@ket Your experiences have essentially become my benchmark or baseline that I would like to test against.
I’m gonna pour over your 590 review again looking for more details/information, but I thought I might ask here; do you have the cpu and graphics scores for those runs available?
Similar to comments you made above, I’m also locked to PCIE 2.0 with an older setup. I’m on Sandy Bridge though but I’m very curious to compete against that FX system of yours, it seems pretty powerful if that’s the system that took your RX580 past a Time Spy score of 5000.
Bonus question: Does that board support resizable bar?
@bikedog The RX580 scores were on my main system which at the time was an R7 2700X @ 4.1GHz and X470 Gaming Pro Carbon with 32GB 3200MHz memory. Basemark GPU takes 99% of the system out of the equation though as the tests run @4K. The FX6300 system is just my venerable backup, she’s old, but still quite mighty I should really drop an FX8320 in sometime to round out the system. If nothing else just to see how it would then cope with Cyberpunk
EDIT: On paper the 990FX can support ReBAR but nobody has modded a BIOS for any old AM3 platform that I know of.
@karmic_koala here is the pic you asked for, the old girl is looking a lot more presentable than she was. I think the only reason she still lives is because of the quality of components, anything less with the abuse she has had, and she’d be dead. Only thing we have to hope for now is that as the cleaning had to be a lot more invasive than what should ever be necessary is that she still lives when I plug her in tomorrow… I have to heat the heatsink with my temperature gun to position the VRM plate better, it fell off and my eyeballing it rather than going to the hassle of exact measurements ended up being slightly out, about 1.5mm.
@karmic_koala Some more information, firstly, the card lives! There’s no reason why it shouldn’t but whenever you have to do an invasive clean like I did there is always a chance it might not make it out the other side because it’s so… let’s call it “battle worn”.
Second, I dumped the vBIOS from the card when I let it stretch it’s legs a bit, not sure if it even has the right vBIOS on there I haven’t looked up serial numbers yet but the card looks like a Nitro+ but with a Nitro+ OC vBIOS… might explain why @ 1.074v the card crashed quickly even at just 1.35GHz from the default 1.342GHz. 1.33GHz (which I’ll probably settle on 1.325GHz) looks promising though, perhaps doable on 1.056v. Samsung being, well Samsung, it doesn’t like OCing. 3 EDCs in a quick FFXV bench @ 2100MHz. Might be able to iron those out and keep going so I won’t write these Samsung ICs off just yet.
Third, ASIC quality is 72.2% which isn’t a bad place to be, higher would be better on a 400 series card as GPU frequency tops out realistically a little over 1.3GHz but anyone who has a 500 series card 69-72% or so is perfect. RX590s… doesn’t really matter their power envelope was thrown out the window you’re better off underclocking those GPUs slightly and profiting from huge power savings and temperature drops.
Lastly, a quick run of FFXV bench at the usual 4K, max settings, nvidia stuff off except for hairworks revealed that the RX470 I also have is batting at Nitro+ RX480 levels in the bench, 1730ish > 1766ish. Not sure if I should be impressed or disappointed by that but that could be just me as my main system has a 6800XT Red Devil and really I wanted a RX590 to play with but hardly any are on ebay right now and the ones that are there are overpriced and most of them are crap, being the nasty XFX model.
Anyway, that’s it so far I’m going to dismount the cooler on this card again to check thermal contact on everything and maybe make a tiny tweak here and there.
4C10N looks like weak link here. If you have DMM check in continuity mode if the outputs from two big inductors are tied to memory power. I would assume only one inductor is for memory
@karmic_koala Yeah, inductors using the 4C10Ns are independent, one powers the GDDR and the other the IMC, so the On-semi parts are (or should) be absolutely fine in that use case neither the GDDR or IMC are going to pull anything even coming close to maxing out the On-Semi parts. The 5335Ms are 40A per stage for a total of 200A for the GPU which is kinda overkill on a 480, as realistically you aren’t going to get much over 1.3GHz without chucking copious amounts of volts at it.
I’m throwing together a modded vBIOS this evening to see what the card can do, got a sneaky feeling it’s got a bit of pep this one on a properly tuned vBIOS.
@bikedog how you getting on? Worked out which memory timings you can try tweaking yet or do you want some hints?
What is the main PWM controller? I think it manages both VDDCI and memory phases. If so, you can find memory voltage in vBIOS.
