Discussion Ryzen 5000 Two chiplet CPUs - differences between chiplets

[Kocicak]

A year ago I played with some two chiplet 3000 series CPUs and I noticed, that one chiplet tends to clock (boost) slower.

Now I already tested two 5900X and one 5950X. All three CPUs confidently reached advertised boost clocks. Both 5900X had the second chiplet approx 100 MHz slower than the first one.

Here is a screenshot of the second 5900X:



It can hit the advertised boost clock and more on ALL its cores, which is in my opinion impressive.

When I got the 5950X, I expected the second chiplet to be weaker as well. To my surprise it is cca 200 - 250 MHz slower. I did not expect such big difference. The first chiplet is hitting 5 GHz on almost all its cores. It looks very nice.

Following are two different monitoring runs of the 5950X:





I noticed, that one review sample had two chiplets capable of hitting 5GHz:



So I arrived to these questions:

How much slower is the second chiplet on your 5900/5950X?

Is it possible that there are some "lucky CPUs" getting both very fast chiplets and some less lucky getting just one?

Is it possible that if a CPU gets some excellent first chiplet which alone can fulfil some defined performance goal, it gets slow brother, and having one markedly slower chiplet is not bad luck but it simply means that the first one is exceptionally good?

Would there be some demand for "5950XX" comprising of TWO highest binned chiplets? How much extra would you pay for such a CPU?

 

[damian101]

I don't see a point in both chiplets reaching very high boost clocks. Workloads that put load on that many cores to make those slightly higher clockspeeds on the second chiplet count would put the CPU in a power or thermal limited state, preventing those extremely high boost clocks anyway.

 

[Kocicak]

Running e.g. 6 very highly clocked threads would be better, if you had them spread across two chiplets, for example for easier cooling or more L3 cash available for each active core.

 

[Mopetar]

I wonder if this is how sites like silicon lottery will stay relevant. It's no longer about getting that last 100 MHz, but being able to get it on both chiplet!

Im half joking, but I have a feeling there will be a market for this.

 

[lightmanek]

Chiplet pairing is not accidental, as one of them is chosen for maximum clock within specific power characteristic and the other is selected for maximum efficiency.
In one of the old interviews with AMD staff this was confirmed, but don't ask me where and when ...
AMD might one day release CPU binned for maximum clocks power be damned, they did it once with their Phenom II X4 TWKR Edition

 

[Kenmitch]

I'm not sure how they pair them up?

Here's my 5900x as an example. I have PBO enabled with a +50MHz boost, core optimizer individual core each with a -10 offset currently. I haven't really done much
playing around with the PBO and offsets yet, but so far so good it looks like. My cores look pretty evenly matched.



Before I messed with PBO and offsets they would all eventually boost to 4950MHz.

 

[Kocicak]

Another 4 hours of recording my 5950x, all cores from the first chiplet boost up to 5050MHz, the second chiplet is SLOOOW.

BTW I am on the newest BIOS 3001 on my Asus Strix X570 F board. I write this because I noticed that the boost behaves differently than before, especially with heavy light load, as 1 thread of Cinebench.

 

[Elfear]

I monitored my 5950 for an hour or so with the following results:



That was at completely stock settings (i.e. no PBO, ram changes, etc.). The first CCD is clocking 100-200Mhz lower than the first although I may not have given it enough time to show the true maximums.

 

[Kocicak]

Chiplet pairing is not accidental, as one of them is chosen for maximum clock within specific power characteristic and the other is selected for maximum efficiency.

Well, if this is still true, one could think about different chiplet combinations aimed at different customers.

For example, a person with workstation type of use running heavy load most of the time would be happy, if his CPU was built from two efficient chiplets, which could enable this CPU to run say 100-200MHz quicker running heavy load with given cooling solution than the universal CPU with both chiplet types.

You could imagine three types of the processors:

1) workstation type optimised for efficiency when running heavy load
2) velocity type optimised for speed
3) universal type, which is possibly the current one.

When you realise that 5950X represents the true potential of the current AM4 platform and that it is the flagship product, I believe that producing special variants of this CPU makes very good sense.

 

[lightmanek]

Well, if this is still true, one could think about different chiplet combinations aimed at different customers.

