Discussion Anyone else bored out of their mind due to mainstream CPU market stagnation?

[igor_kavinski]

That monster would surely have a monsterous TDP and MSRP.

I think they must be really scared of Zen 6 and know something from leaky OEMs that we don't. Maybe 24C/48T dual CCD AMD monster incoming with double the fabric speed and bandwidth.

 

[Hulk]

That monster would surely have a monsterous TDP and MSRP.

Good point but I"m hoping MSRP will have to be competitive with AMD. This ain't 15 years ago when Intel was out in front. Despite what Pat G. said, AMD is not in the rearview.

As for TDP, possibly. But it's more the P's that burn the house down and create the high power draw. 40 sanely clocked E's could provide tremendously efficient MT compute.

As always though, the trade off is die area, which of course scales with cost. If they have the process to make it small enough using in-house (ie less expense than going to TSMC) fabs then they can conceivably leverage their foundry for the first time in a long while against AMD.

It's all speculation right now as I take these leaks as seriously as I take my own speculation, meaning not seriously at all, its just points for our entertaining discussion.

 

[Thunder 57]

Good point but I"m hoping MSRP will have to be competitive with AMD. This ain't 15 years ago when Intel was out in front. Despite what Pat G. said, AMD is not in the rearview.

As for TDP, possibly. But it's more the P's that burn the house down and create the high power draw. 40 sanely clocked E's could provide tremendously efficient MT compute.

As always though, the trade off is die area, which of course scales with cost. If they have the process to make it small enough using in-house (ie less expense than going to TSMC) fabs then they can conceivably leverage their foundry for the first time in a long while against AMD.

It's all speculation right now as I take these leaks as seriously as I take my own speculation, meaning not seriously at all, its just points for our entertaining discussion.

They can barley keep 8 P's under control how are they going to do 16? Plot twist, they are bring back the Extreme Edition CPU and it will get that config . The lower versions should do quite well though.

 

[Hulk]

They can barley keep 8 P's under control how are they going to do 16? Plot twist, they are bring back the Extreme Edition CPU and it will get that config . The lower versions should do quite well though.

I hear you but I'm seeing hot spots/heat transfer being more problematic than total heat load. There are affordable solutions to move 300W out of the case, like a 360 AIO, if you can get the heat into the radiator. That's the real issue. Also, who knows if there will be 16 P's, could be a sh!t ton of E's and 8 P's? Finally, Intel 18A might be good.

So, I could see something like 12+40 being viable with the following caveats:
1. 18A is very efficient
2. The E's are low clocked
3. The P's are spread out between the E's for thermal dispersion
4. TIM and the lid are improved for better heat transfer

Yeah, it's a big nut to crack, especially for Intel as things haven't been going swimmingly for them over the last 10 years.

Coming back to reality means we'll most likely see 8+24 or 8+32. If they can improve the gaming on LC and increase IPC on both LC and Skymont along with 18A being a good node, that would be a good part. AMD would obvioulsy counter with 10950X at 24/48. Either way we win!

 

[Hulk]

I think they must be really scared of Zen 6 and know something from leaky OEMs that we don't. Maybe 24C/48T dual CCD AMD monster incoming with double the fabric speed and bandwidth.

The scary thing for Intel is that AMD wouldn't need heroic efforts to create such a part. Zen 6 is obviously well into it's development and if the next TMSC node can economically hold 12 Zen 6 cores per chiplet then they are "there." This is basically little more than getting to first base, maybe second.

Intel on the otherhand needs to work out the spotty performance with Arrow Lake, get 18A efficient and transistor dense, and make sure the memory subsystem can handle all of those cores. This would be like them hitting a home run. Imagine 8+32, 10% better ST performance over Raptor Lake, great MT performance, and good efficiency. It's quite a task for the "new" Intel. The old Intel? Yeah, they could have made it happen, but the old Intel should never have turned into the new Intel so it is a different Intel. Hey, I'm sounding like Zathros here. That's some deep sci-fi humor

 

[OneEng2]

The scary thing for Intel is that AMD wouldn't need heroic efforts to create such a part. Zen 6 is obviously well into it's development and if the next TMSC node can economically hold 12 Zen 6 cores per chiplet then they are "there." This is basically little more than getting to first base, maybe second.

