Question Zen 4 X3D and speculation about Dragon Range X - Intel's nightmare

[igor_kavinski]

AMD Zen 4 X3D SKUs possibly the Ryzen 7 7800X3D and Ryzen 9 7950X3D reportedly launching early next year to take the gaming crown from Intel - NotebookCheck.net News

I think AMD will rain upon Intel's Raptor Lake parade with Zen 4 X3D samples sent to the press. Release may still happen in Jan-2023.

The possibility of Dragon Range laptops having V-cache excites me like nothing else! It would be the ultimate dagger shaped nail in Intel's coffin, as far as their aspirations for gaming supremacy go.

FR3cm_wXEAA3afw (3271×1753) (twimg.com)



That almost confirms that AMD is bringing V-cache to mobile gaming.

 

[A///]

A lot of AMD's competition is likely to be a significant disappointment unless the ARMy shows up big in the non-Apple PC space. Still waiting for that to happen.

you'll never be able to comprehend how loudly I groaned at your ARM pun. I have mixed emotions on it. Ta M'Lord.

 

[TESKATLIPOKA]

That would be a significant disappointment if that turns out to be the case. 6+8 may have a tough time against even Phoenix (believed to be 8p) in a heavily power constrained implementation as Phoenix will likely have significantly less power overhead as compared to desktop Zen4 with its N6 IOD on top of being on a more mature N5 node.

I wouldn't be so sure about Meteor Lake being worse than Phoenix.
Here is a table about MT performance at different power limits. Link: ComputerBase

As you can see, If you limit the power to 45W then 13900K is the most efficient of the bunch and 6+8 13600K is just a bit worse than 7700x.
This looks pretty bad for Zen4 considering It uses better process for both IOD(N6) and CCD(N5).
Of course the chiplet design is worsening the efficiency of Zen4, so a monolithic 8C Phoenix would perform better at 45W in this table, the question is by how much.
Increasing the power for 7700x by 20W(+44%) increased performance by just ~20%, but the chiplet overhead is hardly 20W, so Phoenix at 45W would be <98% in the table unless the supposed N4 process and additional changes makes up the difference.
Let's say Phoenix will be 20% faster at 45W.

Now back to Meteor Lake.
6+8 Meteor Lake would have the same configuration as 13600K and If you check the above table the 13700K with +2 P cores performs only 1% better at 45W, so only having 6+8 is not a surprise or a disappointment.

Meteor Lake should use Intel 4 and that supposedly reduces power consumption by 40% at ISO frequency or being 20% faster at ISO power.
I made a table based on the above table where I apply Intel 4 gains to Raptor Lake.

	7600x	7700x	7900x	7950x	Phoenix	13600K	13700K	13900K
Performance at 45W	69%	82%	83%	84%	98%	79%	80%	92%
Intel 4 at ISO power	69%	82%	83%	84%	98%	95%	96%	110%
Intel 4 at ISO frequency	69%	82%	83%	84%	98%	79% at 27W	80% at 27W	92% at 27W

Meteor Lake should use Redwood Cove for P-cores and Crestmont for E-cores. That would mean higher IPC than Raptor Lake, which will additionally increase performance at 45W than what I calculated in the above table.
Not sure how the performance at lower TDPs will look like, but Meteor Lake-U has 4P+8C cores. Pretty sure Phoenix IGP will be faster considering Meteor IGP will be limited to 128EU from what I read on net.

P.S. I just noticed, that Meteor supposedly has additional 2 LPE cores inside the SOC. Videocardz

 

[moinmoin]

As you can see, If you limit the power to 45W then (...)

That's the power range of Dragon Range (heh) where that table should also make sense since supposedly it's Raphael reused in mobile. So all these comparison are likely to match Dragon Range, assuming the IOD is not further optimized for mobile.

Phoenix Point with mobile optimized cores (may or may not be related to Zen 4c, I expect they will be) is supposed to go into lower power mobile units.