Yeah, these mosfets look ideal for supplying memory. Until you look more closely at their specs.
46A seems max under cold condition - 25 celsius, but these will generate a lot of current and heat and thus be closer to 80 celsius in operation.
Then there are different specifications as well:
15A at 25C
11.2A at 80C
22.5A at 25C
16.8A at 80C
46A at 25C
34A at 80C
Two of them are used in parallel, that should increase available current but i wouldn’t expect twice the amount of current, maybe 50-70% more.
Almost certain, you will not get 34A with one of these, it is part rating, in actual use the figures are always lower, to keep temperatures down, avoid instabilities etc. so maybe around 25A + current boost due to parallel design and you end up with 50W power available to memory which is tight.
@karmic_koala If we assume a maximum of 50-60w available while that’s tight under a stock configuration it’s not a big deal on a custom vBIOS, most people are going to reduce VDDC. I’ve had a chance to experiment with VDDC on both 400 series cards I have and can reduce it from 1v to 0.85v without it impacting memory OCing at all. It’s also not too hard to give yourself more room to maneuver on those 4C10Ns simply by putting a heatsink on them. Main voltage controller is the IR3567B, no phase doubling going on, switching frequency is probably around 350-400 as thats pretty standard on all Polaris cards. I had a bit of a eureka moment earlier and have cracked Samsungs clock frequency wall - just completed a few tests with zero EDCs @ 2200MHz, timings are decent but obviously not fully optimised yet.
50-60w for stock config is not tight, it is more than enough. Stock config including default memory straps running at 2000mhz max. As you know, the more the clock is raised much more power the memory needs, just as if you’d overclock any other thing - same is with memory.
I did read about this and how power hungry GDDR5 becomes when going high with its clock, the manufacturers didn’t want to deal with it, and made GDDR6 with high bus width so they can achieve 2x - 4x the performance of GDDR5, but even with the GDDR6 power consumption is still an issue… to be continued i guess…
Now I’m curious, what happens on your other samsung card if you apply the timings you had success with here?
Also about 4C10Ns, sometimes conventional “maneuver” methods, such as active cooling don’t work, because VRM may have hard limit (like i had on GPU core), and if there was one in this case i would say it is software limit done by IR3567B. More likely there isn’t and the only thing restricting power is 4C10N silicone quality
@karmic_koala by tight I meant for a card running Samsung memory. Polaris is a strange animal with Samsung, the hardware just seems to be inexplicably inherently unstable. The best I can conclude at the moment is that Samsung draws more power than Hynix\Elpida\Micron as any Samsung based RX470/580 I’ve tested you can get mad tight timings for large amounts of bandwidth at stock, or dial back the bandwidth and get those high clocks to equal bandwidth at stock clocks, not both together at the same time like you can do with the other memory ICs.
Adding another bizarre twist is that Samsung memory produces less errors with tight timings at higher frequencies than it does with more relaxed timings which somewhat also throws the idea out the window that Samsung is more heat sensitive than Hynix\Micron\Elpida. The only whisper of a hint to go on is look at a lot of vBIOSes that exclusively use Samsung for that SKU and you’ll find they all have one thing in common - they all changed the maximum memory clock limit from 2250MHz to 2100MHz, so the manufacturers know something is up with Samsung memory but figuring out what that something is will be the tricky part. The only thing that might provide some insight is if we could somehow get our hands on Samsung recommended timings information for their 9Gbps memory parts.
9Gbps parts are used default in some nvidia 6GB cards, those vBIOSes are not possible to mod as far as i know, but: there are some mining tools that can read timings on the fly and modify the same as well
It would take finding someone with such type of card, who could dump strap and send it to us.
I tried asking in a place or two already, got nothing so far.
Then, i also had some other thoughts as well:
At least one card manufacturer puts out vBIOS with samsung 1500mhz strap used all over up to 2000mhz.
So it looks like overclocking samsung memory to 2250mhz could be done by using 1750mhz strap and tightening it as much as possible, overclocking to 2500mhz - using 2000mhz strap etc.
Maybe that’s a reason there is no factory 2250mhz strap anywhere.
I modified vram info to use these same timings above 1500mhz and this is where i can get up to 2100mhz.