Intel on the otherhand needs to work out the spotty performance with Arrow Lake, get 18A efficient and transistor dense, and make sure the memory subsystem can handle all of those cores. This would be like them hitting a home run. Imagine 8+32, 10% better ST performance over Raptor Lake, great MT performance, and good efficiency. It's quite a task for the "new" Intel. The old Intel? Yeah, they could have made it happen, but the old Intel should never have turned into the new Intel so it is a different Intel. Hey, I'm sounding like Zathros here. That's some deep sci-fi humor

True! Zen 6 isn't revolutionary at all.

N3P should be no issue as it is just an enhanced TSMC N3 node that AMD Zen 5 is already produced on (N3E) for server.

Zen 6 is not a radical departure from Zen 5 either AFAIK. My biggest hope for Zen 6 is that the infinity fabric gets a big upgrade (not core related) and that it is fed with faster, lower latency cache and memory. Still, I am not expecting more than 15% per clock performance increase .... and maybe more like 10% over Zen 5. Perhaps an added 5% clock speed for a total of maybe 15-20% performance lift per core?

In MT, Zen 6 should be a beast though. 24c/48T should rip through MT tasks .... but the question is, how will the MT fare against a potential 48c/48t Nova Lake? Sure, those darkmont cores aren't near a match for a full Zen 6 core, but do a little math ..... currently a Zen 5c is ~1.5 Skymont cores in MT. Now, that sounds pretty good BUT 24 x 1.5 doesn't add up to 48 .

In server for Zen 6 vs Clearwater forest things look a bit different. AMD is doubling the number of cores (and moving to N2 for their EPYC part) giving each CCD 32 cores. If AMD sticks with their current 12 CCD's in the Zen 6 "Turin D" part that would be 384 Zen 6c cores all having SMT and a 512b wide AVX512 unit vs Clearwater forest with 288 Darkmont cores. I can tell you where my money would be on that race .

So it is entirely possible that AMD will be beaten badly in desktop MT applications by Nova Lake, but will continue to pulverize Intel in DC with Zen 6.

I also wonder how well Coyote Cove will fare against Zen 6 in ST performance? My assumption is that Intel will have nothing in Nova Lake that can compete with AMD's X3D parts in gaming.

Things should start getting interesting in 2026. I think we are in for a long, dry 2025 though

 

[Hulk]

True! Zen 6 isn't revolutionary at all.

N3P should be no issue as it is just an enhanced TSMC N3 node that AMD Zen 5 is already produced on (N3E) for server.

Zen 6 is not a radical departure from Zen 5 either AFAIK. My biggest hope for Zen 6 is that the infinity fabric gets a big upgrade (not core related) and that it is fed with faster, lower latency cache and memory. Still, I am not expecting more than 15% per clock performance increase .... and maybe more like 10% over Zen 5. Perhaps an added 5% clock speed for a total of maybe 15-20% performance lift per core?

In MT, Zen 6 should be a beast though. 24c/48T should rip through MT tasks .... but the question is, how will the MT fare against a potential 48c/48t Nova Lake? Sure, those darkmont cores aren't near a match for a full Zen 6 core, but do a little math ..... currently a Zen 5c is ~1.5 Skymont cores in MT. Now, that sounds pretty good BUT 24 x 1.5 doesn't add up to 48 .

In server for Zen 6 vs Clearwater forest things look a bit different. AMD is doubling the number of cores (and moving to N2 for their EPYC part) giving each CCD 32 cores. If AMD sticks with their current 12 CCD's in the Zen 6 "Turin D" part that would be 384 Zen 6c cores all having SMT and a 512b wide AVX512 unit vs Clearwater forest with 288 Darkmont cores. I can tell you where my money would be on that race .

