can't speak to how the 7900XTX is managed under Linux. If I am not mistaken, unless you are purposely overriding power profiles, you should be fine with whatever the stock settings are on the card itself. (7900XTX's do come in non-OC and OC variants or have a switch for a dual bios for either configuration. Manufacturer OC settings can still be quite timid, IMHO.)
Depending on your specific components, the system may crash before it hits a power limit in some cases. If the PSU can't handle the power draw requirements, the power rails might sag a little during a burst and crash the system. I have only seen this a couple of times, but it's still worth mentioning.
Ignoring all of the above, just run some tests under different power loads, and in this case, aim for the lowest power settings you can, just to be on the safe side. While not perfect, invest $20 in a kill-a-watt power draw monitor for system testing. At a minimum, you will get a rough idea of total system power draw, which is good enough here. (I have seen ~+/-5% skew between different kill-a-watts, which isn't an issue unless you are seriously working against a strict draw limit.)
I'm not really interested in shortening the lifespan of my hardware with overclocking, especially considering current prices for replacements, so I'm safe on that front :)
I'll look into undervolting once the card actually gets here and I can test stuff. Sadly the two variants that are actually available for a sane price here (reference design cards are over 40% more expensive) seem to be manufacturer OCd, but afaik I should easily be able to underclock as well using the tools available, in case it becomes necessary.
Thanks for your input :) I'll definitely also look into a kill-a-watt (or whatever similar tool is available here), I was interested in the actual power consumption of my system anyways.
I also forgot to answer your actual question! :)
Trigging the overdraw protection on a PSU is generally not a horrible thing, but it really depends on how the PSU does it. I would generally assume that that Corsair uses a proper current monitoring circuit with my best guess being a resisor shunt of some kind in combination with a proper monitoring chip that triggers a solid state switch to cut power extremely fast. Cheapers PSUs may just use a resettable fuse, which is also "solid state" but may actually fully blow (ie: esplode) or become less effective over time. Absolute worst case, a capacitor may blow its top. (If you have an old PSU you can destroy, flip the switch to 115V and plug it into a 220V circuit. You will get a loud bang and some smoke, but generally, nothing extremely dangerous. Its worth it to see and understand how PSUs may fail catastrophically if you never have blown up a large capacitor.)
The biggest warning sign of overloading a PSU is heat. PSUs get warm naturally and will get hot under heavy load. If the PSU has a fan, just occasionally feel the air and if it's getting uncomfortably hot, you might be getting in the danger zone. Heat will naturally change resistive properties of different materials, leading to variations in current flow. This is normal. However, heat can trigger "thermal runaway" conditions where a component gets hot, starts to draw more current, gets hotter, draws even more current and will eventually melt or explode. Not good, but also normal. Hence: Heat is a good indicator of a device that is about to fail or shut off. (Current overdraw circuits may also use thermal sensors to shut off the device as well.)
If the device gets overly warm, point a fan at it as a temporary solution. It'll give you just a hair more wiggle room during overdraw conditions. Not much wiggle room, but wiggle room nonetheless.
In short: High current loads/pushing components over rated limits is never ideal, but it's not extremely bad either. It shortens the total life of the device itself and its safety circuitry, but it's ok for short periods if you aren't stupid about it.
(Sorry for the TED talk, but it's my way of walking through the different scenarios of what you are trying to accomplish. FWIW, I have a ton of experience blowing up electronics so I am probably more familiar with pushing safety limits than most casual users.)
Bonus: Overclocking/Overvolting is not always going to shorten the life of equipment. Damage is caused by poor and ineffective cooling, but that does take proper planning and testing. Factory overclocked devices are usually still waaaay under their actual (and usually undocumented) thermal/voltage limits. However, you are usually getting really beefy heatsinks and additional fans for those kinds of cards. If you run a "factory overclocked" card at normal load, its additional cooling will theoretically extend the life of the card beyond its actual useful lifespan since it it's being run far below it's rated thermal limits. (This paragraph sounds like a contradiction in itself, but makes sense when taking the total useful life of the card into account, before it actually is completely outdated and is only good for scrap.)
I may be able to provide some personal experience regarding the first part of the first question, "How reliable is the Power Usage Limit I set in LACT enforced?". The answer appears to be application based and I will do my best to explain.
I was having a persistent issue with my GPU PPT0 throttling during gaming sessions with a specific area in a specific game. This was the first time I had ever experienced this odd behavior which initially presented as a rapid increase in GPU fan speeds and a small loss in frame-rate. By using LACT, I was able to confirm why this was happening and the real-time data was indicating that as soon as the GPU got within about 10% of the power cap, it would start to power throttle. The crazy thing was, the 170 Watt cap would not stop the card from drawing over 190W in short bursts. By increasing the power cap to 190W, it did reduce the throttling somewhat but then my GPU was still exceeding this revised cap and I didn't see any real value in pushing the value higher.
I have a decent safety margin with a 9600X and a 9060XT using a Corsair 750e PSU. I imagine your setup has much higher peak power needs so I would suggest reducing the power cap by 10% to start and then incrementally adjusting the core voltage downwards using the undervolt option. This should allow you to achieve a stable operating level if you are careful and find the sweet spot between performance and power draw. Otherwise, getting a beefier PSU may be your safest bet.