AMD 7900XTX Overclocking can lead to LESS performance... Here's how
Вставка
- Опубліковано 25 сер 2024
- Learn more about the Canvas lineup of Monitors from NZXT at nzxt.co/Canvas...
Get your JayzTwoCents Merch Here! - www.jayztwocen...
○○○○○○ Items featured in this video available at Amazon ○○○○○○
► Amazon US - bit.ly/1meybOF
► Amazon UK - amzn.to/Zx813L
► Amazon Canada - amzn.to/1tl6vc6
••• Follow me on your favorite Social Media! •••
Facebook: / jayztwocents
Twitter: / jayztwocents
Instagram: / jayztwocents
SUBSCRIBE! bit.ly/sub2Jayz...
This is not something I have ever experienced with any other AMD gpu series, nor is it something I am experiencing on custom cards... after trying multiple drivers, and unable to find anyone else talking about this, what are your thoughts as to why this happening? I will follow up when we hear from AMD.
I'd imagine it's because the reference cards are very strictly power limited.
The VRAM slowing down is a stability feature, not an issue
If the reference card had a third 8 pin connector, I believe it would've behaved like the Red Devil card to a certain extent.
Isn't this that thing they mentioned that the shader engine is decoupled and will operate with front end frequencies at a higher clock speeds while the back end of the shader engine drops in clock speeds if it's not utilized because of that new power efficiency and power management tuning?
Obligatory Nvidia shill comment
Overclocking RDNA 3 is very different. Messing with the Minimum Clock Speed setting in any way can lower performance and causes all kinds of weird issues. Start with increasing Power Limit to +15 and then lower voltage until you get crashes. With the higher power limit as you lower voltage you will see your actual max clock speeds increase automatically. Once you find the lowest stable voltage in games (short benchmark runs will pass with much lower voltages than actual gaming will allow) you can try increasing the Max Clock Speed setting a little at a time and you might get higher clocks. Most people are finding that memory overclocks top out at around 2750 before performance regression starts. Power Limit can be lowered for lower thermals if needed but that's it for now with RDNA 3. Any other approach right now will most likely cause issues and/or not be stable.
It is exactly the same as with overclocking my RX5700XT. RDNA 1
Thank you, I was just about to post this. Jay is just bad at this.. this was Verge builds a PC tier OC.
The problem is probably the 2x8pin connectors that don't allow enough power for OC. 3x8pin custom cards should allow higher clock speeds, and then we will see the true potential of RDNA 3. I think AMD limited it with the limited power of a 2x8pin...
@@lucidnonsense942 yeah, no, it’s pretty obvious that Jay is a little out of his element when it comes down to AMD GPU OC’ing.
@Lucid Nonsense he's speaking for anyone use to nvidia that is considering switching. It's just another reason AMD isn't taking market share when it comes to enthusiasts
Jay my man, you're trying to OC like its an Nvidia card. You need to undervolt the card to create headroom. undervolt until it crashes, then increase the voltage a bit until its stable, then start messing with the clockspeeds and what not.
Undervolting is almost a must with these cards. My 5, 6 and 7 series cards all loved it. Core speeds haven't had much impact unless they were undervolted. Core min/max sliders on my 7900XT needs to be at least 4-500mhz apart or there's a performance hit. My 6900XT liked them to be as close as possible, (100mhz). My 7900xt likes 1025-1050 for voltage and at least 12% on the power tuning.
Nailed it!
That was the case for Vega cards too, it needed to be undervolted
@@marinmarinhola I will say that my ref 7900xt is a hell of lot quieter than my ref 6900xt. The cooler looks similar, if not the same, but it just doesn't run as hot. I'm limited by my PSU at the moment. 13900K and 7900XT will overload my 750watt corsair fairly easily so I can't really experiment too much.
You are way too bias towards Nvidia. AMD adrenalin software is miles better than Nvidia. BTW, you just need to get familiar with the software and how AMD architecture works. I have owned a 5700xt, 6800xt and 7900xtx and they are easy to OC and UV. Also, I was able to extract between 15-20% increase in performance, easily outperforming the 4080 in raster, and sometimes matching the 4090. I have never experienced any of the issues you show in your video. AMD is the underdog against the evil pricing of Nvidia, and should be supported and commended for their efforts in pricing more competitive than team green. Your video is misleading as you don't know how AMD OC works and just helps Nvidia getting away with their greedy prices
Do you mind helping me with overclocking my 7700xtx merc 310 and ryzen 7 7800x3d
@@Horus1133why would he help you? help yourself and search up how too
No driver issues with the 7900XTX? Thanks.
@@davidsmith4186from what I’ve researched that has been fixed since last year. I’m also considering to move to an xtx but idk. I need more evidence
My xtx is worse after overclocking and idk what to do anyone here have advice or maybe a thread to link to?
Missing the undervolting... the most important part of tweaking a 5000+ AMD GPU
Exactly, he said he doesn't really like overclocking AMD GPUs & it shows cause he doesn't know how to do it
@@chrisz5z This video made me want to punch myself in the dick. Plus, as far as undervolting, adrenaline is wayyyyy easier to use than afterburner.
I haven't had any success undervolting mine. I mean sure the power draw and heat both go down, but the performance always takes some hit as well. Jay has said in the past that performance is his most valued metric for GPUs. With that preference, I don't blame him for ignoring undervolting.
Vega gpus also (56-64-7) huge perfomance loss , all big youtbers missed the chance tweaking the card and showing the full potential
Correct, it would take me 10 seconds to add a 15-20% OC to that card.
Considering that this is a new approach to how GPUs are made, there are different approaches you also need to take as well including core voltage, which i didnt see you adjust at all and left it high.
I love that the community corrected his method, I love my sapphire nitro 7900xtx, i undervolted that behemoth, and she roared. Annihilates anything I throw at it, outside RT.
The size of that monitor and the distance it is from your face is like a monty python sketch xD
I had good results undervolting the gcd cores -60mv the extra thermal and power budget allowed the boost algorithm to boost ~200 mhz higher at stock wattage and ~400mhz higher at +15% power. As for memory it didn’t seem to scale well. Had a 10% fps increase in warzone 2.0. I have the XT which may be different though
I never seen Jay undervolt. The AMD cards since Polaris rx 480 and Vega have responded well to undervolt. He might of undervolted a 3080 or 3090 but never has an AMD card.