On the matter of RTL we still don't know if any RTL dies will be used in mobile units. Seems like a waste to wait for MTL for these improvements.

 

[TESKATLIPOKA]

That's the power range of Dragon Range (heh) where that table should also make sense since supposedly it's Raphael reused in mobile. So all these comparison are likely to match Dragon Range, assuming the IOD is not further optimized for mobile.

Phoenix Point with mobile optimized cores (may or may not be related to Zen 4c, I expect they will be) is supposed to go into lower power mobile units.

On the matter of RTL we still don't know if any RTL dies will be used in mobile units. Seems like a waste to wait for MTL for these improvements.

I also think Dragon Range will be a reused Raphael.
Dragon range under 45W will be pointless, even Phoenix with half of the cores should perform better at 45W. At 88W It looks like a sweet spot, but 65W is also pretty good.
I think Phoenix will be up to 45W.

Raptor is more efficient at lower TDP thank Alder Lake, so I hope they will release them unless they want to release Meteor Lake instead.
Just the process used for Meteor Lake makes a pretty nice improvement, then there is still the improvement from architecture on top of that, so I think It will perform better than Phoenix.
Not sure how much longer we will have to wait for Meteor Lake.
For some casual gaming you should choose Phoenix, but If Meteor can increase the IGP clocks to ~2.4GHz(+50%) then even with only 128EU(+33%) It won't be so bad.

 

[DisEnchantment]

This looks pretty bad for Zen4 considering It uses better process for both IOD(N6) and CCD(N5).
Of course the chiplet design is worsening the efficiency of Zen4, so a monolithic 8C Phoenix would perform better at 45W in this table, the question is by how much.
Increasing the power for 7700x by 20W(+44%) increased performance by just ~20%, but the chiplet overhead is hardly 20W, so Phoenix at 45W would be <98% in the table unless the supposed N4 process and additional changes makes up the difference.
Let's say Phoenix will be 20% faster at 45W.

7700X at 45W barely beats 6900HS at 35W in CB R23 MT. You would think a 1.5x node jump and 13% IPC would give better score, right? There has to be some more parameters and layers introduced in your ML model.

 

[Henry swagger]

7700X at 45W barely beats 6900HS at 35W in CB R23 MT. You would think a 1.5x node jump and 13% IPC would give better score, right? There has to be some more parameters and layers introduced in your ML model.

Isn't dragon range using n4p ?

 

[TESKATLIPOKA]

7700X at 45W barely beats 6900HS at 35W in CB R23 MT. You would think a 1.5x node jump and 13% IPC would give better score, right? There has to be some more parameters and layers introduced in your ML model.

And you got that from where? I found something entirely else.
7700x at 45W managed 15167 points in CB R23. Link
6900HS at 35W managed 11434 and 12438 points at ~45W in CB R23. Link
Let's not forget 7700x is not monolithic, yet at 45W It is still ~22% faster.

 

[DisEnchantment]

And you got that from where? I found something entirely else.
7700x at 45W managed 15167 points in CB R23. Link
6900HS at 35W managed 11434 and 12438 points at ~45W in CB R23. Link
Let's not forget 7700x is not monolithic, yet at 45W It is still ~22% faster.

Looks like the one I saw is not power constrained. But regardless, you might want to take it slow with the performance charts and tables. Might cause some disappointment and rage when it does not pan out. (like in the GPU threads)

 

[moinmoin]

And you got that from where? I found something entirely else.
7700x at 45W managed 15167 points in CB R23. Link
6900HS at 35W managed 11434 and 12438 points at ~45W in CB R23. Link
Let's not forget 7700x is not monolithic, yet at 45W It is still ~22% faster.

45W is also ~28% more power than 35W, so ~22% more performance is still less efficient.

Using the desktop Ryzen 7000 chips to guess the performance of possible Dragon Range chips may or may not end up matching depending on how close those will be to Raphael.

