Who's ready for the new incoming AMD CPU/GPU hardware?

[bluestang]

^ Another reason to dislike RT

[BeepBeep2]

the PBO for single core on 3000 is highly dependent on cooling solution and highly dependent on application. CB single for example don't expect miracles. even 32m pi don't expect miracles. 8m pi yes you can catch the speeds in Hardware info with the refresh speed set to the fastest setting and sometimes catch in in AOD.

With 5000 series chips you don't need voodoo magic to catch peak PBO clock speeds in monitoring programs. They are more than happy to boost to those speeds quite often.

I haven't spent a ton of time yet on the hardware but so far i see a trend with temps vs RT on off. example godfall/dirt 5 run RT and I'm seeing 80c avg where as in pure rasterization i see 74c avg temps.

I'm on a 360mm AIO with decent temps, idle 29c, ~50-55c ST load. Batch 2018 PGT (after XT launch chips) so was hoping for a little better - CHVI Hero latest BIOS. Might go 5000 series later, not sure if it's worth the investment in a new board for me.

CB15/20/23 ST moves between 4450-4525... might see what Pi does after work. Didn't really have a light enough ST workload to get it any higher. Top 4 CPPC cores are all happy to hit 4525 stock, I set hwinfo to update @ 100ms but have never seen higher. It gets a little worse with PBO on for some reason, a lot more like 4450-4500.

I tested up to 4500 on CCD 1 @ 1.34v load voltage but that's too high for more than a CB20/CB23 run.

I wonder how power usage is with RT on, genuinely curious since no reviewer appeared to test perf/w with RT

[chew*]

Pre XT maybe like 3 months prior might have had best batches of 3900 chips. you got to figure they were cherry picking them for XT variants so the non XT's would be the duds after XT launch.

Seems to be about the same just temps go up about 5-6c in dirt 5 and godfall. Dirt 5 is decently optimized can run max setting 1440P np, godfall not quite as well optimized epic @ 1440P is not really playable but stepping down a notch is perfectly fine.

[demonkevy666]

The scheduler works with the ROPs directly, and in the past the ROPs could not do out of order and had to be scheduled in clusters. One of the big reasons that AMD chips did not scale well when they got larger (vega/fiji) was the way the rops worked. It required too much latency so they were bad at high framerates. They had it detailed in the RDNA2 white paper. There was coverage but all I am finding now is about the rx6800 and not the rdna2 reveal.

I found something nice.

https://www.techspot.com/article/215...-vs-amd-rdna2/

Ok now I see why RDNA 2 is lagging a bit in RT.

[chew*]

I tossed a action sequence together in cyberpunk 2077 with 6800XT still not best case scenario just giving the card a workout in various games.

No jerking around just 1440P High preset card @ defaults.

https://streamable.com/0ru98x

I honestly think Godfall is more taxing on epic than cyberpunk2077 on ultra but keep in mind RT is on in Godfall.

Here's a godfall action sequence 1440P high.

https://streamable.com/jwz2a9

Keep in mind i'm recording with MS live so vid quality is not the best then streamable cuts quality even more.

Will be capturing with an elgato 4K 60 card as soon as i set it all up.

[chew*]

AMDs instinct card looks interesting

HBM returns

[Shocker003]

AMDs instinct card looks interesting

HBM returns

i hope they will make a 6900 graphics card with HBM, in the near future

[demonkevy666]

i hope they will make a 6900 graphics card with HBM, in the near future

I doubt it it's CDNA 1.0 That board looks slick though.
CDNA 1.0 is lacking ROP's vs Big navi, it apparently only has 64 ROP's yet it's 3 times the shaders as then 5700 XT.

[chew*]

It's possible for them to release a HBM desktop card, it's also highly improbable that they will unless it gives them some type of edge versus nvidia's answer to 6800xt's value/performance and 6900xt's price point vs 3090.

It's also highly probable they already tested that config and it did not net them the gains they needed or wanted and or effected there bottom line as far as what they make per card vs the performance.

