Discussion in 'Videocards - AMD Radeon' started by OnnA, Oct 29, 2020.
Interesting tidbit regarding AMD's super resolution solution:
It will be similar tech to what we have on consoles IMhO.
Some checker-boarding upscale solution.
Oh, review from "this Digitalfoundry", that everybody is talking about? Good old Gamer for example? They were the first to jump on the NV PR bandwagon, I wouldn't trust them A BIT.
Nice find! Some games already implement a similar way via CAS FidelityFX e.g. Cyberpunk. IMHO SuperRes will improve the performance even further. This is just what we can expect
CAS is decent in cyberpunk. But tbh what makes DLSS stand out even with it's flaws is the blurry af TAA in the game from get go. DLSS makes a big impact in everything that uses crappy TAA tbh it makes it "better" than native because of that. I find it disingenuous while it also punishes everyone not having DLSS.
Mhmm.... I have set Sharpen to 40% in the settings. Never looked better. Might help also with the blur.
Picked up my nitro+ RX 6800 XT today and then the shop called saying they gave me the wrong card. supposed to get the OC Special Edition but they gave me the normal one. Need to go to the store to swap it tomorrow.
@chanw4 : try to overclock this puppy first, for example to 2600-2700Mhz if it holds, since you could get a worse silicon tomorrow!
I wonder why spam those "I just traded" videos... these don't bring any substance to the topic.... and I am reflecting to both of you now
Not too bad so far, going to take some tuning but an hour of testing and 2200 - 2300 Mhz though the hotspot temp is a concern and the fans while quiet needs a pretty good speed to keep it below 85 degrees junction. (~70 surface.)
The settings do behave better than the 5700 XT GPU but beyond that it's too early to tell much else, performance should be a 50 - 60% uplift (70 - 80 for the 6800XT when it performs at peak.)
Will serve until the next generation comes around and however that works out.
Adjustments so far was a simple scalar for the clock speed testing stability against error checking in OCCT simple but effective enough before getting into more sensitive games.
Now to test temps and power draw what's needed and what I can set it to plus how scaling actually behaves it seems like there's more headroom but only so far for the reference cooler models unless you crank up the fan speed.
They're improved but the fans are smaller and the 6800 has a smaller metal cooler at 2 slot rather than 2.5 for the 6800XT thus different thermals although it also has a couple less cores/units but in turn can boost above the 6800XT's clock speeds.
Until you give the 6800XT or 6900 Big Power wattage levels and put a triple slot cooler or a water kit on it at which point the peak 2600 - 2700 clock speeds are within reach if the GPU binning allows for it to be stable.
(Max of 2800 so 2700 - 2750 with the little offset from what you set and what that gives you if that's the same and also where it needs either much more power or even more voltage to gain further.)
And what happens if going the opposite, looks like a 20% performance drop is around expectations if clocking it way down to 1800 at the lowest power draw and voltage so actually less of a penalty than expected.
(Still more than Navi10 though so good to see it scaling instead of a 2 - 5% from really undervolted and downclocked to defaults or even a overclock attempt.)
EDIT: Bit of a difference in voltage set and voltage reported and scaling too.
Going to need some fine tuning here.
Hey mate! Glad you got one too! The 6800 can easily be pushed to 2.5-2.6 GHz on air, just tweak it, uV the hell out of it and be happy .
Yeah got lucky and one of the smaller web shops got in a few reference MSI 6800's with closer to reference pricing (Thus less of a issue with the GPU's being gone a minute or two later. ) so figured might as well go for it it's still a significant performance gain.
I see it also follows the 5700XT patterns and OCCT isn't quite capable of pushing it to full load or it's capping early so I only get part of the full curve after some testing.
In game testing will have to be how this is done finding a Vulkan or D3D12 title to leverage to closer to 100% and see how it operates.
Scaling was around 80% something judging by how it tweaks the set voltage and clock speeds.
(Clock it up see where it stops scaling then clock it down and yep still going lower than set speeds so it's using that curve again.)
EDIT: Was aiming for a 6800XT against a comparably priced 3080 but eh this works.
The price is about the same on this level against NVIDIA's direct competitor with the current pricing situation so it'd have been down to mostly drivers, user feedback and features with the 3080 as the current choice if prices were evenly matched when stock was available.
6800 here should perform a nice bit above 3070 levels but the 3070's supposed to cost less although with the low availability and this as a reference versus a variety of custom 3070's this was lower priced.
6500 SEK so around 650 Euro bit less than the 3070 (Around 700) but if this had been one of the non-reference then yeah anything from 700 up to 800 or more is what I see in the listed prices whereas the 3070 retains the 700 - 750 range making it either worth considering or a way better deal ha ha.
XT's models around 1000 Euro or more hoping these can normalize down closer to 900 at least.
(But for now 10.000 SEK or more just about matched with the 3080's both being non-reference too.)
Anyways heard it's even worse for areas like Eastern Europe and South America and some central European and even the US pricing also varies the way things currently are with the limited availability.
