Review AMD Ryzen 9 3950X Review: 16 Cores Muscles Into the Mainstream

[jimmysmitty]

I really don't think this is true. The interface between each CCD and the I/O die is the same, regardless how many cores are enabled in the CCD. And we know both CCXs are at least partially enabled in both of the CCDs for the 3900X because it has 64 MB of L3 cache.

L3 cache is a last level cache for all cores. It is not core dependant like L1 or L2. Same goes for Intels L3 cache. So the cores could be faulty but the L3 cache could be fine so it should have no determining factors as to active cores.

Most of the cache is L3 for these CPUs to offset the latency gained buy moving the MC to a different die ad connecting with an interconnect. My assumption is that the 3900X is basically two dies that both have up to two faulty cores but the cache is fine.

This does not mean that there are the same connections active to the I/O die. The active cores need to communicate and odds are they would have more connections to keep from saturating others. Its possible the 3900x has the same amount but some are turned off with the cores.

No it shouldn't be on the Cons. In fact that AMD stated out loud that Ryzen 3xxx along with PBO already gives you the best version of the OC CPU out of the box. So whats the point?

Also, one thing I would love to know since there are no temps charts, What was the max temps when runing those test for the Intel Core i9 9980XE? Cause you said the Ryzen 9 3950X gets really hot and you need better cooling than what AMD suggest to get the greateast performance, so What about the others CPU like that Intel Extreme Edition one?

Did you also used the same Corsair 28mm AIO to cool all the CPUs used for the many benchmarks included? (sorry if this is wroten in the article and I didn't see it)

I would say poor overclocking is a con since AMD is heavily marketing towards the enthusiast crowd. The enthusiast crowd is the main overclocking crowd for PCs.

TH is not stating they need better cooling, AMD is recommending a 280mm liquid cooling AiO as the best cooler for the CPU.

The results are from THs other tests. In them they did use the same Corsair H115i for benchmarking. The only time they have not is when they go for top overclocking in which they use custom water cooling, which if they go for higher end overclocking for this I bet they will use a similar setup to get the best results.

 

[cryoburner]

Maybe you could do a review on the impact of different levels of cooling across a few CPUs -- 9900k/ks, 3900x 3950x, 9980xe?

Lower end (but still reasonable) air cooler, high end air cooler, lower end aio, high end 360 aio, and a custom loop with your dual 360s : )

That would actually make a pretty great investigative article that I think would be useful to many people.

That would actually be interesting, but be sure to also throw in a completely unreasonable $10 cooler to see what happens. : P

It wouldn't be bad to see Core i7 and Ryzen 7 models thrown in too, and maybe even Core i5 and Ryzen 5 to show much effect additional cooling might have on performance, temperature and noise levels.

 

[TJ Hooker]

L3 cache is a last level cache for all cores. It is not core dependant like L1 or L2. Same goes for Intels L3 cache. So the cores could be faulty but the L3 cache could be fine so it should have no determining factors as to active cores.

Most of the cache is L3 for these CPUs to offset the latency gained buy moving the MC to a different die ad connecting with an interconnect. My assumption is that the 3900X is basically two dies that both have up to two faulty cores but the cache is fine.

This does not mean that there are the same connections active to the I/O die. The active cores need to communicate and odds are they would have more connections to keep from saturating others. Its possible the 3900x has the same amount but some are turned off with the cores.

L3 cache is exclusive to each CCX for Zen 2 (16 MB per CCX). Having 64 MB means all CCXs are active in the 3900X (which I suppose was already apparent from the core count).

https://www.anandtech.com/show/1452...itecture-analysis-ryzen-3000-and-epyc-rome/12

Infinity fabric links CCXs within a CCD, and CCDs to the IO die. All CCXs are active for the 3900X, across two CCDs, so all IF links must be active. I don't know what other links could be active for the 3950X.

 

[TCA_ChinChin]

Yes, it isn't particularly hard to cool with a 280mm AIO. However, as charted on P2, a beefier cooler extracts higher boosts clocks and also those boosts occur more frequently. And that is at stock settings.

That's true yeah, but that isn't unique to literally any modern processor. A better cooler will always extract higher boosts and a greater boost frequency. All these chip companies have all arrived at the same conclusion, where they must develop clock speed algorithms that must dynamically scale to thermal constraints in the most efficient way possible.

 

[TCA_ChinChin]

This review did the opposite for me, it was counter productive, it actually pushed me to wait for the i9-10908XE.
In almost all benchmarks the 4.4Ghz last gen 18 core won over PBO 3950x, in the final page where it shows combined Geomean stat, it also wins by a lot, the new 18 core will do higher overclock so we can only go higher, sadly even with custom loop AMD wont overclock much.