I believe it's those 2 8 pins on the XTX modal. Those cards really needed at least a 2x8+6 pins or just 3x 8 pins I truly do believe
I have 68XT and I use it for SimRacing Titles only. I use it with Power Minus 20%, drawing 200W, silent operation, and on ultrawide I don't lose much performance, maybe 3-5%. But as I said, I use it only for iRacing, ACC, AMS2. So my FPS is well above 100. RDNA2 is strange, and I know I would benefit more with UV but I'm lazy.
What resolution did you hit the 10% in warzone?
@@kwakes212 1440p, in the benchmark. Also had +50mhz fast timing on memory.
This is just the wrong approach for overclocking AMD cards, you have to raise the power limit and undervolt them first and usually the clock speed will increase on its own
You need to power unlock rdna GPUs to get any real performance boosts. Jay should know this. I can get get close to 20% performance boosts on 6900 XT when power unlocked with more power tool on water. Stop perpetuating AMD can't overcloxk myths...
This is why you're supposed to undervolt...
You're hitting the power limit of the card. There is a way to disable this power limit in the Windows registry though. I've done it with a Vega56 and Gamer's Nexus has a video on it. But if you don't want to registry tweak then undervolt in order to stay within that power limit. We've known this since R9 Fury/Vega but it really became a big thing with the RX 5700 XT.
By undervolting I have a 6800 XT running 2.8GHz on water. Real easy to do.
I have a 6800XT running 2.8hz on air lmfao 😂 IDK what the deal is with undervolting lately but it’s not needed and is pretty much pointless. Saving 50 watts and losing FPS? Just buy a lower tier card lol.
@@LegionGamingTV undervolting shouldn't lose fps unless you're touching the core clock.
@@LegionGamingTV far from pointless, RDNA cards have been shown to run at higher voltage than needed and perform better undervolted
I have stock 6800xts running 2800mhz out of the box.
@@LegionGamingTV you're confusing undervolting and underclocking.
Undervolting is a form of overclocking. Maximizing clock speeds via power and thermal regulation.
What an awesome day it is to have access to all this data and quality content. Sure beats the old days of a magazine how to and a prayer that it works!
Sure, but it seems Jay doesn't know how to overclock AMD GPUs...
@@TheEchelon you mean under clock. Let's be real here. OC on AMD is ghey.
I think they engineered the reference for efficiency because they didn't wanna look bad claiming they were more power efficient than nvidia and proceed to drop a card that matches power consumption. Sticking to 2-8pin connectors is in my opinion what knee capped it. The board shouldn't be allocating power to other places unless the total board power isnt enough. The issue seemed to be fixed once you switched to the 3-8pin red devil.
I mean the 7900XTX is still less power efficient than 4090 if you use sane power limits. The 4090 just has a ridiculous power limit to achieve maximum performance, but you if you cap yourself at like 270W you still retain 90% of the performance. 90% performance of a 4090 is plenty more than the 7900XTX. It seems that the reason RDNA2 was more efficient than Ampere was because the Samsung 8nm node is just that much worse than TSMC N7 (I think I read somewhere that Samsung 8nm is worse than Intel 10nm). Honestly it just seems that most of these performance uplifts come from the node upgrade rather than the GPU makers.
@@lycanthoss That's now an unfair comparison cause you could also power limit or undervolt the 7900 XTX to make it more efficient. You would have to compare one undervolted card against another to get a fair comparison.
@@harryhall4001 I think what he means is if both cards were locked to the same power limit, 270w in this case, the 4090 would be more efficient at said power limit
I think that AMD just trying to weasel through the claims that 7000 cards more efficient than nvidia.
@@KurniawanRamadhan213 it's definitely far from being as effecient as Nvidia when power normalized, AMD is slipping in the gpu market and Lisa su knows it, just watch her body language while presenting these cards and previous cards. She's not happy with it
I know this video was already long enough for most folks, but I really wish he would have tried a shunt mod, to see if increased power remedies the issue.
I was able to get a reference XTX for my nephew for Christmas, and this really makes me wanna put it in my system, see if the behavior is the same as this, and then shunt mod it to see if that fixes it.
A shunt mod won't fix the issue of 2x 8-pins limiting stable power from the PSU. Also 99% of users won't be soldering shit to their $1k GPU so that really shouldn't even be considered when buying it.
@@empireOfLove2 No I wasn’t trying to insinuate that was a “fix”. I’m just curious if giving it more power via a shunt mod would cause the ram to down clock less. I’m curious if it’s really like it seems that it’s stealing power from the VRAM, or if it’s just some kind of bug where added power wouldn’t make any difference.
On my 6900 xt I had zero luck with Clock Overclocks, the card just got hotter and gave no gains and even some lost performance. Undervolting, Ram OC & Timings, AND Fan Curve adjustment got me nearly 11% more performance out of the card. PowerColor Red Devil 6900 xt - 2350 min clock, 2450 max clock, 60Mv reduction, 2100 Ram with Fast Timings, no adjustment to power slider left it at 0, Fan curve Temps 30, 40, 45, 50, 55 and Fan Speed 35, 45, 50, 55, 60. Stock settings 19100 Timespy GPU score to 21600 which was roughly 11% performance increase. Core speed for RDNA cards is not now you OC or tweek the cards.
I have the same card; will have to try this later thank you
I got about a 2000 point increase in timespy on my 6800xt as well by running it at 1025 mv at 2650 mhz with vram at 2112mhz and fast timings enabled. The fact jay didnt undervolt is kinda sad, thats like the main portion of overclocking these cards.
My xfx Merc 6900xt in heavy games have around 2550mhz@1071v. In Uncharted or plague tale 2710mhz@1082v. finding the optimal settings is difficult but the effects are over 10% vs. stock
@@rENEGADE666JEDI Also got the merc 6900xt, would you say these are the most optimal OC settings you have found so far?
This is almost the same settings that I have but I had my memory set to default. I found fast timing for me made it unstable. 2400-2550 clock, 2150 mem, 1130mV
I think the bottleneck may come to the frequency between the connection of GCD and MCDs. Maybe something like AM4 platform, it will have the best performance when the IF and RAM speed become sync.