So it is entirely possible that AMD will be beaten badly in desktop MT applications by Nova Lake, but will continue to pulverize Intel in DC with Zen 6.

I also wonder how well Coyote Cove will fare against Zen 6 in ST performance? My assumption is that Intel will have nothing in Nova Lake that can compete with AMD's X3D parts in gaming.

Things should start getting interesting in 2026. I think we are in for a long, dry 2025 though

AMD would need 32/64 if 48 core Nova Lake is actually going to exist, which I doubt very much.

Intel is in "response" mode to Intel right now. AMD puts down their hand and "calls" Intel. AMD put down Zen 5 and Intel countered with Arrow Lake. AMD will get Zen 6 out before whatever comes next for Intel. If AMD needs to add another chiplet they can make that happen pretty quickly.

Also, and this is kind of interesting, as many around here have noted AMD and Intel know what the other is doing like years before a release. Why would I believe this? Zen 4 and Raptor Lake were just too close in performance for coincidence, same thing for Zen 5 and Arrow Lake.

It reminds me of the Pottsdam meeting in 1945. Truman tell Stalin that we have a new very powerful weapon. Stalin replies, "I hope you will put it to good use." Historians have noted that Stalin most likely knew more about our atomic bomb than Truman did!!!

 

[DrMrLordX]

They can barley keep 8 P's under control how are they going to do 16? Plot twist, they are bring back the Extreme Edition CPU and it will get that config . The lower versions should do quite well though.

The interconnect would be more of a problem. Intel has a really bad track record on that lately. If they could somehow overcome intercore/intertile latency problems, it wouldn't be too hard for them to arrange clocks in a fashion similar to a cell phone SoC. Have 1-4 P cores with high boost clocks, and then gimp the clocks on everything else, with per-core voltage control.

 

[OneEng2]

AMD would need 32/64 if 48 core Nova Lake is actually going to exist, which I doubt very much.

Intel is in "response" mode to Intel right now. AMD puts down their hand and "calls" Intel. AMD put down Zen 5 and Intel countered with Arrow Lake. AMD will get Zen 6 out before whatever comes next for Intel. If AMD needs to add another chiplet they can make that happen pretty quickly.

Also, and this is kind of interesting, as many around here have noted AMD and Intel know what the other is doing like years before a release. Why would I believe this? Zen 4 and Raptor Lake were just too close in performance for coincidence, same thing for Zen 5 and Arrow Lake.

It reminds me of the Pottsdam meeting in 1945. Truman tell Stalin that we have a new very powerful weapon. Stalin replies, "I hope you will put it to good use." Historians have noted that Stalin most likely knew more about our atomic bomb than Truman did!!!

Possibly, but it seems like AMD is much more concerned than Intel on selling chips at a profit.

I question the value of a 48 core chip for anything other than HPC ... and here AMD has threadripper... Most certain to best 48 non SMT cores lacking AVX512.

In the past it was AMD that got out over its skis a bit with the "more cores" approach. Intel remained dominant with single thread advantage.

Seems like all these years later, history my be repeating itself in reverse 😁

Panther Lake does look compelling in laptop though.

 

[Hulk]

Possibly, but it seems like AMD is much more concerned than Intel on selling chips at a profit.

I question the value of a 48 core chip for anything other than HPC ... and here AMD has threadripper... Most certain to best 48 non SMT cores lacking AVX512.

In the past it was AMD that got out over its skis a bit with the "more cores" approach. Intel remained dominant with single thread advantage.

Seems like all these years later, history my be repeating itself in reverse 😁

Panther Lake does look compelling in laptop though.

I think both AMD and Intel are very concerned with selling chips at a profit but AMD currently, for a variety of reason, is better at it.

