Ryzen 9 5950X - All Hail The King!

Endgame

Slightly less n00b
Joined
May 30, 2020
Messages
114
Points
28
I've seen them selling for prices that are as much or beyond that of an RTX 3080's Ebay pricing and a stones throw from an RTX 3090's actual MSRP. No thanks.
Several times per week I've seen 25+ showing up at my local microcenter, and the Amazon prices are down to about $200 over retail. I would offer to drive over to microcenter to buy one for you the next time I see stock after 11am, but I'm guessing they will be available for purchase at retail price very shortly.
 

Space_Ranger

Quasi-regular
Joined
May 16, 2019
Messages
420
Points
63
Several times per week I've seen 25+ showing up at my local microcenter, and the Amazon prices are down to about $200 over retail. I would offer to drive over to microcenter to buy one for you the next time I see stock after 11am, but I'm guessing they will be available for purchase at retail price very shortly.
Which Microcenter?? My store in NY ain't got ****..
 

Endgame

Slightly less n00b
Joined
May 30, 2020
Messages
114
Points
28
Which Microcenter?? My store in NY ain't got ****..
Rockville. Saw it once last week, once two weeks ago, and got mine the week before that on a drop of 25+. I've read they've had additional drops of 5950s at that microcenter, but since I have mine I haven't been watching it nearly as closely.
 

Dan_D

Administrator
Staff member
Joined
May 6, 2019
Messages
1,524
Points
113
Several times per week I've seen 25+ showing up at my local microcenter, and the Amazon prices are down to about $200 over retail. I would offer to drive over to microcenter to buy one for you the next time I see stock after 11am, but I'm guessing they will be available for purchase at retail price very shortly.
I've tried repeatedly at my local Microcenter but all the campers get them first. It's bad at the Dallas location.
 

Dan_D

Administrator
Staff member
Joined
May 6, 2019
Messages
1,524
Points
113
I read somewhere today that AMD thinks 90c is fine, just fine. New normal?
It is fine and has been for many years. Anything below the throttling threshold is perfectly safe for any CPU. The 68c figure that everyone quotes is generally misunderstood and misquoted as their maximum safe temperature. That's not what that value is. It never was.
 

Endgame

Slightly less n00b
Joined
May 30, 2020
Messages
114
Points
28
I found the curve optimizer and it’s been interesting. -15 all core works great and nets me a fairly substantial clock boost (about 150mhz), but cinebench single core workload crashes at anything above -5. Running a mixed negative offset (-5 on star and dot cores and -15 on everything else) doesn’t give the big clock speed boost, but is stable.

I thought there wouldn’t be much tweaking with zen 3, but with PBO + curve optimizer + if speed + tightening ram timings this is likely more to tweak than any platform I’ve ever worked with. Ryzen master is also quite confusing at times as well (with PBO enabled in bios, it tells me auto OC, not PBO, is enabled?!?)

i also read that reducing SOC voltage could help booting up with higher IF speed. That worked great and I managed to boot and compete cinebench with ram and IF set to 2000 (haven’t figured out where to set uclk yet). Unfortunately, moving to 2000 dropped my cinebench score about 300 (29504 -> 29212).
 
Last edited:

Bloax

Slightly less n00b
Joined
Jul 17, 2019
Messages
112
Points
28
Ryzen master is also quite confusing at times as well (with PBO enabled in bios, it tells me auto OC, not PBO, is enabled?!?)
Auto-OC is the Ryzen Master term for having a Max Boost Override above 0, it is all still PBO.

I've also just begun tweaking IF speed - hit 1866 without changing any voltages. 1900+ causes the host to hang, so I'm likely going to need to figure out what to tweak to try to push it higher. Once I've locked in IF speed, I'll start tightening ram timings, which I could also use a good guide for.
I spontaneously found myself with an Engrish source of info on the topic;

lowest voltage for 3800MT/s is
900-980-1050
VDDP, VDDG, vSOC @ 34ohm procODT

Along with some info on Vermeer having aggressive error-correction on the fabric.