They have decoupled core and memory speeds. Could be the Infinity fabric is still a bit raw. Would be bad if AMD just never fixed this with drivers/firmware. Would not imagine it to be a very difficult fix, just to get things stable. And then I will give credence to the "age like fine wine" buzz phrase. But only then. But if AMD does a capitalism and just spins new silicon for new features... then competition is truly lost. No more Nvidia, AMD or Intel... its all a big cartel then
@@priitmolder6475 I think front end frequency and shader core frequency not actually means memory and core speed if I am right. If it is the hardware limit of infinity fabric, I think there is no way to fix it by software. Anyway, we need a official reply on this weird behaviour of the GPU to know what actually happened. I hope it can be fixed also.
@@MJ-nn6cr 7950? 7970? 7990?`Coming full circle names and upgraded hardware to compete with 4090? No air cooled reference design, but a dedicated AIO like the Fury was? To be honest, that would be a neat differential. 240mm AIO would most certainly show 4090 owners and the insane huge cooler sizes...
@@MJ-nn6cr depends on if it’s hardware limitation or firmware limitation
Amd overclocking is always a mess for the first few months. It has been this way since Vega. Be careful with the minimum clock tho. Just try to get max as high as possible, just like an nvidia offset.
With RNDA2, there is morepowertool to increase power limit as high as you want to. I got my 6900xt to pull 600w in furmark, but it always settled between 300 and 350 in games. There will probably be a version for RDNA3 as well.
no one ever uses auto oc hes just doing this on purpose to discredit AMD and make them look bad. He almost always uses Afterburner but for some reason on THIS video hes using the AMD Software Auto OC???? LOL CMON..Hes shilling HARRRRRRRRRRRD for NVIDIA right now. Its so clear. How can you not see that?
@@The_Gent did you take your meds today
Don't think he's shilling for nvidia, more of a tech tuber noticing irregular behavior and presenting it to the public. This was an out of the box approach and using Adrenaline, it's not a terrible program. Yes, there's some "why did they do that?" Quirks, but it displays a ton of metrics that help convey close to real-time, more-so than afterburner or gpu tweak 3.
@@s1n1stersixsgaming8 It's not irregular behaviour. The ref card is capped at 355W and you cannot increase the power limit as the hardware (2x 8pin, PEG) only allows for 366W on the 12V rail.
Increasing the frequency of the GCD requires more power and with a capped power limit there's less power remaining for other components.
@@johnscaramis2515 A GPU designed around optimizing memory bandwidth and cach that tanks VRAM performance as the GPU clock increases is anything but regular behavior. It's not a "hardware limit" it's a power limit set by AMD and he literally showed it topping out at ~395W with the limit maxed out. The Red Devil appears to be hitting that same limit only the behavior doesn't happen.
This is interesting. I leads me to believe that the card is able to quickly predict and adjust to the potential crash. So given a wattage, you could optimize some statistic (fps, the standard deviation of fps, 17% minimum fps) and feed it back into the software and it should be able to find whatever optimizes your criteria; for your particular game/application at your particular settings, all while not crashing. I'd love to see an overclockers database that take a much broader look at hardware and what you can do with it. Are there settings that I can use on a 3 year old machine that I can turn off vsync and get less input lag on Metro. Instead of maxing frames, minimize input lag. Etc.
I have a Sapphire XTX, but I am crashing so I will have to get a new PSU before I chime in about the XTX performance. These things badly need a driver update and it's clear they have been released too early. Launching any game is glitchy and stutters. The 4k performance is still pretty solid though. Superresolution veryyy buggy for me. I can't say I would recommend this card to a friend as it stands, definitely needs some polish. 100% agree about the lack of simplicity of overclocking, although I am just running at default values.
Edit: wanted to add that I have a Seasonic platinum 750w psu and 800w is recommended 🤪
A 750 seasonic titanium should run xtx easily, if not i need a bloody new one grrrr, when me card gets here will find out
I think my RX480 did hit lower vram speeds when hitting the power limit… but it’s weird that he would use the basic and tool when he doesn’t for nvidea… and i think Linus made a good video about OCing RDNA2… maybe that tool works for RDNA3, too
Super weird issue for sure, but I'm still looking forward to getting one for myself. My 2070 is starting to struggle with modern titles in 1440p. This is going to be a huge step up!
AMD makes trash junk card. Stay the 2070 for good experience.
@@taylorwillis5556 lmao
The 6950xt will give u a 3090 gaming performance,a 3070 ray tracing performance n also a 3070workstation performance with way more vram headroom
The 6950xt will give u a 3090 gaming performance,a 3070 ray tracing performance n also a 3070workstation performance with way more vram headroom
@@andrewmutavi590 what about the frames?? would they be the same?
Jay, what about undervolting? Seems to be the most important thing on AMD cards these days, and I'd imagine it would carry over to this card too.
I would love to get my hands on one of these cards. I don't have OCing on my mind, but I wouldn't mind it working as easy as it should. AMD Drivers have come a long way, but they have a few roads more to travel.
Not sure what jay is talking about here but apparently I believe it was techspot that got the 7900xtx from asus the tuf model well above 3 ghz at about that 3.2 ghz mark that jay said it would never get to. Plus he got a good amount extra performance, so yeah these cards can clock a lot higher. Not sure why they were shipped with the lower clocks and a tiny cooler but yeah.
They actually perform poorly above 3000mhz, they only benchmark well, but loses performance in everything else
@@Nobody-su9km what I am speaking of had real world gaming fps scores….
@@Nobody-su9km guru3d literally has a XFX 7900XTX reaching 3ghz and easily gaining 10-15% performance in multiple games at 4k, im not sure what your on about but they dont lose performance
I suspect it has something to do with the chiplet design and the limitation of the Infinity Cache type of interface between the GPU core and the memory caches. It's probably going to take a couple of generations to perfect it just like the Infinity Fabric on Ryzen CPUs. Even as tuned as it is now there is still a memory speed limitation that's not found on Intel CPUs
No idea about RDNA3 but I just retired a Nitro+ 6800 XT and that was a really stable OC card. Maybe I just got lucky with it but In 2 years I only had 1 crash when I was first finding the limits.
can you share the setting for nitro+ 6800XT :)
yes ops post
@@prince-sonawane I've already removed the card and switched over to nvidia. I can't access my adrenalin settings anymore. From memory though on the voltage (yellow bar) I undervolted by 2 increments on the slider. The Vram memory tune I went 2 increments up. The main clock I got up to around 2680-2700 over the stock 2500. It could go over 2700 but I didn't feel it necessary.
