• Guest, The rules for the P & N subforum have been updated to prohibit "ad hominem" or personal attacks against other posters. See the full details in the post "Politics and News Rules & Guidelines."

Discussion Ryzen 3000 series benchmark thread ** Open **

Page 31 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

lobz

Golden Member
Feb 10, 2017
1,401
1,672
106

AMD apparently is now ripping off Intel's low end core marketing with the Athlon Gold 3150U. Not sure if legit. Despite what it says it is likely Pinnacle Ridge based.
WCCFTECH trash tier clickbait 'journalism'... they compare CPUs with way different TDP classes. They are for different form factors. No, I don't know the exact TDP of the Athlon Gold, but the very similar Athlon 300U is a 15W CPU, while the 4415Y is 4,5W. That is a ~3X difference.
I'm sure the new Athlon Gold will be a very nice - and most importantly very affordable - mobile CPU, but this article is a big pile of steaming poo :)
 
  • Like
Reactions: lightmanek

lobz

Golden Member
Feb 10, 2017
1,401
1,672
106
I cannot imagine any reason why the power consumption of the 3950X as a function of the cores used should have this weird characteristic, as measured by Anandtech. Why does it not go like I indicated with the green line?

View attachment 13213
It is called power management. The Zen 2 core is built on a very modern microarchitecture. I see nothing weird there, more cores get hotter sooner, thus the voltage is lowered sooner, which leads to a lower power consumption.
 
  • Like
Reactions: Thunder 57

Kocicak

Senior member
Jan 17, 2019
282
154
76
Heat production is directly proportional to power consumption. One could expect that when utilizing CPU more and more, you would simply hit some thermal limit and you would need to keep the power consumption at a constant level. Whether the power consumption limit would be heat related or specifications related, the curve should look like I indicated with the green curve.
 

Thunder 57

Golden Member
Aug 19, 2007
1,540
1,434
136
Heat production is directly proportional to power consumption. One could expect that when utilizing CPU more and more, you would simply hit some thermal limit and you would need to keep the power consumption at a constant level. Whether the power consumption limit would be heat related or specifications related, the curve should look like I indicated with the green curve.
And power consumption increases exponentially with voltage, and linearly with frequency. Lower core usage boosts higher (frequency) and requires higher voltage. It makes sense when you think about it. We just haven't really seen it before, maybe because we are only now getting into the many core era. Threadripper is the same. It seems it's not as simple as more cores = more power anymore. The times they are a changing.
 

jpiniero

Diamond Member
Oct 1, 2010
8,388
1,422
126
The reviews show FPS in gaming gaining a lot on 9900KS.
Not really seeing that. It looks identical to the 3900X, which is what you would expect. Maybe even a tad slower depending on frequency.

It could just be HP doing it, with AMD being complicit in setting the CPUID name string for a batch of OEM CPUs for them. Remember that thread about the Future Shop ad, with AMD A8 APUs in laptops, being sold as "10 Core". Maybe someone will do the favor of looking that one up. I believe that was also HP's doing.
Yeah you might be right. Specs wise it might be identical to the 300U.
 

Thunder 57

Golden Member
Aug 19, 2007
1,540
1,434
136
I have the same drive as you and was surprised to see some of my numbers being considerably lower. The sequential was maybe 50MB's off as I recall but the randoms were worse. Good, but not great. Well I was in Disk Management a bit earlier and saw my 850 was plugged into a B450 SATA port. Oops!

I'm sure I had it right to begin with, as my two best drives have yellow/grey SATA cables, the rest are red. But somewhere along the way I must have made a newbie mistake (probably when adding my 4TB backup drive) and figured, "meh, there all SATA 6Gbps". I put it back on one of the two "best" ports and now have similar results. Still a bit off on QD32, not sure why, but not too worried about that. Guess we all have those kinds of moments though :laughing: . I figured I'd share in case any one else sees this they may benefit as I now am.