Now, the takeaway is that high FCLKs (above 1900) work, kind of, but more by the power of error correction - worse performance - than through Just Working.
Apparently, VDDP and VDDG CCD like being matched - but I haven't personally messed with what the implications of it are, and decoupling them works fine?
What I did mess with however, is getting 2000 FCLK to perform better than my 1900 setup.

The things which I found on my setup;
An IOD voltage of 1.06v or higher (I ended up running 1.08v) is mandatory; 1.00v was horrendously unstable, 1.05v had performance penalties, 1.06v had more or less the same performance as going up to 1.08v.

Just as demonstrated in the ChipHell image, there was "performance scaling" with higher vSOC voltages; up until I set it to 1.285v, at which point it performed the same as 1.275v.

Handy tool: You can use the Latency Graph function of DRAM Calculator to see if your vSOC/IOD are too low;

This is an example with too little, and the L3 cache portion of the CPU jumps way up in latency compared to where it's supposed to be;

Here's the same CPU settings, the same memory settings, just almost enough vSOC and enough vIOD.

Why almost enough vSOC? Well, because as the first image says; the memory controller likes low (SOC) voltage.
I'm running a cranked b-die setup at 1.63v - on top of it being dual-rank, then also whacking a crazy voltage into the SOC.
It resigned after telling me to go **** myself at vSOC 1.275 :)
So it ended up being this;


And here's the kicker in my case;


The latencies, at the same CPU clockspeed, are functionally equivalent at 1.265vSOC - and mildly better at 1.275v+, but those couldn't run the memory settings ;))
Except at 1900 I get to run much less insane voltages on the internals of my CPU!
ergo it was a complete waste of time and effort, for me at least xddddddddddddddddd

If you're running less intensive memory setups though, there's probably some juice to find there.
 

Bloax

Slightly less n00b
Joined
Jul 17, 2019
Messages
112
Points
28
I found the curve optimizer and it’s been interesting. -15 all core works great and nets me a fairly substantial clock boost (about 150mhz), but cinebench single core workload crashes at anything above -5. Running a mixed negative offset (-5 on star and dot cores and -15 on everything else) doesn’t give the big clock speed boost, but is stable.

I thought there wouldn’t be much tweaking with zen 3, but with PBO + curve optimizer + if speed + tightening ram timings this is likely more to tweak than any platform I’ve ever worked with. Ryzen master is also quite confusing at times as well (with PBO enabled in bios, it tells me auto OC, not PBO, is enabled?!?)
My advice as far as Curve Optimizer tomfoolery goes, is to first do some Manual Clock/Frequency measurements;

What is the highest frequency you can run, at what voltage, in the most demanding workload you regularly find yourself doing?
which to quote past me, was x264 encoding @ 4775 mhz/1.28v in my case (even if it did hit the Thermal Barrier)

If you wish to measure the relative Core performance, you'd have to manually set the CPU Affinity of a process to only use the core you're testing, then using a set voltage (such as 1.25v), find the frequency where that core crashes.
After that, you keep the voltage, and try with the next core at -125 Mhz of the previous crash, until you find its crash-point, and repeat for all the cores zZzzZzzzzz

in my case, I chose HoMM3 as an easy task to test Max Boost on - and all my cores could do 4925 Mhz at 1.25v in that game



Which is why I have my boost offset set to +75, as that maxes my CPU at 4925 Mhz.
I could probably push it higher to +125, or +150 - but the consequences for pushing it too far are quite severe; at +200, I could barely run a curve offset of -10 stable, whereas with a +75 Boost I'm running -26/-28 and one core at -30 (the maximum!)

the result of which - combined with a PPT of 16 W for the SOC, plus roughly 12 W per core - is that I get x264 running at 4.65 Ghz around 1.25v

cheers
 

Endgame

Slightly less n00b
Joined
May 30, 2020
Messages
114
Points
28
@Bloax I see you're using a core by core negative offset. Did you get much in the way of clock speed boost adjusting it independently, or does it lock at the speed of the "worst core"? Ex, I can drop to -20 and get 4.650 ghz on all core load with PBO set to motherboard. If I desync, I can shift multiple cores to -25 with seeming stability, but clocks don't change. If I adjust the * and Dot cores to 0 offset (which they like for low thread usage and can hit 5100 mhz on single core load) and everything else to -20 or -25, the multi core clocks seem to drop to 4500, even though there is the big negative offset on the other cores.
 