The adrenalin software is actually really easy to use I found. I did enjoy my 2 years using RDNA 2, but the main reason I switched back to nvidia this generation is because I was disappointed with the low number of games that had FSR support and I want to be able to play anything in 4K, which I can now, easily. 👍
With the board power limit, we need to likely under volt it. On my 1660 TUF OC (I know), I found that as I pushed ram, clocks dropped further and further. After checking power rail specs, I realized it could handle more. So, I cross flashed the bios and dropped the fans choosing H2O. Freed up quite a bit of power and much quieter.
This is something worth investigating on 7000 series very soon.
@@DaemonForce It must be the power limit, slowing RAM down happens much later on the AIB model with a higher power limit.
Well it's very likely considering PowerColor and Sapphire really push the specs close to the upper limit before shipping them out to the customers but that means pushing voltages as high as you can get away with it. You CAN still overclock these cards but overclockers are not the target audience unless getting into AHOC and at this point you know there's going to be a bios flash or 3 involved.
In my experience with AMD cards, you set the memory first then work from there.
I have a 6950XT Red Devil and I set the minimum clock 100 mhz lower than whatever the maximum clock is and it runs like a champ that way. Also, enable fast timings on the memory before clocking it.
This is exactly how I overclock my 6900xt red devil. You have to work the memory first. Lucky enough I can max out my vram to 2150 and cores at 2615/2715
May not be able to fully see any micro-stuttering from watching this video, but when the camera was panning around the dragon you can easily see chopping within the circular motion.
If only AMD would deliver what they hoped to deliver... 6950xt + 57% = 93% of 4090 as we had (6900xt vs 3090) and 6800xt + 57% = 4080
Instead they delivered 6950xt + 34% = 7900xtx and 6800xt + 34% = 7900xt (wrongly named and wrongly priced)
All current cards (except 4090) are DOA. XTX is DOA, XT is even more DOA, and 4080 for $1200 is also DOA.
The xx80 class card for $1200? A competitor from AMD for $1000? Give me a break. These prices make zero sense.
For your consideration: Economic inflation is out of control. They priced the cards to make money.
Pricing them relative to the previous generation is gonna look crazy, because money was worth more when the previous generation cards were the newest.
Now with money being worth less, they have to charge A LOT MORE to get the same "effective buying power" price.
These prices make sense with respect to the economic situation, what doesn't make sense is the economic situation relative to world events (then again maybe it's Russia making the world economy tank because of their war with Ukraine).
Dumb ass people will still buy them up
You can adjust the fan curve, just expand the fan option to bring up the settings.
To have all overclocked and undervolted, do each 1 by 1, take a note of the values.
Now select the custom settings tab, expand all the settings for gpu, vram, fan speed and enter all the values you noted earlier.
After your done, select to save the profile (small arrow in top right corner), once saved select load profile and click on the profile you just added, this will ensure it stays saved and will use the settings you saved.
Do all that, and it works fine. Cheers.
while this is a great tip, i think his point was that is also should not be this much work to get the result when it could be much simpler
That’s a lot of work just to curve fans lol sheeesh
@@dominiquevanwyk518 exactly too many issues for 1k loving my 4080 more and more
@@dominiquevanwyk518 I agree.
I found this out last week running 3DMark tests and it was driving me nuts all day. I played with the settings for about 3hrs to find out you get maximum score and performance from a slightly lower overclock, full fan, under volt, memory minor overclock.
2128 memory clock, 1050mv, 2500-2600 core clock. It worked the same way on my 6000 series Red Devils as well. I could go up to 2800Mhz+ on my 6900XT/6800XT Red Devils, but would have lower FPS and noticed this a year ago, which baffled TF out of m and took me until last week to realize what it was doing....OOF. LMAO
I think its interesting that it targets the 3GHz mark, which a number of people have also manage to manual overclock the cards to. I really think that something did go wrong with these cards and that was always the target clock speed?
Maybe AMD wanted to avoid the reputational damage of putting the hottest, biggest card out there without taking the outright performance crown in every discipline. Results from more skilled overclockers like over at TUP have pushed custom designs well over 3 GHz with ~10% performance gains. Looks like OC is back with these cards (after years of nV and AMD pushing cards to the absolute redline out of the box, which of course nVidia still does). A partner model with 3 or even 4 power connectors (or even 12VHP) would probably allow cards to get pushed quite far. Very interested to see where this is heading.
Only techpowerup. It’s random. Most ppl have struggled with these cards n opted to undervolt. Jus stop with oc crap. Who knows it will stop working suddenly or not. N then no rma lul
@@anuzahyder7185 If you really think you can't RMA a card that somehow stops working after simply adjusting clock speeds in their own software then yes, you probably shouldn't be OC'ing anything.
@@JJFX- i didnt mean it will jus stop but anyway the clock spd is all over the place for this card. Another substandard product released by amd
I think should be something related to the dual frequency and the chiplets designs
I have to disagree... The 7900xt can hit 3.3ghz quite easily, even without setting a manual overclock it'll been in excess of 3.2ghz frequently and hold it, if the game demands it, most games are limited by the crappy game engine potato code.
Power wise my 7900xt hits 465watt, but it does have 3xPCI power connections.
I'll have to check for the VRAM trade off though, that's very interesting, I'll share how I get on with it.
The other thing I want to see from these new AMD cards is more reviews of the performance when paired with Ryzen CPUs with Smart Access Memory/ Bar Resize on. I'm also curious if using MSI Afterburner works better than Adrenaline for overclocking these new AMD cards because of the fan control capabilities.
Still hoping? Just let it go.
@@anuzahyder7185 Hoping for what? Let what go? What the hell?
Friend of mine just got his 7900XTX and we started overclocking like you should "OC" a RDNA2 card. So first step baseline (its a merc 313 card) 29K3 Firestrike (clocks where close to 2500MHz).
Then we lowered the v-GPU to 1030mv clocks are now above 2800MHz and FS score is 31K2. And this was only with 15 min of effort. So if you know how it works you can get extra performance out of the card. With the Same Power limit we got +7% and this was the first attempt. So i guess 10-15% is on the table for these cards without the memory throttling down. The key is to not be power capped to hard. That was the Same with Hawaii, Fury, Vega, RDNA1, RDNA2 and now RDNA3 seems to act the same. Only the penalty for pushing to hard is now worse than before but it does not crash the card like it would previously.