As to your second point regarding HPC, there is a big price difference between low end HPC and high end enthusiast. I would probably pay $800 for a 32 core Zen 5 part, but I am not paying $2,200 for a 32 core Threadripper on Zen 4. Plus you must consider the added cost of motherboard and ECC memory and lack.

Do I need HPC power? No. Do I and other enthusiasts want it? Yes and we are always wanting more from AMD and Intel at the top of the stack. They of course have to "protect" their HPC parts, while trying to hold onto the enthusiasts and the "glamor" that having the fastest part offers the brand.

Now back to reality. 90% of the time I'm browsing the web, on YouTube, working in Office, CorelDraw, or PS. Any low to midrange current CPU can handle those tasks adequately. It's the 5 or 10% of the time when I am encoding a video, previewing the timeline in my NLE, or batch processing RAW photos when I want more compute. It's not necessary, I can get my work done without it but it makes work like this more efficient and more fun.

Unfortunately for Intel at the moment game performance drives the high end of processors more than any other application for non HPC buyers. Even though only 8 cores are really necessary for gaming, those core types are the ones in the higher core count processors, which in turn makes those processors more performant. My point is that gaming performance is and has been driving desktop performance and all of the other benefits of processors that game will are somewhat of a "side benefit" of good gaming cores.

I've written myself a bit OT here thinking about your post, which is a good one. This discussion brings a good point to light.

What factors determine how much performance AMD or Intel will offer for the desktop before moving into the HPC realm?
I think the main points for consideration would be as follows.

1. How close to HPC performance can we come with our flagship desktop CPU's before we start to cut into HPC CPU sales? If Zen 6 goes 24 cores for say $650 then who is going to buy 32 cores of Zen 5 for $2200? So it's not so much how many cores CAN we offer on the desktop but how many SHOULD we offer? I realize there are other benefits for HPC beyond core count.

2. How important is it for us to have the fastest desktop processor for our brand? If we lose the crown or are permanently looked at as being "2nd best" how much is that going to hurt sales up and down the stack?

 

[OneEng2]

I think both AMD and Intel are very concerned with selling chips at a profit but AMD currently, for a variety of reason, is better at it.

As to your second point regarding HPC, there is a big price difference between low end HPC and high end enthusiast. I would probably pay $800 for a 32 core Zen 5 part, but I am not paying $2,200 for a 32 core Threadripper on Zen 4. Plus you must consider the added cost of motherboard and ECC memory and lack.

Do I need HPC power? No. Do I and other enthusiasts want it? Yes and we are always wanting more from AMD and Intel at the top of the stack. They of course have to "protect" their HPC parts, while trying to hold onto the enthusiasts and the "glamor" that having the fastest part offers the brand.

Now back to reality. 90% of the time I'm browsing the web, on YouTube, working in Office, CorelDraw, or PS. Any low to midrange current CPU can handle those tasks adequately. It's the 5 or 10% of the time when I am encoding a video, previewing the timeline in my NLE, or batch processing RAW photos when I want more compute. It's not necessary, I can get my work done without it but it makes work like this more efficient and more fun.

Unfortunately for Intel at the moment game performance drives the high end of processors more than any other application for non HPC buyers. Even though only 8 cores are really necessary for gaming, those core types are the ones in the higher core count processors, which in turn makes those processors more performant. My point is that gaming performance is and has been driving desktop performance and all of the other benefits of processors that game will are somewhat of a "side benefit" of good gaming cores.

I've written myself a bit OT here thinking about your post, which is a good one. This discussion brings a good point to light.

What factors determine how much performance AMD or Intel will offer for the desktop before moving into the HPC realm?
I think the main points for consideration would be as follows.

1. How close to HPC performance can we come with our flagship desktop CPU's before we start to cut into HPC CPU sales? If Zen 6 goes 24 cores for say $650 then who is going to buy 32 cores of Zen 5 for $2200? So it's not so much how many cores CAN we offer on the desktop but how many SHOULD we offer? I realize there are other benefits for HPC beyond core count.