Might give a try at 1750mhz strap
After a lot of testing today I’ve concluded that if you have a Polaris card with Samsung memory the only way forward you have is to get timings as tight as possible then OC the GPU. If you have a 400 series GPU with Samsung this means you’re pretty screwed as you’re lucky if 400 series GPUs hit 1.35GHz without a ton of voltage, same story if you have a 570/80 as GPU frequency only improved by maybe 70MHz without the need for huge amounts of voltage so you’re going to be GPU frequency limited in those scenarios. A 590 though… that would be interesting running Samsung memory with mad tight timings and the GPU easily hitting 1.5GHz without the need for a lot of voltage.
The best for instance I can give here is using the 480 I’ve tested with , I’ve now tightened timings up to crazy levels of extreme, so much so that just going from 1.32GHz to 1.35GHz bandwidth rockets up by +10GB/s, using a bit of data extrapolation using the current timings, Samsung clocked to an effective 8.1GHz, a 580 clocked to 1.425GHz bandwidth would be about 241GB/s, a 590 possibly as much as 261GB/s, assuming bandwidth scaling is linear but the only way to know that would be to have a 590 with Samsung to see when that scaling starts to taper off.
Ok, so for anyone confused essentially here is how things boil down;
RX400 Series with Samsung: Consider yourself extremely fortunate if you can do 2100MHz stable without EDCs, it simply doesn’t scale well you’ll be relying on optimised timings to do all of the heavy lifting and whatever extremely limited headroom you have left on core frequency. Prioritise bandwidth efficiency because you aren’t going to have much core or memory frequency headroom.
RX500 Series with Samsung: Almost the same story here, consider yourself extremely fortunate if you can do 2100MHz stable without EDCs, it simply doesn’t scale well you’ll be relying on optimised timings to do the majority of the heavy lifting. Prioritising bandwidth efficiency should still take precedence. The small reprieve you have is that you’ll have a bit more core frequency headroom to play around with to alleviate bandwidth bottlenecking due to insufficient core frequency.
RX590 owners with Samsung: Things get interesting here, you can “load balance” responsibilities if you will, get those timings as tight as they will go for bandwidth efficiency and then let the rather large core frequency scope pick up the lack of memory frequency scope to really rocket bandwidth to the moon.
Cards with Hynix\Micron\Elpida: The former two especially are great overclockers and you can set very nice timings, even if you lack core frequency scope these ICs can give you both high memory frequency and very good timings.
@karmic_koala maybe you can help on this question; as a final redemption arc for Samsung I searched the vBIOS for voltage reference, if I remember correctly GDDR5 has a default voltage of 1.5v so that’s what I looked for in Hex (05 DC) and found 2 references, one at offset 0000B900, the other at 00011280. I’m fairly certain the one I need is the first but not entirely sure. I’ll attach a copy of the vBIOS for this card, you thinking the same as me or have I got this one wrong? OC.rar (104.2 KB)
Patched amernime drivers BIOS signature check and loaded. No improvements in opengl. Asian modders probably can’t do it.
If you have newer driver than 22.7.1 maybe you can check if the WDDM is still 2.7 or 3.x
For memory voltage, you’d be looking for DC05 hex value. I looked quickly at your BIOS it appears memory voltage is not set in usual places, like asic init, firmwareinfo etc.
I’ll look more closely----> you should check voltageobjectinfo table
I compared it to Sapphire.RX460.2048.160714.rom there is object to set memory voltage in sapphire bios but i don’t know which controller is used in sapphire card.
In your bios there are what appears to be VDDCI voltages in the end of that table, but i think you should focus on beginning ofthat table there are values like 14 and 1C that translate into SVI2 voltages of 1.375 and 1.425. DMM would be good to confirm if changing this takes effect
@karmic_koala I might have a few more vBIOS for you to look at I suspect on closer inspection the one I dumped from this card isn’t intended for it. Dangers of buying used, especially when the card turns up in the condition this one did, one of the vBIOS was so wrong for the card it even artifacted and crashed on the desktop.
Warning for the masses: DC05 is actually backwards if you put that in the vBIOS left to right you’ll get 5.63v lol, 05DC as it’s displayed in the vBIOS is 1.5v, you have to read backwards in AMD vBIOS to make any sense of it, I don’t know why AMD do all the code backwards.
The system where you benchmarked opengl with doubled up performance is running win11?
Do you get any improvements in compute performance, e.g. oclmembench ?