Attempt 2: 1000mv now 2900+ MHz FS 31K7 now +- 9% faster than stock. And these cards are a bit faster than reference because of the higher power targets. Power is now @ 393-404 watt just below the cap for the OC bios of the card.
The AMD reference XTX is already pulling max power available to it via the 2 8-pin connectors. It can't be a surprise that pushing the power consumption up causes problems, it's just weird that AMD chooses to siphon power from VRAM rather than cause it to crash. The XTX has 2 8-pin connectors, each capable of 150W for a total of 300W. Then add in the PCIE slot itself is rated for 75W for total available power of 375W compared to the total board power of 355W. There's essentially no room for overclocking. That extra 20W is nothing.
HOWEVER, any AIB cards that use 3 8-pin connectors should have enough power available to run overclocks without cutting power to the VRAM, which explains why custom cards are noticeably outperforming the AMD reference card.
That extra 20W is something, its not nothing.
I'm in a similar situation of power limitation with my 3090 Vision that uses 2 8-pin connectors. The card can max draw 390W, but Gigabyte's BIOS updates that adds ReBar limits the power draw to 350W.
So I weight my situation on going without ReBar and keeping the 390W power draw, or taking the 40W loss and using ReBar.
@@vh9network you do know the GPU also pulls 75w from the PCIE connector, and 2-8pins can pull between 300-445w, depending on the cable your 8 pin connector is rated between 150-225w each + the 75w PCIE connector, you shouldnt be capped and shouldnt have issues, try better PSU cables so they can handle the extra load
I don't see this as a problem at all. If you want to overclock your GPU, be sure what are you doing and how it works. Not that you read somewhere that it is possible and start pushing the card to unrealistic numbers.
Anyone who knows some of this stuff will notice the VRAM clocks going down and realizes that the card hit its power limit and set the clocks a bit lower. It has to draw power from somewhere, in this case, the VRAM most likely. For the rest there is auto overclock for you.
Memory on my RX580 will let me go to around 2300MHz and while it looks stable, the performance drops because of the memory errors. "stable" but less performance, kinda the same issue but not considered as ISSUE.
Undervolt and overclock! Always. Especially on RDNA GPUs. I was able to get my 6800 xt to 2.7/2.8ghz at ~300w, maxing the voltage out on these cards you will 100% hit thermal/power limits before you even come close to the full potential of the card
Funny that AncientGamerPlays managed to undervolt a Sapphire reference model
I have both a 6600xt and a 3080 10gb, I found the AMD drivers easier to work with than the nvidia ones to overclock. Using riva tuner or other tools means installing more software, it kinda conflicted with fan control, and the fans didnt start anymore and obviously a few mns after launching any game I'd crash with the 3080 (and it went to 100C before crashing... not good!).
So yeah, maybe rdna3 is a b*tch, and it's missing options, but for casual users it's way easier to use than nvidia methods!
Exactly. Undervolting AMD gpus is much easier in the Radeon software compared to the frequency curve editor in MSI Afterburner you have to use with Nvidia GPUs.
I found the opposite, my 6600XT will initially tolerate a small overclock and be stable for a few gaming sessions only to crash for no reason and sometimes requiring the drivers to be reinstalled. Nvidia seems that bit more stable and easier to overclock. Jay's is right.
Proper overclocking requires no less than 4 softwares running concurrently, unless you have your power data displayed via hardline connection, then 3 softwares.
PCs were better when the masses couldn't keep conflating things, when they knew less. Now people just know what theyre marketed or enough to get in trouble lmao
When was the last time we saw a card that was this much of a joke? Radeon VII??
@@mattgreenfield8038 tell me you dont know anything about GPU's without telling me you don't know anything about GPU's
99% of users dont overclock. I havent overclocked anything for more than a year due to the massive increase in power costs in Europe.
you make it wrong it force clock and reduce memory because you are limited by wattage if you reduce mV you eat less energy and can push it harder still to 400w limit. correct me if I wrong but I seen video when they reduce mV to 1090 and achieve better performance than in 1150
this didn't age well
I saw a few video about undervolting
10-15%+ performance increase.
I read people get better results on the reference using afterburner. Also with an undervolt. You should revisit this trying that to see if anything changes.
I can confirm After burner is easier to oc with
@@DeeBatch good to know. My reference card direct from AMD arrives tomorrow.
cope
Mine arrived today, so keep an eye out.
@@Partyhardy28 Bro your whole life is going to videos about RDNA3 and telling people to cope, get a job 😭💀
12.00 That gap is insane, for perspective my GPU is 38th in port royal for GPU model and there is less than 400 between me and first!
Thanks for this video. I just built a new rig with a 7700x and a ASRock 6800 XT PG-D OC. This 8s my first AMD card, and like you, I found the overclocking to be much less straight forward than with Nvidia (OCing the CPU is vastly different from Intel as well, but that's a different story). But I digress, I was combing forums and online articles trying to figure out the best way to get a stable OC and came across a Reddit post where someone said they just do a "lazy" OC. What they do is go into Adrenaline, then go through the auto OC settings for GPu, undervolt, and VRAM, record those numbers, and just go into the custom setting, I put those values, switch the VRAMto Fast Timing, make a more aggressive fan curve, max out the power limit, and they're good to go. I tried this and saw a 10+ FPS bump in both the RDR2 and Cyberpunk 2077 built-in benchmarks with all graphics settings cranked to max (sans RT, because honestly the hit to FPS just isn't worth it) and GPU temps never went above 71° C. I still have to run Heaven and Superposition, but I fully expect to see significant improvements there as well. For anyone looking for an easy, stable OC method, I highly recommend this.
Yup
i hope its something the drivers can fix
really seems like its build quality issue unfortunately might want to wait on AIB boards as the reference cards are screwed
That's one thing that gives me hope for the 7900, it may age like a fine wine once the drivers get better.
You and me both brother
It's not an issue. The card behaves like it should
The software isn't ready AMD will age like fine wine! The card cost is great. What game can't it play? Best card for the money even saves on power. NVIDIA is bad but great drivers.
can play any game is a pretty low standard for a $1000+ graphics card lol.