Phoenix Point most likely will have a significantly different v/f curve. And not only because it's on another node, previous mobile Zen cores also had significantly different v/f curves despite using the same node. So I don't think the existing desktop Ryzen 7000 chips can give us much of a hint for those.

For example this is how much Zen 2's v/f curve differs as part of desktop 3950X vs mobile 4800H (taken from Chips and Cheese's article on ADL's power efficiency):

 

[TESKATLIPOKA]

Looks like the one I saw is not power constrained. But regardless, you might want to take it slow with the performance charts and tables. Might cause some disappointment and rage when it does not pan out. (like in the GPU threads)

What I wrote about Meteor Lake is only my speculation, and I didn't even speculate about the final performance, just applied the supposed process advantage to Raptor Lake.
I won't be disappointed, there is still AMD.
As for RDNA3, I really believed the clocks will be a lot higher for N31, didn't happen, and I was angry and disappointed. If nothing else, AMD really could have gone and made a bit bigger GCD(64WGPs). Maybe with RDNA4 they will be more daring.

 

[TESKATLIPOKA]

45W is also ~28% more power than 35W, so ~22% more performance is still less efficient.

Using the desktop Ryzen 7000 chips to guess the performance of possible Dragon Range chips may or may not end up matching depending on how close those will be to Raphael.

Phoenix Point most likely will have a significantly different v/f curve. And not only because it's on another node, previous mobile Zen cores also had significantly different v/f curves despite using the same node. So I don't think the existing desktop Ryzen 7000 chips can give us much of a hint for those.

For example this is how much Zen 2's v/f curve differs as part of desktop 3950X vs 4800H (taken from Chips and Cheese's article on ADL's power efficiency):

22% was at the same TDP and I also mentioned 7700x is not monolithic unlike 6900Hs.
What I wrote is still only my speculation. We will see if Raphael = Dragon Ridge or there will be some improvements made.
Phoenix will mostly gain additional efficiency due to being a monolithic design compared to dragon ridge.

 

[leoneazzurro]

Why thinking Dragon Range will be limited to 8 cores? There were rumors of it going up to 16 cores and 65W. It is clearly a solution aimed at the high-end gaming and workstation market.
If that is the case, the efficiency discussion will change completely, not even thinking that applying desktop results to mobile binning may be quite misleading.
Not that for the high-end mobile market efficiency is a big selling point, anyway.

 

[TimCh]

Why thinking Dragon Range will be limited to 8 cores? There were rumors of it going up to 16 cores and 65W. It is clearly a solution aimed at the high-end gaming and workstation market.
If that is the case, the efficiency discussion will change completely, not even thinking that applying desktop results to mobile binning may be quite misleading.
Not that for the high-end mobile market efficiency is a big selling point, anyway.

According to AMD Dragon Range has the “Highest core, thread and cache ever for a mobile gaming CPU” so it will definitely not be limiting to 8 cores.

 

[TESKATLIPOKA]

Why thinking Dragon Range will be limited to 8 cores? There were rumors of it going up to 16 cores and 65W. It is clearly a solution aimed at the high-end gaming and workstation market.
If that is the case, the efficiency discussion will change completely, not even thinking that applying desktop results to mobile binning may be quite misleading.
Not that for the high-end mobile market efficiency is a big selling point, anyway.

Who says Dragon Range will be only 8C16T? Phoenix will be limited to 8 cores, Dragon will have more cores or there would be no reason to release It.

You are right about binning, but we don't have anything better to compare. Let's say that desktop results are the worst possible outcome for the mobile variant.

 

[leoneazzurro]

The reason I thought people was saying Dragon was limited to 8 cores is the pointing continuously to the 7700X as a reference point. Which is not clearly the best from an efficiency point of view due to the thread deficit compared to Intel's lineup. Better comparison should be those where the thread number is more similar, that is, 7900X and 7950X.