Well time to see what this can do in some more testing and what else to expect, 5700 XT Pulse lasted well enough for what punched above a very well priced mid-ranger hitting high-end performance figures but it had it's quirks though I am glad AMD really started to resolve the driver side after the initial period of all kinds of weird problems.
Next up either RDNA3 against Hopper or this rumor about Lovelace GPU getting put before Hopper whichever ends up being accurate.
Also possibilities of a soon to be revealed 3080Ti and from AMD I guess that's Zen3+ as the next CPU reveal.
So in a less stress-test condition junction is at 70 degrees Celsius and 2200 for GPU clock speed yeah that's pretty good mild undervolt (1025mv -> 1000mv) but can be tuned further.
(2050 to begin with for memory see how that goes guessing 2075 - 2100 is about where it caps and then it either stops scaling or decreases if error correction kicks in.)
Just have to fine tune it a bit.
Pity you haven't told me earlier. I was selling my 6800xt two weeks ago / once I have bought the 6900xt. For the price I have bought it in the shop (MSRP +12% overhead but ohwell... ).
And IMHO Lovelace is the non MCM design... I think something has gone wrong and NV scratched the MCM design for now.... And I kinda am interested now how the RDNA3 MCM design will fare against Lovelace AMD is back in the game, they need to implement "RDLSS" now to shut up all the naysayers and then can focus on the MCM + RT improvements.... Not that RT matters for now...
IMHO the 6800 can uV even higher.... A few pages back there was a recommendation from someone... can't recall atm....
Plus you can bump the mem to 2100 without breaking a sweat and set the GPU MHz limits between 2350-2550 plus bump the power limit to +10 or 15%
My usual store can't get a 6800 xt or 3080 but has 6900 in stock lol
I DONT WANT to throw 1200€ lol god give me strengh
Yeah I've done a bit of testing now and so far it hits 2300 easily (2350 set in Wattman.) there's a few differences versus the values the 6800 XT gets but the GPU still scales really well even without advanced tweaking like More Power Tool.
Has a few less cores but the resulting increased clock speed and tweaks to voltage can be pretty impressive 0.900v at 2200Mhz appeared to be stable but needs further testing to confirm fully and then finding where it does start either requiring more or outright hitting errors. (2450 to 2500 seem to be where that's usually at.)
Overclocking has more gains too trying to see if I can balance a 5% general increase against a undervolt then work on memory a bit seeing where that goes. (2100 as the frequency I'm seeing in many of the reviews or the full 2150 it's not much but there's some small gains from it.)
Mostly the usual incremental 25 or 50 here and there test for 30 minutes in some of the more stressful D3D12 or Vulkan titles and see what Wattman reports and if there's any gains or you just hit a software crash ha ha.
EDIT: Shipping costs here have increased a bit not exactly sure on the number though but I've seen figures around quadruple the usual cost so sending stuff across the border is difficult no wonder if the report is accurate that some of the incoming shipments have been slow.
Also curious as to how this chip design is going to work for GPU's suppose there's a chance AMD's going to be going for it with RDNA3 unless it gets pushed back.
Maybe a area where a whole lot of fast cache / infinity fabric comes in handy but going by the little info I've seen on it they take up a fair amount of GPU die space so you can't just add a huge amount of it either.
6850 XTX with a 512-bit / 512-MB IF would be fun though, see what the GPU can scale to when the cache memory can cope with 4096x2160 or even higher output.
(Well 3840x2160 up to 5120x2880 display resolution rather. )
Using resolution scaling it holds up but there is a more noticeable performance drop as well as this limitation starts showing up a little bit more.
Might be something the RDNA3 cards have a solution for alongside I presume also the second generation ray tracing enhancements making it a bit more viable on AMD GPU's.
Well it works but a 30 FPS cap might be a idea if there's more than one effect in action currently and the settings this uses plus the type of effect though all are fairly heavy to process.
(More of a turn it off and utilize the GPU power for the framerate instead here personally but probably going to at least test it and see how it goes.)
EDIT: Stock performance isn't bad or anything either, I like to do some tweaking and there's usually some headroom plus configuring fan speeds depending on efficiency and against how loud this becomes over AMD's often fairly conservative default scaling.
It's nice that the card can get close to the stock 6800 XT performance though it's also nice that scaling and efficiency leverages the additional cores better so the 6800 XT and even the 6900 XT can utilize these more fully for a nice further uplift to overall performance.
Driver limit on power draw and some of the caps for voltage and max speeds unless using higher limits via software adds some limitations. (Though some are capped and can't be increased or if increased the GPU clocks lock at a low value.)
But the 6800 at least has enough room where it scales up nicely even without these changes.
Some of the custom 6800 XT's and 6900 XT's should also have a higher TBP value in the bios already which can then be given the 15% power draw increase if required allowing for some extra gains.
Voltage might be a bit of a limit when there's 72 CU's or the full 80 over the 60 in the 6800 for overclocking but the performance is also a nice bit faster by default already since the scaling is not as limiting as it was with Vega or earlier.
Lots of reading to though but from what I'm getting thus far for what's going on with this.