If you can't deliver enough watts to the card (2x8pin) then it will downclock in areas to compensate.
So what people recommend doing FIRST is finding your LOWEST voltages possible without crashing, once found, try from there. This gives more headroom. If your lucky.
yeah, I'll keep looking into a 4080. The limits and issues with the 7900XTX just are not worth saving 200 bucks.
Your junction temp was hitting 93C. Are you sure that wasn't what was throttling the VRAM? It's more than likely temp throttling using some sensor. 93C seems a bit too close to where you don't want to be. GPU temp was fine....that junction sensor was HOT! Seems to be a trend with the cards.
3090s where known to hit 100-104C with very minimal loss. Micron I believe has those vram capable or running up to 100c without issue.
my 5700xt starts clocking down at around 100 gpu hotspot memory dont usualy get as hot on mine so not sure on that part
@@insiainutorrt259 yeah I remember the 3000 series cards from nvidia ran high 90s into 100s and micron said it was normal
@@Annihilator105
They say its normal but ive hat to repaste and change all the cooling pad on mine because the cooling just got worse and worse over time til i changed it
@@Annihilator105 I think der8auer figured out the issues with these cards. He got his hands on a few to check. He also found a huge difference between vertical and horizontal mounting. Essentially vapor chamber issues causing GPU hotspot throttling. Someone needs to throw a water block on the 7950XTX and figure this out. My thoughts is AMD!
It would be interesting to run and test the XTX at the same power consumption/limit as the 4080(both reference) and see the performance difference between them and the XTX between itself at stock and limited
What would be the point of that? We already know that the 4080 is more power efficient
LMAO @6:00 it's only doing it because camera is running...
i know that feeling xD
Two things for two cents:
1) AMD has an option for ""fast timings"" with no transparency about what's happening. Could that help alleviate slow memory speed?
2) do fans take up board power? Would running external fans bring some power budget back?
Question, what does the vram tuning do?
Very interesting video btw
It opens a Box to enable "Fast timings" in the momory. I think I heard it doesnt do much for the rdna 3 cards, but on like the 6000 series GPU s its a must and increases fps.
It likely does that because it only has two 8 pins, the aibs with 3 8 pins have a higher tdp. Why it’s still loosing score is likely a driver issue but who knows.
I have noticed that in 5000 and 6000 series but the MPT from Igor's Lab helps balance the frequencies if you know what to do. I got close to next tier performance with power tweaks in furmark. At stock 125 fps, with oc 132 fps but with MPT tweaks I could go to 152 fps @85c, all the way to 162 fps @100c. Clocks went from 2300mhz to 2850mhz. Ended up settling at 2750 mhz @75c for most stable performance at 145 fps.
Each section of the chip likely has separate power and clocking capabilities and this means the management core can manage thermal throttling of each section separately. Something like the memory management section of the chip is going to have a smaller surface area on the chip but is going to be working harder when the clock speeds are increased, so reducing clock speed of the memory/memory manager could be due to it working even harder and getting hotter as the cores request data faster than before... faster than the smaller surface area can cool down between requests? idk Just a theory...
Looks like my vaga 64 is going to get another year or two. It may be time to re paste her
Mine's shot. I have to use the power saver bios. It normally crashes just being on desktop over night. Any load is a crap shoot.
Lian LI a4-h20 case
x570 strix-i
5800x3d
Vega 64
all full water blocked inn this case with a FLT80
+2 LOL, don't forget to UV, increase power limit and OC the HBM
That's not shocking. You always had to validate result after applying OC on AMD. I think it's good that it tries to maintain stability. We will see if models with more power are able to OC better. Yes - it takes time, it isn't NVIDIA.... well, I'm sorry but OC was always a bit of challenge. You had to find the best configuration. If you prefer NVIDIA (crank up everything on max, hit apply and see if it is stable) - then this is fine. But don't be mad at AMD that they don't do OC the exact same way NVIDIA does.
I once read a comment of someone complaining that he had problems OC'ing a Ryzen CPU. What did he do? He simply applied the same methods he used for Intel CPUs on AMD. But Ryzens simply work different, you have to adapt, just like control elements in cars vary from brand to brand. But I guess this guy would also have had problems if OC'ing Intel CPUs worked slightly different from gen to gen.
Afair there is also performance regression on Nvidia on the 40 series (and previous series), if you do the OC the wrong way.
Cope
@@johnscaramis2515 Yeah. OC can go badly. That's sort of normal. My first OC of then new Zen+ ended up in regression. But I learned. Jayz says that some of this behavior is new to him and that it only happens on certain card - okay this might be interesting, but when you watch gamers nexus doing OC of AMD cards, they always show what they punched in and whether it gave good performance or not. Saying that increasing clocks leads to performance regression isn't nothing new on AMD. You have to pick your Frequencies very precisely and test them. It's good that he did PSA, but I'm also worried that now everyone will think that AMD cards cannot be OC'd. Which isn't true.
Well, techpowerup was able to get some insane OCs on the Tuf 7900 XTX and the Merc 7900 XTX - the tuf in particular was within 0.8 fps of the 4090 in Cyberpunk @ 4k, something that I am still shocked by. The Tuf was also around 3.2 ghz (GPU clock) stable and was drawing around 400 watts though.
@@jannegrey593 The problem here is also that afaik the power limit of 355W of the reference cards cannot be increased, as there are only 2 8-pins and the PEG-slot, so in total 366W on the 12V rail.
And the rest is simple physics/maths: if you power is capped and you increase the frequency of the GCD which requires more power as a consequence (if you don't undervolt), then there is less power for other components.
It looks Jay is starting to rap at 15:46. "(...) so performance has dropped but the clocks are up (...)". Jayminem in the house! lol
Almost looks like a remnant of the HBM era, when they've definitely had to pull down the memory in order not to overcook the heatsink.
I just don't overclock at all. I like my perfectly fine 3090 ti GPU at 22 percent power with fans at 75% playing 4K 60 frames.
Bullshit
@@PaytonPreslee-xl3bp you try it? 👀 Turn off anti-aliasing
I either hope this is an issue that gets resolved or that coverage of AIB cards is thorough, because this should be an avoidable issue
EDIT: Is it possible there's error correction happening in the GPU cache and to compensate, it's slowing everything memory down?