 

[inf64]

Dragon Range should be 16C/32T as per AMD's slides, but who knows. They still can go mental with Zen 4c and cram 32C in the same mobile package, we can dream at least

 

[TESKATLIPOKA]

The reason I thought people was saying Dragon was limited to 8 cores is the pointing continuously to the 7700X as a reference point. Which is not clearly the best from an efficiency point of view due to the thread deficit compared to Intel's lineup. Better comparison should be those where the thread number is more similar, that is, 7900X and 7950X.

I was using 7700x as a reference for Phoenix not for Dragon Ridge.
If I wanted a reference point for Dragon Ridge then I would naturally use 7900x or 7950x at 45 or 65W, which were also present in the table.

 

[leoneazzurro]

Yeah but that comparison is tricky. Apart the rumored N4 usage, core should be identical but way different optimizations, i.e. RAM controller is different in more than one aspect (LPDDR5 for an instance), power/frequency curve will be also not the same. Yes, we could expect performance per clock to be similar (except cache dependant workloads) to the 7700-7700X but power draw should be lower similarly at what we saw with Rembrandt compared to Vermeer.

 

[eek2121]

Not sold on a V-Cache mobile part (unless AMD announced it explicitly and I missed it)

Apparently, according to a leak, desktop is getting a 6-core and 8-core X3D part. No 12-16 core parts.

Also note that AMD doesn’t need to use V-Cache to claim they have the “Highest core, thread, and cache ever for a mobile gaming CPU”.

They simply need the desktop chips with the normal amount of L3. Previous mobile chips had half the L3.

The 7950X has 64mb of L3 Cache, which was more than Alder Lake. It also technically had both the highest core count and thread count until Raptor Lake.

Basically, a mobile 7950X would make that statement acccurate.

 

[moinmoin]

The selling point for V-Cache in mobile parts would that it increases power efficiency of the CPU as well as allows using slower but more power efficient memory without significantly hurting performance.

So V-Cache is a card AMD could play if it needs to push publicity for highest possible performance in a mobile form factor with great battery life. It may or may not feel the need to play such a card.

 

[MadRat]

It would be logical to follow in the footsteps of Intel's EMIB strategy with their own. Looks about as optimal of layout as one could go for graphics performance. Could a spiralized layout shorten connections I wonder, by capitalizing on PCB layers. A straight line looks good until I think about cooling. Would love to see AMD simplify to one monolithic cache rather than use two, and God forbid go for three.

 

[lopri]

That computerbase.de chart is so nonsensical. At 45W, 7700X, 7900X, 7950X performs basically identical, yet 7600X somehow falls behind?

 

[TheELF]

That computerbase.de chart is so nonsensical. At 45W, 7700X, 7900X, 7950X performs basically identical, yet 7600X somehow falls behind?

It is the lowest tier though, so why do you think it's nonsensical?
That's what binning is all about, sort out the ones that need more power for a certain performance target.

 

[BorisTheBlade82]

That computerbase.de chart is so nonsensical. At 45W, 7700X, 7900X, 7950X performs basically identical, yet 7600X somehow falls behind?

At 45w the 7900X and 7950X get pretty much starved for power because of the Chiplet-Penalty. But 7700X is not in this state and able to profit from 33% more cores in these multithreaded Tests.
More cores at ISO-Wattage almost always means more performance because of the nature of the V/F-curve. At 65w you can observe the expected scaling.

 

[TESKATLIPOKA]

That computerbase.de chart is so nonsensical. At 45W, 7700X, 7900X, 7950X performs basically identical, yet 7600X somehow falls behind?

There is no problem with that table or charts.
That's an average score from 10 MT tests they used. They have charts for those too, you just need to select them Link.

Here is just Cinebench R23:





























7600X is a 6C12T, so It's not surprising 8C16T 7700x 45W is faster, but only by 19%, which means 7600x can clock higher within the limited budget in CB R23 because It needs to feed less cores.
7900X and 7950X don't perform anywhere near what they should, but I can only attribute It to chiplet design.
A shame they didn't measure the frequency in R23 for example, It would be very helpful in this case.