Bloax

Slightly less n00b
Joined
Jul 17, 2019
Messages
112
Points
28
@Bloax I see you're using a core by core negative offset. Did you get much in the way of clock speed boost adjusting it independently, or does it lock at the speed of the "worst core"? Ex, I can drop to -20 and get 4.650 ghz on all core load with PBO set to motherboard. If I desync, I can shift multiple cores to -25 with seeming stability, but clocks don't change. If I adjust the * and Dot cores to 0 offset (which they like for low thread usage and can hit 5100 mhz on single core load) and everything else to -20 or -25, the multi core clocks seem to drop to 4500, even though there is the big negative offset on the other cores.
Based on quickly toying around with restricting an instance of Memtest to a single core, and Prime95 small FFTs on another;

Per-Core PBO Curves work weird: The maximum frequency for any core is dictated by the most heavily worked core and its (negative) curve offset.
i.e. if I'm running a Max Boost task on Core 1 - a worse core than Core 8 - and Prime95 on Core 8;
Core 1 will be boosting higher than if I was running Prime95 on Core 1 while letting the Max Boost task run on Core 8.

So humourously enough, you let your best cores do the dirtiest work - and your worst cores, the easy work, for maximum overall CPU performance.
 

Endgame

Slightly less n00b
Joined
May 30, 2020
Messages
114
Points
28
I feel all left out. No Ryzen 5000 series for me. :(
I would be playing a whole lot more if I could get the kids consistently in daycare. As it stands I spend a lot more time daydreaming about tweaking my 5950 than actually working on it. I still haven’t even wired up all the fans, USB, etc yet, and I’m looking at maybe moving to an aquaero 6LT to control the fans and adding a flow meter .

To top it off, another new BETA bios just popped up for my crosshair hero viii with the new 1.2(?) AGESA version, which might stabilize the infinity fabric at higher speeds.

So much to try, so little time
 

Peter_Brosdahl

Moderator
Staff member
Joined
May 28, 2019
Messages
2,161
Points
113
I feel all left out. No Ryzen 5000 series for me. :(
Me neither and even though these 5950s look great, I just can't justify it. Not too mention I got a crap board for it so I'd have to step up for that too.

I would be playing a whole lot more if I could get the kids consistently in daycare. As it stands I spend a lot more time daydreaming about tweaking my 5950 than actually working on it. I still haven’t even wired up all the fans, USB, etc yet, and I’m looking at maybe moving to an aquaero 6LT to control the fans and adding a flow meter .

To top it off, another new BETA bios just popped up for my crosshair hero viii with the new 1.2(?) AGESA version, which might stabilize the infinity fabric at higher speeds.

So much to try, so little time
Don't have kids but my other job has run me so ragged most days, and weekends, it's work, eat, sleep, repeat. Been a bit better during the holidays but still.
 

Bloax

Slightly less n00b
Joined
Jul 17, 2019
Messages
112
Points
28
I do no official work, I barely play things.

however I have only spent the past week or so relentlessly digging away at this silly stuff as if I'm paid to do so
alas!
 

Dan_D

Administrator
Staff member
Joined
May 6, 2019
Messages
1,524
Points
113
Me neither and even though these 5950s look great, I just can't justify it. Not too mention I got a crap board for it so I'd have to step up for that too.


Don't have kids but my other job has run me so ragged most days, and weekends, it's work, eat, sleep, repeat. Been a bit better during the holidays but still.
Well, I would buy one if I could find one.
 

Peter_Brosdahl

Moderator
Staff member
Joined
May 28, 2019
Messages
2,161
Points
113
Well, I would buy one if I could find one.
Day of release I spotted some at various Microcenters around the country but unfortunately, none were close to me. The nearest one was a seven-hour drive each way and at my age, plus the pandemic, wasn't the kind of thing I'd hit the road for. In my twenties-thirties, I would've been there though. I'm continually miffed at all those lucky people who live near one considering how often I see things instore/instock online.
 
Become a Patron!
Top