Navi RDNA Owners Thread, Tests, Mods, BIOS & Tweaks !

Come join us..... Seriously.... we got cookies and stuff.....lol

 

Thanks, mate. It seems that once again, overclocking the memory is gonna be key, just like with Vega.

I pulled the trigger yesterday on a Sapphire RX 5700 XT Pulse for 389€ on an online store here in Spain. The lower price is due to my living in the Canary Islands (we pay lower taxes), but the downside is that the card will take 2 weeks to get here, since the order won't be shipped by boat until next week It would haven been shipped this week, but the deliveries to my region from that particular online store are done on Thursdays, and this Thursday is a public holiday in my country.

Well, I guess by the time I get the card there will be at least a newer set of drivers

 

No Comment.

 

At least the situation is improving nicely for these Navi10 GPU's and I'm guessing these will also improve the launch situation for whatever the "Navi12" and "Navi14" GPU's end up being. (5800 and 5600 or something else.) though from the Linux code commits VCN was updated to 2.0 but it might not end up for the Navi 1x series though we'll see when it's released though that's just the video decoder/encoder and it's capabilities.
(I assume a key difference between Vega lacking certain supported features and Navi having them comes down to that too even if support was initially confirmed for Vega but perhaps it's software and hardware mix and VCE hardware just couldn't do it sufficiently enough. - Video Codec Engine and Video Codec Next I think with Next replacing the earlier versions supporting up to 8k hardware decoding of certain formats.)

Also remains to be seen if AMD will tinker with the drivers to the extent they did for the Fury and Vega cards altering functionality and significantly adjusting overclocking functionality though I suppose Overdrive 8.0 found in VII and Navi is currently in a pretty good state and OC wise from what it looks like it comes down to primarily memory bottlenecks so a few tweaks to clock speed and boosts and undervolting could see a minimal or no performance decrease but a nice reduction in power usage over stock but on the flip side the cards might only show more significant performance scaling with memory bandwidth and speed increases.

Suppose that could also change for high-end Navi if they use 384 or 512-bit or just faster clocked VRAM and a stronger stock cooler to keep it under whatever thermal threshold there is.
(Not sure if GDDR6 is as sensitive to this as HBM2 was, have a lot of reading to do for a eventual Navi 20 upgrade next year ha ha.)


Guess there's not much to say will be interesting to see results from the Navi12 if that's a higher performance variant of the Navi10 chip and then the big Navi20 next year if that's the plan and the driver situation is whatever it is although support wise Polaris looks like it's struggling a bit although 19.7.x to 19.8.1 here have mostly focused in various issues with the new graphics cards and some general smaller fixes and game support plus a couple of additional Vulkan extensions though the runtime remains at 1.1.109.0 for now. (Not that it's all that important other than newer extension support and gaming wise little would be using this outside of development environments.)

AMD probably has nothing to say for now about unlocking the sharpening effect from a driver setting on earlier hardware either and then we'll see what happens with the December driver now that voting is ongoing for which features are popular.
(Could go either way really depending on how it would then be to implement these.)



EDIT: Pretty sure crypto currency GPU mining is also less active so unlike the Vega situation availability should be alright and waiting for driver updates or custom designs won't see the entire stock disappear or prices just shoot up due to demand and low availability.
(Unlike the CPU situation but that should fare better after a few months though the 3950 will probably be a bit of a rarity if binning sees priority towards Threadripper and such first.)

 

Well tomorrow I am gonna do another cheap mod on my 5700 XT I am gonna replace the Artic IV vers 2 with an Artic Accelero Hybrid III 120 GPU Cooler I have on one of my old 580's While the artic IV is a good choice for around 50.00 USD for cooling as long as you put thermal pads on the back plate and a small fan blowing across the back plate.

Anyhow this Hybrid I have worked very well on my highly overclocked and overvolted 580 and well why not do it and see what happens. The Artic Hybrids are cheap to get and already just bolt onto the 5700s and GN used a different version but I do think they should all work if they fit a 580 or 290 or the other gpu models. Anyhow I had been thinking about it since I had it and ran across this GN vid today lol. I already get the same temps as he did, but I think I will be able to improve on it once I start the project.

If I am still not quite happy will have to go full water cooling which I already have a pump to use. Just would need the water block and a rad.

And of course I will report back.