I wish I had an exact before and after, but it was significant enough to make a difference. If anyone just needs to know I guess I can get that.

Clipboardcd.jpg

Capturecd.JPG

Mine first, yours below.
 

joesiv

Member
Mar 21, 2019
75
24
41
I cannot imagine any reason why the power consumption of the 3950X as a function of the cores used should have this weird characteristic, as measured by Anandtech. Why does it not go like I indicated with the green line?

View attachment 13213
I'm in the same boat as you, I can't imagine why this makes sense. I'll bet that it's just a bug, and will probably be fixed in the near future via agesa updates. Either they'll allow the higher core loads to consume more power, or more likely they'll dial back the mid core usage (which will affect performance during those scenarios).
 

amrnuke

Senior member
Apr 24, 2019
871
1,107
96
Not really seeing that. It looks identical to the 3900X, which is what you would expect. Maybe even a tad slower depending on frequency.
Anandtech, it's below 3900X in most games.
KitGuru has it ahead of 3900X in 3 of 8.

However...
Hexus, it's ahead.
ComputerBase.de it beats 3900X in 5 of 8 at 1080p (lowest-res).
Guru3D has it ahead of 3900X in 5 of 6 at 720p (lowest res).
HotHardware had it ahead on both gaming tests at 1080p.
Techspot has it ahead of 3900X on 6 of 6 games at 1080p.
Tweaktown has it ahead in 3 of 4 games.

Overall it seems like the consensus is that it's better for gaming than the 3900X based on the reviews I've seen.
 
  • Like
Reactions: lightmanek

scannall

Golden Member
Jan 1, 2012
1,709
1,104
136
Anandtech, it's below 3900X in most games.
KitGuru has it ahead of 3900X in 3 of 8.

However...
Hexus, it's ahead.
ComputerBase.de it beats 3900X in 5 of 8 at 1080p (lowest-res).
Guru3D has it ahead of 3900X in 5 of 6 at 720p (lowest res).
HotHardware had it ahead on both gaming tests at 1080p.
Techspot has it ahead of 3900X on 6 of 6 games at 1080p.
Tweaktown has it ahead in 3 of 4 games.

Overall it seems like the consensus is that it's better for gaming than the 3900X based on the reviews I've seen.
With such a scattering of results I'm wondering if backing down to 3200 ram, and lowering the voltage some might result in higher clocks. Some of these motherboards seem awfully aggressive on voltages. Pushing up thermals.
 
  • Like
Reactions: lightmanek

amrnuke

Senior member
Apr 24, 2019
871
1,107
96
With such a scattering of results I'm wondering if backing down to 3200 ram, and lowering the voltage some might result in higher clocks. Some of these motherboards seem awfully aggressive on voltages. Pushing up thermals.
Here are the listed RAM MT/s for each reviewer:

Anand: 3200
Kitguru: 3200

Hexus: 3200
CB.de: ??
Guru3D: 3200 and 3600 (though in results I don't see delineation)
HH: 3600
Techspot: 3200
TT: 3600

I'm not sure we have enough data points to see if that makes a huge difference. I don't see much else either - most are using X570 high end boards, 2080 Ti. Unclear why the disparate results.
 

amrnuke

Senior member
Apr 24, 2019
871
1,107
96
Might be down to the mobo.
Actually going back and looking at test beds, noted the following differences:

Motherboards: all X570 -- ASRock Taichi, GB Aorus Master (x 2), Asus Crosshair VIII Hero, MSI MEG Godlike (x 2), GB Aorus Extreme.

Cooling: Kraken X62 (x 2), Corsair H100X w/ SP120L fans, Noctua NH-D15S, AMD Wraith Prism, Corsair HydroX with XC7 block + XR5 360mm radiatio + XD5 pump/reservoir, Corsair H150i Pro.

The 3950X lost with the Taichi, both won and lost with the GB Aorus Master, won with the Asus, MSI, and GB Aorus Extreme.