2. How important is it for us to have the fastest desktop processor for our brand? If we lose the crown or are permanently looked at as being "2nd best" how much is that going to hurt sales up and down the stack?

Excellent commentary and points.

AMD is currently making big inroads in DC at Intel's expense and making hefty margins there. AMD is also making advances in desktop and laptop .... but desktop is a shrinking market .... and Intel has pretty compelling products in Laptop IMO to fend them off with (unlike DC).

Still the main point remains, from a product line management standpoint, while AMD "COULD" produce a 64c/128t, 3 tile desktop chip for Zen 6, would it make any sense? It would require the compute tile to be on a very expensive N2 process, and would pretty much eliminate the need for Threadripper. Of course, I could be oversimplifying here as I am not sure you can feed 64c without more channels of memory than desktop/laptop can provide cost effectively.

It is so easy to slip off into the "can it be done" and forget all about the "is it a good idea financially to do it".

Your last point is also hard to put your finger on. How important is it to be seen as the "leader"? How much does it cost you NOT to be seen this way?

I am certain that these are questions pondered at the highest levels of AMD and Intel (and pretty much any CPU maker I would think).

 

[Hulk]

Excellent commentary and points.

AMD is currently making big inroads in DC at Intel's expense and making hefty margins there. AMD is also making advances in desktop and laptop .... but desktop is a shrinking market .... and Intel has pretty compelling products in Laptop IMO to fend them off with (unlike DC).

Still the main point remains, from a product line management standpoint, while AMD "COULD" produce a 64c/128t, 3 tile desktop chip for Zen 6, would it make any sense? It would require the compute tile to be on a very expensive N2 process, and would pretty much eliminate the need for Threadripper. Of course, I could be oversimplifying here as I am not sure you can feed 64c without more channels of memory than desktop/laptop can provide cost effectively.

It is so easy to slip off into the "can it be done" and forget all about the "is it a good idea financially to do it".

Your last point is also hard to put your finger on. How important is it to be seen as the "leader"? How much does it cost you NOT to be seen this way?

I am certain that these are questions pondered at the highest levels of AMD and Intel (and pretty much any CPU maker I would think).

The desire to "be on top" does seem to be quite strong still and it is good for the consumer. If AMD "broken the rules of the game" by offering 16 core parts at relatively affordable prices for the desktop in order to get past Intel then Intel would not have responded with 8+16.

Now that AMD had ramped up Zen 5 IPC and Intel has lost a bit by forgoing HT and monolithic Intel again might need to respond with more E cores, which of course would force AMD's hand to do the same.

I don't know what the area factor is between Zen 5 and 5c but AMD should be thinking very hard about something like 12+20. That would put Intel in a very tough position. My reasoning is that any application that is going to benefit from more than 12 cores is going to benefit more from 20 Zen 5c cores rather than an additional 12 Zen 5 cores. If the cost in die area is the same then there is no reason to go Zen 5 only.

 

[OneEng2]

The desire to "be on top" does seem to be quite strong still and it is good for the consumer. If AMD "broken the rules of the game" by offering 16 core parts at relatively affordable prices for the desktop in order to get past Intel then Intel would not have responded with 8+16.

Now that AMD had ramped up Zen 5 IPC and Intel has lost a bit by forgoing HT and monolithic Intel again might need to respond with more E cores, which of course would force AMD's hand to do the same.

I don't know what the area factor is between Zen 5 and 5c but AMD should be thinking very hard about something like 12+20. That would put Intel in a very tough position. My reasoning is that any application that is going to benefit from more than 12 cores is going to benefit more from 20 Zen 5c cores rather than an additional 12 Zen 5 cores. If the cost in die area is the same then there is no reason to go Zen 5 only.

It will be interesting to see how the two different approaches between Intel and AMD iron out. AMD is maintaining E cores that are fully instruction compatible to the P cores while Intel's E cores are a totally different architecture than their P cores. Intel appears to be able to pack a ton of those little E cores into a design, but I am thinking that where it really counts (data center), AMD's approach may still win the day. We will see.