For example, a person with workstation type of use running heavy load most of the time would be happy, if his CPU was built from two efficient chiplets, which could enable this CPU to run say 100-200MHz quicker running heavy load with given cooling solution than the universal CPU with both chiplet types.

You could imagine three types of the processors:

1) workstation type optimised for efficiency when running heavy load
2) velocity type optimised for speed
3) universal type, which is possibly the current one.

When you realise that 5950X represents the true potential of the current AM4 platform and that it is the flagship product, I believe that producing special variants of this CPU makes very good sense.

I agree, but this also creates more SKU's which then have to be marketed, manufactured, binned and kept in the supply chain. There are things which could be done, but simply don't make business sense or would be lose leading products.
Anyway, these are my readings from 5900X, so probably terrible compared to others, but I'm still waiting for a BIOS with Core Optimizer feature to play with. Anyway, my fast CCD has two cores which can boost to 5100MHz and the rest will do 48xx to 49xx. Second CCD is almost always below 4850MHz.

 

[Kocicak]

Well, if you look at the amount of old gen CPUs they are still selling (and making?), I do not believe that one or two more products would make any difference and would bother them.

Actually, even more sense would have these 4 processors:

5950X universal
5950X workstation - binned for high efficiency and performance with very heavy load

5900X universal
5900X gaming - binned for highest speed

 

[hillrock100]

Here is my 5900X, with PBO offset +150 MHz and Curve Optimizer -10.
CCX0 is boost much higher than CCX1.

 

[jamescox]

Isn’t it the case that the higher clocking parts often have higher leakage and therefore higher power consumption? Not an overclocker. If they paired two high leakage, high clocking die, they would probably be over the TDP. If you have something capable of using all of the threads, you are generally going to hit power and cooling limitations, so I don’t know if there is much point in pairing two high power, high clocking parts. Overclockers think they would want that, but for two such high binned parts, it would probably need to be prohibitively expensive or possibly a lot less profitable for AMD. They can mix and match to give the desired boost clocks while still delivering a lot of cores at base clocks.

 

[Kocicak]

If they paired two high leakage, high clocking die, they would probably be over the TDP.

Overclockers think they would want that, but for two such high binned parts, it would probably need to be prohibitively expensive or possibly a lot less profitable for AMD.

I do not understand the TDP specification of these processors much. These 12 and 16 core CPUs are severely constrained by the TDP specification. Even a cheap 30 USD air cooler can cool 150W without any problems.

AMD would do that because it would be more profitable for them. After the demand/supply imbalance stabilises, it is something they should consider in my opinion.

 

[Kepler_L2]

The AM4 spec has a power limit of 142W. Many CPU/motherboard combos can easily go above this with PBO/OC, but it's not something that AMD wants to break at stock settings.

AM5 will likely increase socket PL to 180-200W.

 

[Hulk]

Chiplet pairing is not accidental, as one of them is chosen for maximum clock within specific power characteristic and the other is selected for maximum efficiency.
In one of the old interviews with AMD staff this was confirmed, but don't ask me where and when ...
AMD might one day release CPU binned for maximum clocks power be damned, they did it once with their Phenom II X4 TWKR Edition

Ian wrote about this in his Zen 3 review. His thought that the higher speed (higher leakage) part is the first chiplet while the 2nd chiplet would be the more efficient one that might not be able to reach the high single core clocks required for the spec of the CPU.

 

[Topweasel]

The only importance of the second chiplet in a properly configured machine (RE: running a Ryzen power config), would be that the second chiplet can run at boosted all core frequencies considering the power cap. There isn't anything to be gained from their point of view running two high powered CCD's with high perf caps, when the only time it can run that high is when over clocking and basically doubling up on the power cap.