It also won with both AMD Wraith Prism as well as the custom HydroX solution, both won and lost with a Kraken X62, and with the various Corsair solutions.

I can't make heads or tails of any of it, other than to say: while the general consensus is that the 3950X is faster at gaming than the 3900X, your mileage may vary depending on RAM MT/s and timings, motherboard, cooling solution, and which direction the wind is blowing in Nova Scotia.
 

Topweasel

Diamond Member
Oct 19, 2000
5,327
1,525
136
I'm in the same boat as you, I can't imagine why this makes sense. I'll bet that it's just a bug, and will probably be fixed in the near future via agesa updates. Either they'll allow the higher core loads to consume more power, or more likely they'll dial back the mid core usage (which will affect performance during those scenarios).
Amp usage. I mean in theory it should level out. But as the core count goes up the current goes up. So AMD pulls back on voltage and that causes a more dramatic drop in power usage as the cores lower in speed. The end result is AMD is taking like 10 different types of information from tons of inputs to judge when a CPU is at its limit (artificial or actual). So after about 5 cores they back off a little bit for the whole package. It might be along the lines of their 1-2c clocks as well. All about maximizing light loads for games (and benchmarks against Intel's CPU) and figure that as core count goes up its better to balance out the package stability and not push upper limits.
 

joesiv

Member
Mar 21, 2019
75
24
41
Amp usage. I mean in theory it should level out. But as the core count goes up the current goes up. So AMD pulls back on voltage and that causes a more dramatic drop in power usage as the cores lower in speed. The end result is AMD is taking like 10 different types of information from tons of inputs to judge when a CPU is at its limit (artificial or actual). So after about 5 cores they back off a little bit for the whole package. It might be along the lines of their 1-2c clocks as well. All about maximizing light loads for games (and benchmarks against Intel's CPU) and figure that as core count goes up its better to balance out the package stability and not push upper limits.
Sure, that sounds great. But it's not a pattern that we've seen in any other processor, so it seems like conjecture. I realize that these processors are complicated, and there are many things that can affect power draw, but as you say, they "back off a little bit", perhaps they just didn't back off enough at the mid end, or are backing off a bit more than they need to at the high end. It's not like AMD has a golden record for agesa code, they've been tweaking these parameters since before launch of the 3000 series!
 

Abwx

Diamond Member
Apr 2, 2011
9,113
889
126
Dramatic perfs enhancement for Ryzens when using Intel s MKLs once the CPU dispatcher is neutered :



And a review of the Athlon 3000G, that s a 3000 serie after all... :

 
  • Like
Reactions: lightmanek

Topweasel

Diamond Member
Oct 19, 2000
5,327
1,525
136
Sure, that sounds great. But it's not a pattern that we've seen in any other processor, so it seems like conjecture. I realize that these processors are complicated, and there are many things that can affect power draw, but as you say, they "back off a little bit", perhaps they just didn't back off enough at the mid end, or are backing off a bit more than they need to at the high end. It's not like AMD has a golden record for agesa code, they've been tweaking these parameters since before launch of the 3000 series!
You act like this is a run of the mill CPU. We are talking about feeding 3 different dies in a desktop CPU with desktop limitations with a certain degree of hotspoting being a possible issue as well. Why they have it peaking at that could be one of several reasons. My personal guess is that boundaries are pushed more in light loads and it backs of not just because of power usage, but die temp, current, package voltage and several other reasons for stability. We only have to look at the max core speed (4.7GHz and 4.6GHz for the 3900x) and max all core overclocks (which is as a side comment is generally lower then the max clock of any given core). There is a huge rift there. Why because while AMD uses their 7nm advantage to create some of the best all core speeds south of the 9900KS and specially for anything with its core counts. The boost system prioritizes low load boosts. In the end its probably comes down to a yield thing. As core count keeps going over 10c, the more uniform the clockspeed between cores are and probably getting clocked to the lowest common denominator within the power limits set (which again doesn't just involve voltage).