Zen 5c on N3E is 1.7mm2 compared to Skymont at 1.15mm2 on N3B (which is a slightly more dense process than N3E). To me, it appears that AMD's approach yields a good deal better results for performance per area as Zen 5c still enjoys AVX512 (full) and SMT. Still, its close when you take into account 1 Zen 5c in MT is ~ 1.5 Skymont.

1.15 * 1.5 = 1.73. Pretty close result for two such disparate approaches.

 

[Hulk]

It will be interesting to see how the two different approaches between Intel and AMD iron out. AMD is maintaining E cores that are fully instruction compatible to the P cores while Intel's E cores are a totally different architecture than their P cores. Intel appears to be able to pack a ton of those little E cores into a design, but I am thinking that where it really counts (data center), AMD's approach may still win the day. We will see.

Zen 5c on N3E is 1.7mm2 compared to Skymont at 1.15mm2 on N3B (which is a slightly more dense process than N3E). To me, it appears that AMD's approach yields a good deal better results for performance per area as Zen 5c still enjoys AVX512 (full) and SMT. Still, its close when you take into account 1 Zen 5c in MT is ~ 1.5 Skymont.

1.15 * 1.5 = 1.73. Pretty close result for two such disparate approaches.

I think Zen 5c has lower clocks than Skymont, right?

 

[GTracing]

I think Zen 5c has lower clocks than Skymont, right?

Yes, that density comes at a cost. Zen 5c max boost is 3.3GHz in Strix point, and 3.7GHz in Turin dense.

A bigger issue, in my opinion, is the comparison leaves out the area used by the L2 and L3 cache. Zen 5 has 4MB L3 per core, Zen 5c in Turin dense has 2MB L3 per core, and Skymont in Arrow Lake has an abysmal 0.66MB L3 per core. Spitballing it, I think the size comparison would be around to 2:1 if you include cache.

 

[DrMrLordX]

I don't know what the area factor is between Zen 5 and 5c but AMD should be thinking very hard about something like 12+20.

Intel's problem is that their Coves are not terribly area or power efficient. Zen5c has area efficiency on its side, but power efficiency? Maybe not so much? A standard Zen5 core can be run at a lower clockspeed to achieve similar effect without introducing the constraints imposed upon Zen5c.

In other words, if Intel goes wide enough that AMD starts losing productivity benchmarks left and right, they can just at more of their mainline cores, which is probably what they'll do with Zen6 anyway.

 

[Hulk]

Intel's problem is that their Coves are not terribly area or power efficient. Zen5c has area efficiency on its side, but power efficiency? Maybe not so much? A standard Zen5 core can be run at a lower clockspeed to achieve similar effect without introducing the constraints imposed upon Zen5c.

In other words, if Intel goes wide enough that AMD starts losing productivity benchmarks left and right, they can just at more of their mainline cores, which is probably what they'll do with Zen6 anyway.

We need good numbers to actually analyze this in a meaningful way. I'll try. Assuming Skymont clocks to 4.6 and 5c to 3.7 then Skymont has a 24% advantage.
Using "throughput" numbers from our Handbrake test, this would reduce the IPC advantage for 5c to about 25%.
If 5 Skymonts can fit in the area of one 5c then when it comes to encoding x265 we have parity. But if more than 5 Skymonts can fit in the space of four 5c's then Skymont will have a MT advantage.

Despite most people losing faith in Intel's E's I think Skymont is still the dark horse for Intel. I have owned a 13900K and 14900K and those 16 Gracemont cores were very good at keeping things moving during heavy multitasking. Skymont is about 50% better IPC-wise and clock higher. They are a not-so-secret weapon for Intel.

Chatgpt (yeah I know) is telling me Skymont has a 2:1 area advantage over 5c. If that is true or even close to true Skymont is considerably more area efficient than Zen 5c.