It's an interesting point that you make. x299 will have more PCIE lanes but its not as bad as it might seem since it is PCIe 3.0 vs PCIe 4.0. Of course the 18-core Intel will do better, but it will also cost 250$ more as the 10th gen 18 core from Intel should cost about 1000$. Lastly, if you are looking at workloads where you are legitimately considering these 16/18 core processors, considering the new Threadripper will almost certainly be worth it. Especially if you are concerned about PCIe, where the new Threadripper should absolutely destroy Intel's HEDT lineup.

TLDR : If you have work that can scale to 16/18 cores, surely the new Threadrippers will be even better.

 

[Arbie]

Page 4s top says "Insert Heading".

So its about as expected with better boost rates but requires much better cooling due to the amount of cores.

Yes, it's about as expected: winning on everything. Kinda worth writing about, I think.

And... much better cooling than what, exactly? Better cooling than an Intel chip with the same number of cores? Or just something that can be made to sound like a negative? (See your other comment on OC).

 

[Arbie]

Considering it is an enthusiast CPU, yes. Overclocking is a core of the enthusiast platform. Hell it used to be one of AMDs biggest points was that all their CPUs could overclock on all boards. But with Ryzen its been limited due to the CPUs being clocked to the upper limit.

It has "limited overclocking headroom"... because it has top-notch thermal / clocking control built in!! It takes far better advantage of the silicon you've bought, automatically. How can this conceivably be presented as a negative ??

??

???

Would you not buy it because you can't overclock it? Do you not buy SSDs because you can't defrag them?

 

[Arbie]

I would rather state that instead of the 3950X being superior the 3900X is inferior.

A telling remark.

 

[jimmysmitty]

L3 cache is exclusive to each CCX for Zen 2 (16 MB per CCX). Having 64 MB means all CCXs are active in the 3900X (which I suppose was already apparent from the core count).

https://www.anandtech.com/show/1452...itecture-analysis-ryzen-3000-and-epyc-rome/12

Infinity fabric links CCXs within a CCD, and CCDs to the IO die. All CCXs are active for the 3900X, across two CCDs, so all IF links must be active. I don't know what other links could be active for the 3950X.

The links to the CPUs themselves and the cores themselves. They are obviously turned off. Not every link would have to be active for the L3 cache to be usable.

Yes, it's about as expected: winning on everything. Kinda worth writing about, I think.

And... much better cooling than what, exactly? Better cooling than an Intel chip with the same number of cores? Or just something that can be made to sound like a negative? (See your other comment on OC).

Never said it wasn't worth writing about it just that it is as expected.

And per AMD the 3950x is recommended to have liquid cooling. I made no mention of it being better or worse than Intel in any way.

It has "limited overclocking headroom"... because it has top-notch thermal / clocking control built in!! It takes far better advantage of the silicon you've bought, automatically. How can this conceivably be presented as a negative ??

??

???

Would you not buy it because you can't overclock it? Do you not buy SSDs because you can't defrag them?

I apologize. I guess making the statement that overclocking being important to the enthusiast market is bad. Except it is. Do you know why the Q6600 did so well? The G0 stepping gave 600MHz free. No need to change voltage at all just set FSB to 333MHz and bam. 3GHz quad core. If you wanted you could go further with most easily hitting 3.2GHz.

I would say its a negative since AMD touts that there is no requirement, beyond buying a quality board, for overclocking their CPUs.

Now for some people this may be great. No need to mess with the CPU at all. But a lot of enthusiasts enjoy the ability to take a product and extend its life through overclocking.

The negative is not that its a well tuned clock speed its that the product has no range beyond that. And if you do it might hinder you more than help, the 3950X OCed to 4.3GHz on all cores wasn't sipping power anymore.

Seriously when did the enthusiast market become satisfied with what they were given?

A telling remark.

Telling how? It is technically correct. If the 3900X is using the same power as the 3950X it is most likely the 3900X is an inferior CPU. It is in essence a 3950X with 4 bad cores.

I honestly would expect the 3900X to use less power with 4 less cores even if the disabled cores do still use some power.

 

[wwwkennyseqcom]

As a test it is nice and well written but I am waiting for encoding tests in OBS Studio because that's why I would like to buy this processor. Unfortunately, they are not here and the test loses its meaning to me

 

[TJ Hooker]

The links to the CPUs themselves and the cores themselves. They are obviously turned off. Not every link would have to be active for the L3 cache to be usable.