All of this applies really only to the 3900x and 3950x as a single die probably doesn't require nearly the same level of management on the power end (voltage and current) compared to the rest of the 3000 series. If you look at the peak power (and this is before agesa updates continued to tweak lower core usage up more), the 3900x sees a dip going into the 12th core in load. The answer is that at 105wTDP settings (the power group AMD tosses it into). The limits (even with BPO and autoOC enabled) AMD set on the package are barely rubbed up against on the 3900 and punched in the face by the 3950x with its 33% more cores.
 

amrnuke

Senior member
Apr 24, 2019
871
1,107
96
Stock memory on the Coffee Lake Refresh is only 2666. That makes a big difference versus 3200 in some games.
Correct, but in this line of discussion that started with my reply to your "It looks identical to the 3900X, which is what you would expect.", we aren't comparing 3950X to Coffee Lake Refresh. We are comparing it to 3900X.
 

moinmoin

Golden Member
Jun 1, 2017
1,883
2,090
106
But it's not a pattern that we've seen in any other processor, so it seems like conjecture.
I'm honestly not aware of other processors that did away with essentially any hardcoded boundaries like turbo tables and manage the system based on live statistics to the extend Zen 2 chips do. Hotspots are a serious issue on these small nodes, and the system is reacting to those.
 

joesiv

Member
Mar 21, 2019
75
24
41
I feel like you're getting really defensive on this topic. I think it's amazing what AMD has done here, and I love that they're taking a more programatic approach to squeezing as much performance out of their chips compared to old simple turbo tables.

No need to get defensive, I love AMD, and the 1950 is amazing. I feel like you're arguing over nothing really.

I probably should do this, but you said it yourself, quite well:
(and this is before agesa updates continued to tweak lower core usage up more)
AMD is tweaking via agesa updates, and recently it's been to adjust lower core count loads. Which means they're still tweaking... where that tweak falls off, is all worked out using algorythms that react to sensors (temperature mostly I believe) all over the cores. Tweaking can be done, and tweaking had to be done because it wasn't working within specs, and further tweaking could still be needed. That's all I'm saying. I'm fully happy with the performance either way.

No need to jump all over me (this seems to happen a lot on these boards, if you don't beat the same AMD does everything right drum). I'm more of an AMD fan than an Intel one, but I like discussion and I agreed with some things that @Kocicak said, that's all :)

Something relavant:


So here is an interesting case where we do see a lowering of total power after 36 threads (also with a bit of a dip after 28 threads), but in the end, at 64 threads, it does go back up to nearly the peak that it hit at 28 threads. Definitely not a linear graph (I'm agreeing with you by the way, no need to jump on me ;) )

To me I'm not sure if the peaks at 30T and 36T are intentional (in the threadripper chart above), and would ideally be smoothed out, I guess that's the crux of my thought process.
 
Last edited:
  • Like
Reactions: Elfear

joesiv

Member
Mar 21, 2019
75
24
41
Wow, the power distribution vs loading of the 2990WX is a mess!
No kidding, though the last gen threadripper was an interesting beast with some compromises with how they setup inter-die communications. It'll be interesting to see how the new threadripper does in this regard.
 

Topweasel

Diamond Member
Oct 19, 2000
5,327
1,525
136
No need to jump all over me (this seems to happen a lot on these boards, if you don't beat the same AMD does everything right drum). I'm more of an AMD fan than an Intel one, but I like discussion and I agreed with some things that @Kocicak said, that's all :)
No one is saying AMD is doing everything right. Just saying that Linear power load considering what the CPU is and how it manages turbo frequencies probably isn't realistic. This thing is figuring out how whats the best clock for max performance within all (not just max W which is ~140w for 105w CPU) perimeters. The assumption that its not doing something right and that AMD should have this correct because well all other 3000 CPU's don't see it (well the 3900x does) is a bad assumption based on bad data and is just as dangerous as saying everything AMD does is right. They still have some things to iron out with Agesa, but don't assume this is one of them, and 1.0.0.4b was written specifically for this CPU (while again ironing other issues). This is like people getting angry they couldn't get max overclocks to match top clock speeds. The 3000 is new and different beast, the 3900x and 3950x take that new setup to the extreme and AMD has been developing a turbo method that maximizes silicon without dramatically affecting yields almost to pefection on this CPU. Whether its the right way to go, thats for people to discuss for years. But its completely new and alien to everyone and its dangerous to assume that because it isn't working like you expect it to that is in fact a problem and not how it was designed.