I re-realized how powerful Skymont is I saw Igor's Skymont only results for our Handbrake test. They contribute more to the total score than the P's and he is running 6+8. Intel's E's have become a real weapon. If not for the gaming "glitch" with Lion Cove performance Intel could claim world's fastest CPU.

I think I am being fair in my analysis having owned both a 9950X and many Raptor Lake CPU's. AMD and Intel are still neck-and-neck outside of gaming, which I realize IS a big deal in today's marketplace! Honestly, outside of competitive gaming, framerates are so high with high end CPU's I'm not even sure the CPU matters that much, but I'm not a gamer. I remember when 40fps was a decent frame rate and now I see people argueing how 210fps so "so much better" than 175fps. Both seem wildly beyond human perception to me.

 

[OneEng2]

We need good numbers to actually analyze this in a meaningful way. I'll try. Assuming Skymont clocks to 4.6 and 5c to 3.7 then Skymont has a 24% advantage.
Using "throughput" numbers from our Handbrake test, this would reduce the IPC advantage for 5c to about 25%.
If 5 Skymonts can fit in the area of one 5c then when it comes to encoding x265 we have parity. But if more than 5 Skymonts can fit in the space of four 5c's then Skymont will have a MT advantage.

Despite most people losing faith in Intel's E's I think Skymont is still the dark horse for Intel. I have owned a 13900K and 14900K and those 16 Gracemont cores were very good at keeping things moving during heavy multitasking. Skymont is about 50% better IPC-wise and clock higher. They are a not-so-secret weapon for Intel.

Chatgpt (yeah I know) is telling me Skymont has a 2:1 area advantage over 5c. If that is true or even close to true Skymont is considerably more area efficient than Zen 5c.

I re-realized how powerful Skymont is I saw Igor's Skymont only results for our Handbrake test. They contribute more to the total score than the P's and he is running 6+8. Intel's E's have become a real weapon. If not for the gaming "glitch" with Lion Cove performance Intel could claim world's fastest CPU.

I think I am being fair in my analysis having owned both a 9950X and many Raptor Lake CPU's. AMD and Intel are still neck-and-neck outside of gaming, which I realize IS a big deal in today's marketplace! Honestly, outside of competitive gaming, framerates are so high with high end CPU's I'm not even sure the CPU matters that much, but I'm not a gamer. I remember when 40fps was a decent frame rate and now I see people argueing how 210fps so "so much better" than 175fps. Both seem wildly beyond human perception to me.

I believe this is a more compelling analysis than SPECINT (at least it is to me).

The one thing I would like to point out is that Handbrake is likely more bandwidth constrained per core than it is compute constrained .... so those Skymont cores don't have too much being asked of them other than to move memory in, tweak it, and move memory out.

That isn't to say that this isn't important. There are likely many other loads that are like that .... but when it comes to more heavy computation and branching, I suspect Zen 5c will excel. Of course, I could be wrong.

I wonder if the product mix is just different for different workloads and markets. For something like Handbrake, lots of space efficient cores capable of moving lots of information in and out of main memory is the most important.

For gaming, a core with high IPC, clock rate, and low latency to memory is most important.

For workstation, many cores of high IPC and clock rate.

I wonder about DC. If Darkmont doesn't need to compete in IPC with Zen 6, but rather is more like Handbrake, then massive core count and memory bandwidth will win the day there as well.

In desktop, the die size is hyper critical as high volume and lower pricing force performance compromises to be made based on what can be produced at a profit. Perhaps Intel's approach of a few big fat powerful cores combined with a metric crap ton of lower performance, but space efficient and power efficient cores will win the day?

 

[Hulk]

I believe this is a more compelling analysis than SPECINT (at least it is to me).

The one thing I would like to point out is that Handbrake is likely more bandwidth constrained per core than it is compute constrained .... so those Skymont cores don't have too much being asked of them other than to move memory in, tweak it, and move memory out.