I get WDDM3.1 for RX460 in win11, but APU drivers are probably bottleneck being stuck at WDDM2.7
It appears only graphic rendering would be improved and probably because of updated WDDM driver.
What has WDDM got to do in headless design (laptop) i can’t tell for sure, there is D3D render performance counter in task manager and it’s up to 100% load, but my guess is all depends on actual display GPU and thus no performance improvement can be utilized in this case
@karmic_koala Nope, system is W10 because W11 sucks, W10 isn’t great but far better than W11 at least, driver should be WDDM 3, any driver package from 22.7.1 onwards will have the OGL fix which AFAIK applies to Polaris and RDNA architectures. Bad OGL performance was purely down to a shitty implementation in the driver, I can’t speak for mobile drivers beyond I’m pretty sure those drivers are completely different from the desktop versions. Attached are the two vBIOS that I think this card is meant for, both are a clean version I got from elsewhere in case the one that did “work” dumped from the card has been messed with and the other is what I think this card probably is meant to have, GPU is slightly lower clocked but voltage-wise has scope where the other vBIOS that uses a higher GPU clock has zero scope the card runs maximum voltage constantly to maintain 1.34GHz any adjustment of it and it crashes. While shitty silicone does make it into models it shouldn’t this card just feels like the higher clocked vBIOS version is pushing the card too hard and too far for what’s good for it. So in short, ignore the other vBIOS it’s the two in this .rar archive.
Interestingly I was playing around with the 470 I have yesterday and so far in early testing reducing TDC a bit seems to help with Samsung memory EDCs, which is very odd. I’ll be seeing if I can push the memory further later I’m still trying to come up with a magic mix to get Samsung to 2150MHz stable but honestly I don’t think it’s going to happen Samsung memory just sucks I’ve had brand new 500 and 400 series cards with Samsung chuck out EDCs at entirely stock settings. Sapphire RX480 vBIOSes.rar (104.1 KB)
Yeah, same impression on win11 here. It’s ugly and sucks so much like switching from vista to lubuntu.
Now I’m not sure what is the deal with nimez drivers. Exact same version i used on win10 and win11, in win11 polaris was identified as WDDM3.1, in win10 WDDM2.7
So i don’t think win10 includes anything above WDDM2.7 and those opengl improvements are still possible with it.
It’s just something else, either the headless design of polaris cannot utilize them due to crossfire with display GPU, the drivers don’t include improved version, or they still have some bugs with it…
@karmic_koala Have you tried the regular Adrenalin drivers? I used to use modded drivers way back in the day when it was still ATi, the Omega drivers were the only drivers to have, but honestly I’ve just got finished writing most of a driver roundup review comparing Adrenalin and Pro drivers progress over the last year, the latest 23.4.3 are actually the best and there has been a steady marked step up in driver quality over the last year so it is definitely at the very least worth keep the 23.4.3 drivers around as a solid fall-back.
I’m also done with trying to get Samsung past 2100MHz. I tested some timings that literally any other IC would be cool with but not Samsung, tens of thousands of errors still in minutes. I intentionally used timings that scaled very predictably and @ 2150MHz I had bandwidth a little under what I can get completely stable @ 2050MHz with much tighter timings so yeah, Samsung can’t scale for shit and this story has remained unchanged across all the Polaris GPUs I’ve tested, it’s the memory, not the cards, and if I had to point my finger I’d point it at TCRCRL and TCRCWL the problem is if you change those timings at all the card will completely freak out.
Anyone that has Samsung memory will only see the full benefit of my timings on an RX590 because you’ll need the core frequency scalability to pick up the slack of the near zero scalability of Samsung ICs. On a 590 with my timings bandwidth with easily scale into the 240-250GB/s territory, if not 260GB/s, with the GPU at 1.5GHz+. Dammit I really want a 590 with Samsung to test that out, if anyone has one kicking about they don’t mind donating for science hit me up. These cards are so much more interesting to play with than my 6800XT, modern GPUs are boring AF being all locked down.
Anyway, let me know when you have looked over those couple vBIOSes, in the meantime I’m going to finish optimising this 470 then move back to the 480.
I should really drop an FX8320 in sometime to round out the system. If nothing else just to see how it would then cope with Cyberpunk 
and my eyeballing it rather than going to the hassle of exact measurements ended up being slightly out, about 1.5mm.