Could be something related to that - would make sense as to why AMD didn’t just crank the voltage and let these things fly to 3.4 ghz with full ram speed
Or the issue is Jay doesn't know how to overclock AMD cards by undervolting, and increasing power targets, and is trying to use NVidia's "just shove move" style of overclocking, and getting confused when the results are totally different?
Imagine a card with a built in failsafe for user stupidity? That future, is now!
@@bigbuckoramma Okay, sure, people do things wrong all the time. Where's your video showing him how it's done?
@@RubyRoks Why do I need to make a video on a subject that has already been made, years ago, by multiple other TechTubers that aren't NVidia fanbois?
It has long been well established, since Fury years ago, that undervolting and increasing power target is the first step to AMD overclocking, before you ever touch the frequency.
There is no excuse for missing this information. Especially if you claim to be a system builder and overclocker, such as Jay does.
@@bigbuckoramma Let me reiterate my point clearer then. Can you prove you actually know what you're talking about, or are you just upset that he's not validating your purchase?
Me loving that I have a brand new red devil 6950xt arriving today. Paid 775 for it from Amazon
I bought a 4080. Sick of buying AMD cards at launch and waiting a year for drivers to actually clear up. The fact so many are having issues while reviewing is not a good sign for this generation. 4080 has similar performance and sits at 300 watts in most games.
nice. good choice. i got a 4090 myself and im loving it. out of curiosity.... what model did you buy??
Hey Jay. Thanks for doing this video. This really helps me understand why I'm seeing weird performance behavior with my AMD GPU when overclocking.
What if you tried to overclock the memory to force the memory clocks to stay up while overclocking the gpu?
This. Why didn't Jay try this??
It will not work anyways. The card just doesn't have enough power for both GPU and MEMORY. It takes power from MEMORY to feed the GPU.
My RX 570 was super easy to overclock. Not just measurable gains but also noticable more FPS. My current RX 7900XT (power cowercolour - Hellhound) is more of a chalange. The clue is indeed undervolting. The RX 7900XT does not liike power being shoved down it's throught. Before undervolting I would get an increadable 380Watt+ (even peaking @ 400 watts) power usage on the board. After undervolting I measure 340watts max, temps are down and FPS is up.
In 3Dmark12 I used to get a 12900-ish score, now its a 15009 'score :)
I need lots of amd content i haven't been using them since 7870
I am not sure if undervolting it a bit while overclocking will help out the performance. Theoretically, a lower voltage means higher GPU efficiency, and in this case it might leave more power budget for the VRAM and infinity cache to clock up to their rated max, but seeing how these cards can run as high as 3400Mhz on the core leads me to believe that a slight reduction in voltage (1000mV) should still allow the cards to hit 3Ghz+, and if it is not bottlenecked by the low VRAM or infinity fabric clocks it might still gain quite a bit of performance.
I'm so happy that i went for my amazing ASRock RX 6900 XT OC Formula 3 months ago for 899$ ❤💙
You should stay away from these first batches, wait for a custom with 3 x 8 pin, support in morepwertools, and then you can get the most out of it. You probably need to boost power limit and undervolt core/ memory to get the best out of it.
Yeah, even my 3700X CPU from first batch is not that great. Sillicon with limited OC while newer batches and other newer SKUs easily clock higher.
@@96kylar Im not saying this card is good/ power efficient. Just that it could maybe perform better. More power intake gives you more room to tweak core upwards. If you are limited by physical aspects that it. 2x 8 pin will only allow you to safely draw 375w, while 3x 8 pin can do up to 525w...As this card seems to be quite power hungry it might matter. Of course, you still need to adjust voltages, power limit, etc.
sooo.. basically AMD is giving all users almost full juice out of the box, simplified overclocking to the most trivial possible so anyone could do it (and has almost no chance to mess things up), but overclocking nerds are not happy because AMD got all the work done for them and they couldn't do better, right?
Jay, you never once attempted to undervolt, and you also left the VRAM slider locked. The performance is going to be bad if you do that. You also thermally tanked the VRAM by shooting the GPU clock so high that the VRAM had to downclock to save itself.
The card doesn't steal power from the VRAM for the GPU. You overshot the VRAM thermal capacity limiter. You also boosted the min clock, not the max clock.
Undervolt the card as you bump the boost clocks in small intervals and boost the VRAM. Do NOT touch the minimum clock. Once you find the equilibrium in boost clock, VRAM, and voltage, you'll find the sweet spot.
Curious to see more custom cards and what may change with drivers. Also curious to see if supposed new revision is to rectify issues and bring higher clock models.
I'm debating between a 4080 and a 7900XTX in a few months when I get a bonus at work. Hopefully more is known about this issue by then. The fan speed control is going to be a non-issue either way because I'll be using a water block on it.
You going to water block a GPU? Just buy a 4090 at that point. Trust me. It'll be the same cost and better performance
Trust me, 7900 xtx release only made people buy the 4080, performance out the box and you have options like the aero, master and so on that are just beautiful... You get really good gaming compatibility, is plug and play, im all team red for cpus. but this is dumb the glitches and everything..
You mean when you get your tax return.
@@spankbuda5760 I mean either way would be his money... he paid the taxes, is the same thing. whatever people choose to do with theirs is fine. Most people finance a car and lose it by summer since cant make payments... so well.
@@txkingg No.
An air cooled GPU would not fit into my system as it sits right now, or work with how I use it. I would have to get a new case, a new CPU cooler, and probably have to upgrade my PSU as well since 4090s are so thirsty. I'd also have to run my air conditioning harder to make up for the fact that I can't dump the heat from my system into another room. An air cooled 4090 is NOT going to be the same cost.
I wonder what Jay was expecting? The reference XTX has 355W and only two 8-pin connectors. There's not much room for increased power limits, as two 8-pins and the PEG slot can only deliver 366W maximum on the 12V rail (2x150W and PEG 5.5A * 12V = 66W).
And if you are power limited, using more power in one spot must lead to a decrease in power elsewhere. That's simple physics or more like basic maths.
Just like high RAM speeds on Ryzen CPUs increased the IF frequency leading to higher power consumption of the IF, potentially leaving less power for the CCDs.