[Shocker003]

It's possible for them to release a HBM desktop card, it's also highly improbable that they will unless it gives them some type of edge versus nvidia's answer to 6800xt's value/performance and 6900xt's price point vs 3090.

It's also highly probable they already tested that config and it did not net them the gains they needed or wanted and or effected there bottom line as farzi as what they make per card vs the performance.

I strongly believe they will. It might be their ace card. They are waiting to see how nvidia react to the 6900 price. Without all the eye candies it's almost on the same level minus the 499$

[Shocker003]

The 6900 is one hell of a card. The ref. Card is well designed. With the fans at 50 percent i could not get it to exceed 61?c while playing forza Hori.4 at 1440 ultra settings. There is no need to watercool the ref. Cards. If you play games with light oc. But when shooting for records. In benchmarks, one can put it under water or whatever

[zanzabar]

I strongly believe they will. It might be their ace card. They are waiting to see how nvidia react to the 6900 price. Without all the eye candies it's almost on the same level minus the 499$

They would not put out a card with HBM and infinity cache, and I dont think they can up the ROPs like that without the infinity cache or a bunch of HBM stacks. Maybe next gen.

[chew*]

DISCLAIMER
I may list voltages in these posts that may or may not be safe........I would not take my advice as far as voltages being safe as I'm just turning knobs to see what makes thing stable or achieves XXXX speed.
I would definitely take TheStilts advice as far as safe voltages go as he has more insider info than I do


Just a heads up and it's not really anything bad just saying don't don't don't believe the hype.

yes you can do 2000 fabric, this chip can do up to 2066 board dependant but not without cpu bus/interconnect errors @ anything over 1900.

Might be board specific might be cpu. Needs further testing. I've found ways to tune it away but not in real stability tests.

Older variant x570 or at least ch8 hero wifi can also do 2000 but require around 2.1v pll to get around the 07 post code issue.



Example of ch8 hero wifi doing 2k with 2.1 pll volts.

[demonkevy666]

I'm curious why people think Infinity Fabric is a bottleneck. Surely if it was truly a bottleneck you wouldn't have gotten 19% increase in IPC out the new 5,000 series CPU's. It would have been nothing if that was true.

[chew*]

I'm curious why people think Infinity Fabric is a bottleneck. Surely if it was truly a bottleneck you wouldn't have gotten 19% increase in IPC out the new 5,000 series CPU's. It would have been nothing if that was true.

it's a bottleneck for sure if it's not stable. there's gains to be had compared to defaults loaded, the gains from 3600 to 3800 are small now at = settings

[demonkevy666]

it's a bottleneck for sure if it's not stable. there's gains to be had compared to defaults loaded, the gains from 3600 to 3800 are small now at = settings

I mean it's not HUGE, but there is some differences.
Before I forget. Can you test that 5800X running as 4 core vs 8 core in single thread? I wanted to know if the cache size change will increase the single thread any.

[chew*]

HMM i guess so. Let me toss this up first and let prime run while i sleep to verify the below is a reality stability wise.

I experimented with Fmax scalar by stilt in bios, experimented with PBO alone then experimented with PBO curve optimizer.

By far the curve optimizer works hands down the best.

[BeepBeep2]

Have learned some about Zen 2's quirks tonight...probably should have picked one of these up a long time ago.

PBO is not useless but with some settings it seems useless / will perform worse than stock, especially when set too high. It appears there are limits in AGESA/SMU regardless of what is actually set with PBO to protect the chip from exploding.

The SMU / FIT will only ever allow it to hit ideal TDC/EDC/PPT numbers and certain settings skew it slightly worse even if not tanking it, so it takes lots of testing, combined with offset undervolt.