Here's what you originally said:

You forget it has 4 additional cores which mean additional links to the I/O chips as well compared to the 3900X. That adds up and uses more power

Is see no evidence that there is any difference in the links between the CCDs and the I/O die between the 3900X and 3950X. Any power required to move data to/from an individual core (which I imagine would be small relative to moving data between dies), would just be part of the overall power of a core, i.e. it'd be expected to scale linearly with core count. Compared to inter-die communication, where you'd expect to see a non-linear jump going from 8 to >8 cores.

 

[jimmysmitty]

Here's what you originally said:

There is no difference between the links between the CCDs and the I/O die between the 3900X and 3950X. Any power required to move data to/from an individual core (which I imagine would be small relative to moving data between dies), would just be part of the overall power of a core, i.e. it'd be expected to scale linearly with core count. Compared to inter-die communication, where you'd expect to see a non-linear jump going from 8 to >8 cores.

I could be wrong. Without detailed specs or a in depth analysis of it I wont say for sure. I still suspect there are some things the 3950X has that the 3900X does not that could cause higher power draw.

I think the best way to properly test it would be to take the exact same setup, put both chips in and run the same tests. Then set the clocks to the same speed, I would say 4GHz as a base, run the same tests and see if there are any power differences. Then get both to the same max and do the same.

If power remains the same across all tests then we can assume that everything except the cores is still active. Even so I would expect 4 cores not being active we should see lower power draw from the 3900x even if by only a few watts per core.

Heat should also be slightly different as the additional cores would create more heat in the areas which would need to be cooled.

I admit I am not an expert on any of this. Just going by what I have seen. While core counts doubling doesn't mean doubling of power draw it normally does have some increase in power and heat. Its why Intels 9900K can clock and boost higher than its 9980XE.

 

[Dark Lord of Tech]

Can't wait to receive mine , it's on the way!

 

[jimmysmitty]

Can't wait to receive mine , it's on the way!

You have any plans for the 3800X? Or are you doing a new build?

 

[Dark Lord of Tech]

Another Build , my youngest sons Birthday is coming up soon , I told him he can have the 3800x setup , minus the card.

 

[jimmysmitty]

Another Build , my youngest sons Birthday is coming up soon , I told him he can have the 3800x setup , minus the card.

Damn. Kids get dibs though.

 

[Ninjawithagun]

I do want to point out that it's a bit misleading by AMD to label the 3950X as a 105W TDP CPU when in fact that when all cores are being used, TDP jumps to 140-150W. You will definitely need some very good 3rd party AiO liquid cooler or custom water cooling to get the full potential from the 3950X. Just saying...

 

[Ninjawithagun]

Considering it is an enthusiast CPU, yes. Overclocking is a core of the enthusiast platform. Hell it used to be one of AMDs biggest points was that all their CPUs could overclock on all boards. But with Ryzen its been limited due to the CPUs being clocked to the upper limit.

Yes, just one more reason to wait for Zen 3. It will be out in about 8-9 months, so why even consider buying a 3900X or 3950X Zen 2 CPU? Just saying....

 

[joeblowsmynose]

I do want to point out that it's a bit misleading by AMD to label the 3950X as a 105W TDP CPU when in fact that when all cores are being used, TDP jumps to 140-150W. You will definitely need some very good 3rd party AiO liquid cooler or custom water cooling to get the full potential from the 3950X. Just saying...

That's nothing compared to a 95w TDP 9900k jumping to 250w under heavy AVX load. As far as level of misleading goes, Intel wins that category hands down.

But more seriously, TDP <> power consumption, and I feel there needs to be a 3rd party standard created for TDP numbers - maybe something with average and max values? Right now they are pretty much just marketing numbers that have no real meaning, and certainly do not equate to power consumption (or even really cooling requirements for that matter - of which that number is supposed to represent), and certainly can't be used in comparison of AMD vs Intel.

In short, don't take TDP numbers too seriously -- they are pretty much meaningless.

 

[NightHawkRMX]

Yea, i ignore tdp anymore. Neither intel nor AMD are accurate.

 

[jimmysmitty]

Yes, just one more reason to wait for Zen 3. It will be out in about 8-9 months, so why even consider buying a 3900X or 3950X Zen 2 CPU? Just saying....

I think anyone with Zen+ wastes money upgrading to Zen 2. Short of cores and slightly better boost clock speeds its like someone going from a 9700K to a 9900K. Just not worth the cost.

Yea, i ignore tdp anymore. Neither intel nor AMD are accurate.

TDP at stock is normally pretty accurate. We can't really take benchmark or stress tests numbers as more accurate. Benchmarks are normally pushing hardware while stress tests is absolute worst case scenario.