That isn't to say that this isn't important. There are likely many other loads that are like that .... but when it comes to more heavy computation and branching, I suspect Zen 5c will excel. Of course, I could be wrong.

I wonder if the product mix is just different for different workloads and markets. For something like Handbrake, lots of space efficient cores capable of moving lots of information in and out of main memory is the most important.

For gaming, a core with high IPC, clock rate, and low latency to memory is most important.

For workstation, many cores of high IPC and clock rate.

I wonder about DC. If Darkmont doesn't need to compete in IPC with Zen 6, but rather is more like Handbrake, then massive core count and memory bandwidth will win the day there as well.

In desktop, the die size is hyper critical as high volume and lower pricing force performance compromises to be made based on what can be produced at a profit. Perhaps Intel's approach of a few big fat powerful cores combined with a metric crap ton of lower performance, but space efficient and power efficient cores will win the day?

Good analysis and I agree regarding the relevance of Handbrake. At the end of the day, when we are talking about MT performance on the desktop I think we need to focus on the applications on the desktop where the additional MT performance is actually beneficial. Handbrake or media encoding in general is one example.

 

[DrMrLordX]

We need good numbers to actually analyze this in a meaningful way.

While that's interesting, my point is that AMD will probably just not use Zen5c cores on consumer desktop at all, and may not use the Zen6/Medusa equivalent in that role either, since they can just use more Zen6 cores if they feel the need to go wide(r) and clock them accordingly to easily stay within TDP targets. So while you may have an Intel CPU with two separate compute tiles for 16P + whatever e (20? 40?) configs, AMD will likely be content to keep the consumer market at 24c Zen6 and leave Zen6c (if that's what they do next gen!) on mobile/handheld SoCs. Depending on the size of the CCDs and everything else on Medusa, it may even be feasible for AMD to go with a 3 CCD setup for 36c Zen6 if they feel the need to do so.

But at that point will the memory interface keep up? The answer is very likely "no". It probably will struggle with 24c Zen6!

 

[Hulk]

While that's interesting, my point is that AMD will probably just not use Zen5c cores on consumer desktop at all, and may not use the Zen6/Medusa equivalent in that role either, since they can just use more Zen6 cores if they feel the need to go wide(r) and clock them accordingly to easily stay within TDP targets. So while you may have an Intel CPU with two separate compute tiles for 16P + whatever e (20? 40?) configs, AMD will likely be content to keep the consumer market at 24c Zen6 and leave Zen6c (if that's what they do next gen!) on mobile/handheld SoCs. Depending on the size of the CCDs and everything else on Medusa, it may even be feasible for AMD to go with a 3 CCD setup for 36c Zen6 if they feel the need to do so.

But at that point will the memory interface keep up? The answer is very likely "no". It probably will struggle with 24c Zen6!

Yes, I agree.

I think there is a reason we don't see Zen 5c on the desktop. I don't think 5c saves enough area over 5 when you factor in the frequency reduction to make it useful for the desktop. So, while you may be able to fit more 5c cores into a given area than 5, when you factor in the frequency reduction for 5c there is no gain in going with 5c. In addition, if you have an appliction that uses say 12 cores, then you are losing performance by adding in those 4 lower clocked 5c cores to the mix.

For mobile it's a different dynamic. 4 or 6 Zen 5 cores for ST performance consumes most of the power envelope so whatever power is left is best spent efficiently on lower clocked cores. If you are going to clock at 3.7GHz there is no point in using Zen 5 or Zen 5c.

Unlike Lion Cove/Skymont the area scaling for Zen 5/5c just isn't there to make hybrid beneficial for the desktop. A Skymont core is about 1/3 the size of Lion Cove. 5c on the otherhand is 3/4 the size of Zen 5, which is about the difference in clock speed. Hence, 5c is there for power efficiency, no area efficiency for MT compute as is the case with Intel.

That's my theory.