PBO Scalar actually does something, but raises the Core VID to unsafe voltages (1.41-1.45v+ all core) past ~4x, but can be mitigated with heavy offset undervolt, increasing MT performance significantly but at the risk of unstable ST due to max VID the chip ever chooses being 1.475-1.50v = low actual ST voltage. Had a hard time booting to windows or even posting with too much undervolt or combination of undervolt and PBO settings the SMU didn't like, - ie. Auto OC @ -0.125v offset and +75 MHz fine, but even -0.05v offset and +100, or +150 still completely unstable...or set PPT to 200w+ after -0.125v undervolt and also unstable, it is changing something with single core / light load boost aggressiveness and breaking everything.

More vdroop with PBO is better, any LLC is bad - it's possible to push to the limit of stability if undervolted as much as possible with PBO using a high scalar number, the chip will just keep going for roughly the same TDC/EDC/PPT targets and adjust it's MT Core VID to the moon.

All core OC basically useless, CCX OC not as useless, but depends how ballsy you want to be with voltage and what you are doing - CCX OC is okay if just gaming / no heavy all core loads. Example this CPU does 4.3 all core or 4.45/4.3 CCX OC @ 1.31v load - this is too high voltage for P95 Small FFT but fine for almost everything else (Blend/Cinebench/AIDA stress loop forever, etc.) but hits 75-85c on 360mm AIO in those. 4.4/4.25 can be had at 1.27v actual.

CPPC on, started at TDC 300a/EDC 300a/PPT 300w Scalar 10x, -0.05v undervolt, worked down in 20a/20a/20w increments until I gained a little more performance at TDC 180a/EDC 180a/PPT 180w.

P95 FFT 128k after 2 minutes
1.38v VID 1.23v actual - temp 75c
Clock - 3975
TDC - 121a
EDC - 157a
PPT - 173w (this was lower ~165-170w @ 200-280w PPT set)

Continued undervolting with P95 128K FFT running using ASUS AI Suite until I lost a worker thread in P95 @ -0.1375v...settled on -0.125v.

PBO TDC 180a/EDC 180a/PPT 180w, Scalar 10x, -0.125v offset, Auto OC +75MHz

P95 FFT 128K
1.41v VID 1.2v actual - temp 76c
Clock - 4100
TDC - 121a
EDC - 156a
PPT - 173w

CB R20
1.45v VID 1.25v actual - temp 68c
Clock - 4150
TDC - 95a
EDC - 161a
PPT - 150w

PBO TDC 130a/EDC 170a/PPT 180w, Scalar 10x, -0.125v offset, Auto OC +75MHz

P95 FFT 128k
1.41v VID 1.2v actual - temp 76c
Clock - 4100
TDC - 119a
EDC - 157a
PPT - 174w

CB R20 - 7433 score
1.45v VID 1.25v actual - temp 68c
Clock - 4175
TDC - 97a
EDC - 163a
PPT - 154w

PBO - TDC 130a/EDC 160a/PPT 180w, Scalar 10x, -0.125v offset, Auto OC +75MHz

P95 FFT 128k
1.4v VID 1.19v actual - temp 75c
Clock - 4100
TDC - 116a
EDC - 153a
PPT - 168w

CB R20 - 7497 score
1.44-1.45v VID 1.24-1.25v actual - temp 66c
Clock - 4200
TDC - 96a
EDC - 160a (capped)
PPT - 153w

CB R20 ST - score 527
Clock - 4500-4600 - temp 50c
1.469-1.5v VID, 1.35-1.375v actual
PPT - 45w


I've now seen peak ST frequency of 4650 with these settings as well outside of R20.

My max scores in R20 -
7867 MT / 518 ST with CCX OC @ 4450/4300 1.306v actual
7497 MT (-4.7%) / 527 (+1.7%) @ 4200 MT 1.25v actual and 4500-4600 ST 1.35-1.375v actual.