Sadly the 19.8.1 driver have a bit of worse performance whell quite a lot on Navi then then 19.7.4. I cannot no longer hit 29k with my 5700 XT



He did make an interesting point about validating actual performance with your overclock on your GPU. Since the RX 290s days I have seen stable overclocks that were higher perform worse, however sometimes it was the core clock sometimes it was vram and sometimes it was both. Higher is not always better. And this is especially true with how Vega and Navi handle boosting.

Currently I pushing 1220 on the core volts with a 70% power limit which pushes the card to 268 watts, thermals for me are fine...but I am hitting a hard limit on performance with core clock as I have stated before Navi 5700s are VRAM bandwidth limited.

Once you get your Vram to the highest stable clock, you get to the point with the core clock that even pushing it higher yields no gains. Unless you can manage to actually reach a higher stable vram clock.

I am testing something with all this power I am using and pushing through the my gpu and seems to be getting some interesting results which may end up mean for a power user like me a waterblock will make sense not due to temps. For the more normal user, the artic IV would be enough.

More power plus good thermals really help to be able to push the Navi.

So I already know I am at the peak my particular 5700 XT is gonna perform.

Hellm explained a few things to me this morning as well.

 

https://www.computerbase.de/2019-08/xfx-radeon-rx-5700-xt-thicc2-test/

 

It's certainly the best looking card XFX has put out for quite some time imo. I have to say though I'm really not that impressed with any of the aftermarket versions so far. None of them perform any better than the reference design from what i have seen. They may run a bit cooler, but that's pretty much it. Might as well just get a reference card really.

 

The only to get is the pulse, it get Vram and vrm cooling done right, it has proper coolers on them. Really hoping the Devil card will be better. And I need to see if the 8 powerconnectors will give more headroom with the powertable.

 

As always sapphire cards are the one to get. Let them bring out the nitro+ with two 8pin power.

 

I really do not know about pulling excessive current. Once can get higher clock w/o big issues. But 250W instead of 180W for 10% higher performance does not feel like good trade.
I was quite impressed when I enabled chill and got average of 135W at 120fps limit in Vermintide 2.

I only miss ability to run chill at same time as lag reduction.

 

Depends who you ask. I would take that 10% performance extra and hit 2200mhz. Its enough to actually reach 2070S.

 

You can hit 250+ on just the 8 + 6 with proper cooling and thermals. I am already hitting 268 with my powercolor ref and aftermarket Artic IV.

However as I mentioned previously, the core clock scales with the vram clock almost linearly. The vram is the true bottleneck. As I also pointed out. You can set you core clock to what you want but if you can get your vram to clock higher then the core will just pretty much run as fast as it can with vram bottleneck. And also the gddr6 is very sensitive to temp. People are able to achieve higher vram clocks on waterblocks though as I have seen I max out currently at 1810 to 1812 on my vram with my core clock set to 2190 with 1220 mv on the core and a +70 power limit.

One last note the vram voltage is not running at .850 as reported in gpu-z and hwinfo64 hellm confirmed this with me yesterday, he nor the authors of those tools knows exactly what is reporting that or where that .850 mv is coming from but the vram is running at some other voltage we do not know.

And I really do not get people wanting efficiency so badly and give up performance that is on the table. I still remember happily running R9 290x in crossfire overclock and overvolted. Very happily. As Hilbert pointed out overvolting a gpu is gonna add about 10.00 USD a year to your annual power bill.

Well I would say PowerColor or Sapphire...the red devils are monsters at cooling and are very overbuilt. But wait until the teardowns to verify.

Both Navi 5700s and Vega do not pull excessive current they are smart with power consumption and only pull the voltage they need.

Be careful with XFX Navi as they are using steel plates for the cooling plates for the vram and vrms....not a even a decent way to go, much less cooling.

 

https://www.reddit.com/r/Amd/comments/cq35sn/gamers_nexus_sapphire_rx_5700_xt_pulse_teardown/

Teardown of the RX 5700 XT Pulse via Gamer Nexus.

I found this comment pretty interesting as well besides looking at the PCB and what Sapphire has done here.

https://www.reddit.com/r/Amd/commen...s_sapphire_rx_5700_xt_pulse_teardown/ewtswf2/

So there's a 5700 Pulse and a 5700 XT Pulse and the underlying cooler is different between the two including reducing the amount of heat pipes and the fins on the heatsink which seems like important information for the difference across the two so it's not just clock speeds or some binning between the GPU chip here.