Secondly. I wouldn't say that out loud. He hasn't made a whole lot of fans. Not everything he says is crap or unreasonable. But he has been dieing on a lot of mole hills he should be smart enough to recognize as worthless.
 
  • Like
Reactions: Elfear

joesiv

Member
Mar 21, 2019
75
24
41
No one is saying AMD is doing everything right. Just saying that Linear power load considering what the CPU is and how it manages turbo frequencies probably isn't realistic. This thing is figuring out how whats the best clock for max performance within all (not just max W which is ~140w for 105w CPU) perimeters. The assumption that its not doing something right and that AMD should have this correct because well all other 3000 CPU's don't see it (well the 3900x does) is a bad assumption based on bad data and is just as dangerous as saying everything AMD does is right. They still have some things to iron out with Agesa, but don't assume this is one of them, and 1.0.0.4b was written specifically for this CPU (while again ironing other issues). This is like people getting angry they couldn't get max overclocks to match top clock speeds. The 3000 is new and different beast, the 3900x and 3950x take that new setup to the extreme and AMD has been developing a turbo method that maximizes silicon without dramatically affecting yields almost to pefection on this CPU. Whether its the right way to go, thats for people to discuss for years. But its completely new and alien to everyone and its dangerous to assume that because it isn't working like you expect it to that is in fact a problem and not how it was designed.

Secondly. I wouldn't say that out loud. He hasn't made a whole lot of fans. Not everything he says is crap or unreasonable. But he has been dieing on a lot of mole hills he should be smart enough to recognize as worthless.
I don't know much about who you're talking to, I don't really track users and their previous replies. I saw a graph and decided that it would be really interesting to investigate, and had a quick theory on what could happen in the future. Which is fun to add to a discussion to see if it comes true in the future. And to be clear I also didn't say that it must be a bug and must be fixed. I'm not mad.

Anandtech's latest article on favored cores and windows is very interesting and goes to show that there is often much more going on under the hood of not just the processors, but the OS/agesa than we could even imagine. Basically AMD is forcing windows to favor a single CCX before spilling into the second CCX. I guess that's obvious they would want to do that, keep data as local as possible. But If you consider that information, it may shed some light on what we're talking about.

Let's look at the 3950 again with that in consideration. In particular consider that each CCX could have it's own power curve, and the transition to adding a second CCX (at 18T), Look specifically at the Core 0 voltage trend.


As more cores are added, watts for core 0 is decreased consistently, until you hit 9 cores (18T), where it jumps up. That is the transition from using 1 CCX to using two. Then we see the watts continue it's downward trend. A bit of a sea-saw. It's almost like the first 2 cores on each CCD are allowed to consume more power, relatively speaking. Note how there is a steap drop off from 2 cores to 3 cores as well.

Though it's confusing because If we can assume that the core number is consistent, and 0-7 are CCX1 (favored ccx), and core 8 is now on the second CCX, the core that's actually on the second CCX is actually consuming less watts (~12w vs ~14w) than the ones on the first CCX. So that throws a spanner in the works for my theory of the first two cores on the second CCX being ran at higher power :(

Or maybe it's a bug? Afterall, in my small uneducated mind, I would think that a fresh CCD with low core utilization could accept more watts just strickly on thermal headroom.
 
Last edited:

ASK THE COMMUNITY