The PBO boost is even better with lighter loads ... only ~50-100 MHz off from manual CCX OC. Considering my temps are much lower and the chip will probably degrade less this way, I think I'll stay with PBO and -0.125v w/ 10x scalar.
12-thread CB20 = 4250 1.28v actual
8-thread CB20 = 4300-4325 1.3v actual
6-thread CB20 = 4300-4425 1.32v actual
4-thread CB20 = 4400-4450 1.34v actual



Something to note in this screenshot is the power reporting deviation - the PPT number is not correct under load, but the since the SMU is always hunting for a similar max, these PBO settings happen to skew the power reporting deviation even more than usual resulting in the higher boosting behavior with PBO and the ridiculous VIDs that would be nearing 1.35v actual voltage all core with no offset. With these settings + the voltage offset, the CPU is always sitting higher up in the AVFS curve than it should be, at a lower actual voltage than expected, while being told it is sucking less power from the VRM than actual.

[chew*]

ill give you an idea of how drastic the difference is between the cpu's. I'm just using PBO/curve optimizer as there is no point in manual ocing.

Bios bug forcing 3733 fabric currently on 5900x. 3800+ post all the way to 4066 but 3800 has a clock hole apparently.




Here's the curve optimizer + PBO 100 on the 5800x through prime 95.

Note the avg clock speed and the peaks..... Hwi opened after prime started running to log clock speeds.

[BeepBeep2]

ill give you an ide of how drastic the difference is between the cpu's. I'm just using PBO/curve optimizer as there is no point in manual ocing.

Bios bug forcing 3733 fabric currently on 5900x. 3800+ post all the way to 4066 but 3800 has a clock hole apparently.


Here's the curve optimizer + PBO 100 on the 5800x through prime 95.

Note the avg clock speed and the peaks..... Hwi opened after prime started running to log clock speeds.

Clock speeds on the new 5000 series with PBO + Auto OC are insane - biggest thing is it seems like they are way easier to OC that way and looks like the MT clocks hold up to whatever you could do manually? 3000 series left some on the table for both ST and MT.

Even though I got a good result for the most part, the fact that turning PBO on initially loses performance without mega tweaking on 3000 kind of sucks. I believe this CPU is capable of 4750 on the two best cores, but the algorithm doesn't go there and with the PBO scalar used to claw back MT performance with the offset voltage, ST voltages are too low to really get above 4600 anyway. Can't have it all on these like Zen 3 does it.

[chew*]

chip can manually OC to 4.5 prime stable. thermal shutdown above that. I can live with an avg clock of 4.5 through prime considering it clocks even higher with lesser demanding iterations. Primes not realistic but it's important to verify stability. It means on avg it will clock even higher when say gaming.

[chew*]

Fed ex just woke me up so excuse my bed hair reflection ( i work nights ). As the gamers tend to say nowadays. LETS GO!!!!!!!!!!!!



Now I need to install that over here in this system swap PSU out to something more tolerable to my ears and get it all in a case for real world results. Hmm maybe wall mount monitor while I'm at it to give me a little more room to work.

[demonkevy666]

Fed ex just woke me up so excuse my bed hair reflection ( i work nights ). As the gamers tend to say nowadays. LETS GO!!!!!!!!!!!!



Now I need to install that over here in this system swap PSU out to something more tolerable to my ears and get it all in a case for real world results. Hmm maybe wall mount monitor while I'm at it to give me a little more room to work.

First pic RTX on
Second pic RTX off lol
Have fun when you get the time.

There's only one game That I know of that supports Raytracing with mGPU O.o "Shadow of The Tomb Raider"

[chew*]

does multi gpu even work anymore? not like its needed anyway i already tested that to be good for 4K60. I know godfall has RT on ( amd title ) as well as dirt 5 ( amd title ).

That monitor is also HDR and HDR is not 100% free I found that out working on the helium pc with a 2080Ti. Seems it's not 100% free on AMD either so I'll probably do an on/off comparison at some point for that.

Corsair PSU inbound something a little more real world wattage wise and certainly more real world volume wise, Big thanks to Corsair and ASUS in helping get this show on the road.

[Mehdi_FXX]

Hi

Ryzen 5000 FTW! Amazing IPC performance and best price but hard to mem OC
I have heard AMD RX6000 has lower image quality in games! is that so?!