EDIT: There also seems to be some confusion around memory temperature and reported values versus actual values from reading threads on other custom Navi GPU's but someone who actually has one would have to confirm it further.

https://www.reddit.com/r/Amd/comments/cq9cxx/igorslab_de_msi_radeon_rx_5700_xt_evoke_oc/

https://www.reddit.com/r/Amd/comments/cq9cxx/igorslab_de_msi_radeon_rx_5700_xt_evoke_oc/ewuspqx/

EDIT: Which could also affect performance if the sensor is forcing it to throttle depending on if it's actually going that far or if it just reports the incorrect value but is operating within parameters for what the real temperature reading is.


EDIT: Oh and one of the AMD engineers also answered questions and some additional information from the recent blog post.

https://www.reddit.com/r/Amd/comments/cpdjrh/amd_radeon_community_update_more_control_over_gpu/

Though a lot of things can't be confirmed of course and are still under wraps but there's a few tidbits like the way integer scaling might be coming in later drivers and Wattman though overclocking is as usual unsupported and AMD can't officially support it due to liability.

 

Timely post. I was just considering this exact senerio while walking one of my dogs this morning and was about to look into vram clocks and actual temps. I need to dig out my some of thermal diodes and attach them to my vram cheaps and my fan controller. If it is misreading and reading higher then it would throttle or crash. I believe the max safe temp samsung specifies is 90 degrees and micron is 95 degrees on gddr6

We already know that all the software is reading something else for the vram voltage so this is not too far fetched.

I just looked up the specs on my PowerColor reference 5700 XT and it states this about the GDDR6

"GDDR6 Memory for Advanced Gamin Equipped with 8GB of advanced GDDR6 memory to provide high bandwidth of up to 448 GB/s" What this tells us that while the vram speed is showing half and has to be doubled it is running at the correct stock vram clock of 1750 Mhz, so at least we know this is correct.

 

Maybe a misreading with the sensors although if it's only the third party designs then it might be software related from say GPU-Z and others though it seems a bit off unless these use a different GDDR6 memory supplier and something here has changed but I don't know it wouldn't seem very likely that even with different suppliers the sensors would differ significantly either.

It will be interesting to see what it comes down to and if it's investigated further, I know AMD's CPU with Ryzen had a bit of a extra from how software read the temp value so it could be something like that as well or the sensor is more embedded and surface readings will be different from what the actual hardware is getting even if it's then quickly dissipated via the cooling solution.
(Though this too has been a bit hit & miss with some of the third party designs although if the reported value isn't quite correct even if the cooling might not be quite optimal the actual difference and reported temp value could be less than some of these early benchmarks for Asus, MSI and XFX and probably Sapphire too then are reporting.)

EDIT: Well I am still learning and it changes every time a new product is launched and AMD has significantly altered how this all works from Vega to VII and then again with Navi where the junction temperature and sensors all come together with the card design, software and bios to adjust GPU clock speeds, voltage and more over the formerly fixed DP states.
(VII has some of this but Navi utilizes it even more from the AMD representative in that Reddit topic on the recent community blog post they made on the subject.)


Could be important though and we'll probably hear more about this and a eventual fix as well wherever it's actually coming from software or hardware or the bios although that too seems unlikely though I suppose either one is a possibility and that also includes the display driver itself if that can influence how data is reported if you don't monitor via a deeper level if that's possible.



EDIT: Which yeah sounds like everything but it's a complex piece of hardware and software and the way it can go through ADL or perhaps lower level and how the sensor is read could be affected by a number of things so as a general user knowing some few things here and there this is mostly theories or what-if and then if there is a problem here it'll likely be investigated and the root cause found and corrected or worked around whichever happens.

I just hope it doesn't outright affect performance from throttling if it hits the temperature threshold but then I suppose it's more likely to hit the junction temperature threshold anyway so overall performance might be identical but measuring programs could report the wrong value or a higher than expected value which might make it appear like the VRAM's are improperly cooled.

Possibly, we shall see I suppose and if this gains traction AMD will likely also pick up on it and investigate on their end if it's something serious requiring attention.
(So many different factors and possibilities after all and what it could affect or lead to.)

 

Well bad news...while doing my 5700 XT mod half way through attaching the AIO block it slipped off my counter and cracked the die.

 

Thats a game over. Sad stuff man, try putting it back together and ask for a new one (if anyway possible).

 

It was newegg and just a few days outside the return time. Right now my vega 64 is back in my system. About to make me a drink and drown my sorrows.

 

Even i feel bad so i can imagine how you feel. Dont drink too much.