In the long run how these CPU's run it doesn't make any sense to overclock, as it murders single and low core count boosts and only marginally increases all core perf. And in that sense using two premium clocking dies seems to A.) And efficiency killer b.) just wasteful in general.

they did it once with their Phenom II X4 TWKR Edition

As for this. Got to remember market times. One was late into the Core 2 life and AMD was struggling to keep up. The 9k Bulldozer was also a OMG we need something to show CPU. A absolute power be damned Ryzen doesn't do anything. Doesn't improve game play. Just makes it a better MT CPU. They sell better MT CPU's. And really they sell basically every chiplet they make and then some. Its one of the reasons we didn't see smaller TR's. The 24 and 32c versions were selling like hotcakes. Why make a liability risk high powered slightly better MT desktop CPU when they have better MT CPU's for someone to buy? Heck they could of with the XT cpu's last year, but instead worked up the ST clocks a bit and called it a day.

 

[Kocicak]

There isn't anything to be gained from their point of view running two high powered CCD's with high perf caps, when the only time it can run that high is when over clocking and basically doubling up on the power cap.

In the long run how these CPU's run it doesn't make any sense to overclock, as it murders single and low core count boosts and only marginally increases all core perf. And in that sense using two premium clocking dies seems to A.) And efficiency killer b.) just wasteful in general.

This is confusing. I do not discuss overclocking. Overclocking has no place in this topic, I am discussing stock performance.

Having two high performance (quicker) chiplets on board instead of just one would increase performance. This is obvious, right?

Intel is burning silicon and killing polar bears while trying to squeeze every drop of performance out of their CPUs, and AMD should not make any high performance version of their CPUs? Why?

BTW it would not be necessary to increase TDP of such processor much, because TDP is given by all core heavy load, and this CPU would be geared towards boosting performance in middle or lower intensity loads. You could also drop all core heavy load frequency in order to not having to increase TDP.

I already wrote, that for heavy load the other CPU would be useful, made from two high efficiency chiplets.

 

[Mopetar]

I mean AMD could and I suspect that there's a market for such a part if websites like siliconlottery.com are any indication. It would require an extra binning step at the very least and it's hard to say what percentage of chiplets can sustain enough of an extra boost to make such a part worth bringing to market. Sure it might suck to have a CPU with one golden chiplet that's being held back by something in the 10th percentile of the bin, but without better statistics it could be quite rare to see a disparity beyond 100 MHz or so.

 

[Topweasel]

This is confusing. I do not discuss overclocking. Overclocking has no place in this topic, I am discussing stock performance.

Because I think that adding two high leak CPU's would only accomplish one thing and that is under overclocking only you would get better all core clock speeds.

Having two high performance (quicker) chiplets on board instead of just one would increase performance. This is obvious, right?

It doesn't and you know it doesn't. Not with Windows running properly with the right drivers installed and scheduling plan set up. The second CCD will realistically not have a chance to clock any of its cores higher.

Intel is burning silicon and killing polar bears while trying to squeeze every drop of performance out of their CPUs, and AMD should not make any high performance version of their CPUs? Why?

Because they need to. Once upon a time they had a Arch design, Process, and clock speed lead. During this time they kept core clocks low to increase yields as high as they could and give them room to increase performance later. As AMD has caught up to them, closed in on clocks, and gotten ahead of them in IPC. They are using some of the tricks AMD has used to get max performance. That means running as close to max in single core performance and lowering that for multiple reasons as more cores are used. What intel doesn't have is a second CPU die in their desktop processors. If they were going to do a 16c 11900kRocket Lake with two dies, they would do the same use a 11700k quality Die and a 11700. Why because if its not possible to run the second die faster, why use a die that is capable.

BTW it would not be necessary to increase TDP of such processor much, because TDP is given by all core heavy load, and this CPU would be geared towards boosting performance in middle or lower intensity loads. You could also drop all core heavy load frequency in order to not having to increase TDP.

Lower Tier and lower intensity loads would be handled by the first CCD for the most part. But if you aren't trying to increase all core frequency why would you care to have a second "fast" CCD?

I already wrote, that for heavy load the other CPU would be useful, made from two high efficiency chiplets.

These aren't efficient dies. The exact opposite. High clocking dies tend to be higher leakage die. So part of this choice could also be in allowing them a balance of a really fast Single and limited MT performance on the first CCD, and to balance that out with a more efficient second CCD to get better overall all core clocks within their power allotment. Where two higher leakage CCD's might clock higher in overclocking. They would have to run the all core clocks lower to keep it under their TDP envelope.