Yeah I can't wrap my head around it. He has a lot of experience with overclocking, and yet he still publishes this video, where he basically ignores everything he knows, but why
@@1337Ox I'm not sure how much he actually knows and how big his experience actually is. When watching videos where technical background is required, he often fails, because he obviously did not prepare.
I mentioned same behavior some 3 years ago in one of your videos. I did not find anyone else complaining so I thought my gpu was weak. I might be wrong but I think rx series cards are also affected by this. Cards like rx 560, 570, etc. My gpu was an RX 560
I have some information that can help with overclocking. I’m currently running a XFX 6900xt Speedster Zero with Liquid Metal applied between the waterblock. What I had to do to get those higher frequencies with VRAM and clocks was to flash the bios with the more power tool. My video card can now hit 2800mhz and a 2150mhz for memory. The stock bios that AMD cards come with even a overclock capable card retains that same conservative bios. You have to get help or figure out what is safe power and voltage for these cards. Maybe it’s different on the 7000 series but I did make it work with 6000 and got a really good increase in performance.
I like how the narrative has been for years "YES AMD DESTROY THEM!" Why were AMDs overclocking issues never mentioned in any review, ever by anyone till now?? How can anyone recommend products to the general public by keeping information to themselves??
This shows reviewers across every industry on YT can't be trusted for recommendations. This narrative is widespread.
And what overclocking issues are you talking about?
The card simply hit its power limit, cannot draw any more power. For the card to be able to hit those core clocks it needs more power, so it takes it away from memory, resulting to lower memory clocks and lower performance overall.
The solution? Try to undervolt the card and make some power headroom if possible.
When you hit power limit on nvidia card, it will start dropping the core clocks (not memory clocks like in this 7900XTX case), doesn't matter if you set the core to 2800MHz for example, the card will completely ignore it and boost only where it can depending on the power available.
It's like setting your power limit to -50% and then complaining about the card performing worse because it dropped frequency... in order to maintain the max power budget...
Only difference to "normal oveclocking" is that the card doesn't crash, but it still gives you info that it's not working optimally.
Filmed on the 16th, released on the 18th?.... Good job guys!
There are SO many 'news' type channels I watch who could really do with taking a leaf out of your book.
but he did it wrong
You need to under volt like every RDNA card before it to get the most out of the card, Stability comes from the under volt.
Stability comes from the under volt? looool 😂😂😂😂
I think to push faster clocks on both GPU and VRAM there will need to be an AIB version with three 8 pin connectors, the card doesn't have enough power maybe. Probably a conscious decision by AMD
i notice the VRAM Timing has a disable button, would turning this on stop the VRAM from going down.
Jay the ability to hut higher click depends on the program. In passmark compute I was getting over 3.2. Time Spy Extreme just under 3 on 7900 Xt.
I'm playing around with one right now for a customer build. Personally I've found the best results are typically coming from an undervolt then overclock liked you'd typically tend to do with a low to mid range Nvidia card. My personal rig runs a 3060TI so I'm used to this approach. Performance wise? It's meh, you'll notice on an FPS meter but not in any tangible practical gain in usage, but the power savings and resulting thermals? Much better.
give use some bench mark information and information on games played what setting in game what fps are you achiving?? not just meh
AMD: Gives you the performance you pay for
NVIDIA: Overcharges you and requires overclocking to get the actual performance.
Looks like people claiming that we should undervolt these cards to get better performance when overclocking
I got an easy 2971 OC on my reference 7900 xt by simply using adrenaline auto OC. But in game it’s really hitting a steady 2800mhz. By the way thanks for mentioning that you can’t control your GPU’s fans if you’re using the auto OC feature. I was wondering why afterburner wouldn’t hold my fan speeds I was setting.
Really interesting findings Jay! By all accounts, it does sound like RDNA3 was rushed to get products on shelves and sales to wrap up Q4. I’d love to see UA-camrs revisit benchmarks and overclocking in a few months when the drivers are fully baked. If we could see performance more in line with the stand out titles like Modern Warfare 2 after a few driver updates, it’d put the card more in line with what we were expecting from all the 2022 leaks
Pretty sure MW2 is just an outlier due to bad driver optimization on Nvidia cards. And at 4k max settings the 4090 is still 10% faster than 7900xtx on MW2.
7900xtx is somehow faster than the 4090 on MW2 on low settings though. So probably just a driver issue for Nvidia. MW2 was bad on Nvidia GPU’s from day 1. It had stuttering and random white flashes which has now been patched
Do you think it could be as simple as because of the smaller cooler and only having 2, 8 pin power connectors that the card is doing the best it can? I am just amazed that they were able to keep that card stable without crashing? If you want RTX performance, than yes Nvidia is the way to go!
I havent had a chance to try 7000 series AMD gpus but i think this is due to the drivers. My reference watercooled 6900 xt was such a beast when overclocking. My power draw was 330 watts for 24/7 (mined ton-eth while i was not busy on the PC.) After a year or so, the GPU degraded and i could only use it at 1.8 GHz before having a bluescreen so i bought a new one with xtxh chip. I could go above 2.8 GHz with stock power draw of 370 watts until a month ago. Now, radeon software crashes and reboots itself after the startup. I dont want to turn to the older versions of radeon software but if things dont go right, i might have to.
My radeon software crashed on start up too until the Dec 2 drivers, 6800xt (some time close to that date. Didn't/ have a slow down problem though, do you have a brace, and what aib card did you have and currently have? It can make a big difference. For instance, Zotac needs their Nvidia rights taken away, their cards are put together with trash bin components versus like EVGA. I've seen
@@theodorepollock1273 I have aorus 6900 xt xtreme waterforce, currently using it with stock power draw and 2.65 GHz clock, it was the cheapest in turkey with a custom water block. The latest update was about November which started the issues.
I don't know about that. An OC'd and undervolted 7900 XTX Tuf card was able to get 10% more performance compared to the stock Tuff and around a 23% performance increase over the 4080 reference at stock in Cyberpunk 2077. It also was also 1.2% slower than the 4090 in Cyberpunk 2077. With that said, it was pulling around 400 watts as well.
Side note: GPU clocks were at 3.2 ghz and the mem clocks were at 2.8 ghz.
wasnt there another publication that said that overclocking with the AMD software right now wasnt working but something like MSI Afterburner didnt show the same issues. Not discounting that AMD needs to correct this but if its limited to how their software is applying the overclock it something that can be worked around.