The second that you overclock however TDP is worthless. Even Ryzen when set to a single clock speed ups power draw quite a bit over stock settings. In most cases when overclocked a lot of power optimizations go out the window.

 

[joeblowsmynose]

The second that you overclock however TDP is worthless. Even Ryzen when set to a single clock speed ups power draw quite a bit over stock settings. In most cases when overclocked a lot of power optimizations go out the window.

I have the best overclocking Ryzen part ever made ... the R7 1700 ... 3.0 base clock, and I got it up to 3.9 all core 24/7, four hours of 3D rendering stable (which stresses CPU more than CPU-Z stress test)

I noticed you said back there that Ryzen's don't overclock at all - this model is the exception - some user got theirs to 4.0+ but I'd question the stability at that speed ... I could do 4.1 but its not fully stable - maybe good enough for light gaming.

I was doing some testing after updating bios the other day and I noticed that my cooler seemed to be struggling to keep it cool for some reason ... I checked HWMonitor and clocks were normal at 3.9, but power draw was 170w (default TDP is 65w) - which explained the cooling issue. After a bit of poking around I saw my voltage was at 1.6 volts ... when I applied the OC with MSI command center it also applied voltage as "auto" and I guess the CPU or bios figured that was a good voltage to use for 3.9ghz. I'm surprised it even worked - I guess a testament to my coolers abilities.

With a more reasonable voltage (but still pretty high - 1.375v) I can draw pretty much double the rated TDP when OCd to 3.9.

Just an anecdote ...

 

[NightHawkRMX]

I got my Ryzen 5 1600 which has a base of 3.2 and 3.6 boost up to nearly 4.4ghz and it was stable enough to boot Windows and browse the web. It would run 4.1ghz stable with reasonable voltage.

My 2600 with 3.6ghz base and 3.9ghz boost runs 4.2ghz stable and probably could go further, but idc.

1st and 2nd gen non x parts could overclock a good bit, but no Ryzen 3000 CPUs seem to be the same way.

 

[joeblowsmynose]

I got my Ryzen 5 1600 which has a base of 3.2 and 3.6 boost up to nearly 4.4ghz and it was stable enough to boot Windows and browse the web. It would run 4.1ghz stable with reasonable voltage.

My 2600 with 3.6ghz base and 3.9ghz boost runs 4.2ghz stable and probably could go further, but idc.

1st and 2nd gen non x parts could overclock a good bit, but no Ryzen 3000 CPUs seem to be the same way.

Yeah but with 3rd gen ryzen, RAM OCing and tweaking is where you can squeeze extra performance out -- and it does make quite a bit of difference on all Ryzen parts. There's a Ryzen RAM calculator that will give you the best sub-timings to make really tight timings and high clocks work properly (and tight timings on Ryzen are must, I'd say). I lucked out with "unsupported" corsair ram that runs at 3200 with tighter timings than XMP - and I haven't even tweaked any sub-timings yet. After a bit of testing I can confirm that RAM tweaking on ryzen is worth it.

So now Ryzen 3xxx people must learn to tweak RAM to satisfy their OC hunger.

 

[jimmysmitty]

I have the best overclocking Ryzen part ever made ... the R7 1700 ... 3.0 base clock, and I got it up to 3.9 all core 24/7, four hours of 3D rendering stable (which stresses CPU more than CPU-Z stress test)

I noticed you said back there that Ryzen's don't overclock at all - this model is the exception - some user got theirs to 4.0+ but I'd question the stability at that speed ... I could do 4.1 but its not fully stable - maybe good enough for light gaming.

I was doing some testing after updating bios the other day and I noticed that my cooler seemed to be struggling to keep it cool for some reason ... I checked HWMonitor and clocks were normal at 3.9, but power draw was 170w (default TDP is 65w) - which explained the cooling issue. After a bit of poking around I saw my voltage was at 1.6 volts ... when I applied the OC with MSI command center it also applied voltage as "auto" and I guess the CPU or bios figured that was a good voltage to use for 3.9ghz. I'm surprised it even worked - I guess a testament to my coolers abilities.

With a more reasonable voltage (but still pretty high - 1.375v) I can draw pretty much double the rated TDP when OCd to 3.9.

Just an anecdote ...

From base, sure. But from rated all core boost? Not by much. With Ryzen it is essentially better to just leave it be and use PBO.

As you said if you set an all core clock then power draw increases pretty insanely. I saw one review showing the 3950X at 4.3GHz match a 9980XE at stock.

I have no idea what Zen 3 will bring. I like anyone have to wait for its release and until then expect the same crap rumors as before like it will be a 32core 6GHz+ monster with 64 PCIe 5.0 lanes with 128GB of integrated HBM 3 memory on die.