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

[igor_kavinski]

Just wanted to add to this thread that Intel stagnated the market even more with their Raptor Lake degradation fiasco. Really hope the only thing left for them to salvage this situation is to lower the prices on their workstation class CPUs with AVX-512 and AMX, just to prevent enthusiasts from turning to Ryzens.

 

[Nothingness]

Just wanted to add to this thread that Intel stagnated the market even more with their Raptor Lake degradation fiasco. Really hope the only thing left for them to salvage this situation is to lower the prices on their workstation class CPUs with AVX-512 and AMX, just to prevent enthusiasts from turning to Ryzens.

I thought Xeon were not plagued by the issue.

 

[mostwanted002]

How many of you find this fascinating?

Installing update:

Downloading update:

No need to run Cinememe to watch threads in action.

It was the hard way of multiple installs that I found out that consumer windows aren't the way to go with CPUs like EPYC. It (windows) automatically downloaded incompatible PSP drivers, which caused BSODs and boot loops. Had to run a server Evaluation copy to run some benchmarks. Now permanently moved to Ubuntu Server LTS

 

[MadRat]

Microsoft and Intel in many ways got taken over by foreign interests. The world population out numbers the Americans by 30:1. The heyday of new developments were Americans building for their own interests and letting the rest of the world subsidize their own needs. Some of the best American companies found niches paid better and abandoned the rat race for home desktops. We allowed Intel to monopolize markets to squeeze out the rest of what was left. And that wasn't enough. Now they churn out incremental upgrades because they can. And because they want foreign sales, true effort is focused on that market because American consumers are chump change. The EU, China, India, etc. all seem to push for turd implementations. They do not send their brightest ideas, just their most loyal. They all want to be able to disrupt the others as this past week's Windows kernel implementation under EU interests proved.

 

[igor_kavinski]

I thought Xeon were not plagued by the issue.

I'm hoping they aren't and then we can have those juicy W series Alders and Raptors with twice the AVX-512 power plus the unobtainable AMX accelerator, for the price of Core i9 so Intel bent enthusiasts have an excuse to avoid 7950X/9950X (it doesn't have AMX!!!). And then we get those workstation mobo prices lowered too and get all the goody goody PCIe lanes too

 

[igor_kavinski]

It was the hard way of multiple installs that I found out that consumer windows aren't the way to go with CPUs like EPYC. It (windows) automatically downloaded incompatible PSP drivers, which caused BSODs and boot loops. Had to run a server Evaluation copy to run some benchmarks. Now permanently moved to Ubuntu Server LTS

I'm running the Workstation edition so maybe that's why I didn't have anything bad happen. It runs just like a normal consumer PC would be expected to run and 3D graphics benchmarks run fine too. Haven't run into any situation yet where the application got confused by the CPU name (AMD Engineering Sample and some long string). But then I haven't used it as much as I would like to.

 

[carancho]

What's the current single threaded performance annualized improvement rate for client, since the release of, say, the Ryzen 3xxx line? And for Macs?

 

[igor_kavinski]

From the Geekbench 6 average score:

3950X 1717
5950X 2190 +27.54%
7950X 2941 +34.29%
9950X 3477 +18.22%

M1 2336
M2 2595 +11.08%
M3 3065 +18.11%
M4 3675 +19.90%

 

[Wolverine2349]

Just wanted to add to this thread that Intel stagnated the market even more with their Raptor Lake degradation fiasco. Really hope the only thing left for them to salvage this situation is to lower the prices on their workstation class CPUs with AVX-512 and AMX, just to prevent enthusiasts from turning to Ryzens.

That will not work well for those which there are many who want the good latency of a ring bus CPU and not the crap mesh architecture latency and gimped IPC of Xeon Enterprise class Golden Cove.

No option exists on a ring bus with reliability with those right now which is a shame.

As an enthusiast I want more than 8 P cores on a single tile or ring bus that is stable and reliable. Arrow Lake may be an option if they released it a 12 P core option.

Bartlett Lake I dunno. Have to wait 1 year at least probably for 12 + 0 on the ring bus and can it even be trusted if its going to use Raptor Cove on same 10nm process which appears to have issues. Or maybe and hopefully a new die and stepping will fix it.

Afterall it appears Alder Lake is fine and reliable. And it was on 10nm and was an 8+8 die and a 6+0 die.

Correct me if I am wrong but Raptor Lake thus far has only had one 8+16 die and no others and the 13th Gen rebrands some were 8+16 dies with cores disabled and some were 8+8 Alder Lake dies with parts disabled so Intel was kind of lying by branding some non K 13 series parts as Raptor when they were really better binned Alder Lake parts? So maybe just a bad die and a new stepping needed which Bartlett Lake will fix hopefully.

 

[ondma]

That will not work well for those which there are many who want the good latency of a ring bus CPU and not the crap mesh architecture latency and gimped IPC of Xeon Enterprise class Golden Cove.

No option exists on a ring bus with reliability with those right now which is a shame.

As an enthusiast I want more than 8 P cores on a single tile or ring bus that is stable and reliable. Arrow Lake may be an option if they released it a 12 P core option.

Bartlett Lake I dunno. Have to wait 1 year at least probably for 12 + 0 on the ring bus and can it even be trusted if its going to use Raptor Cove on same 10nm process which appears to have issues. Or maybe and hopefully a new die and stepping will fix it.

Afterall it appears Alder Lake is fine and reliable. And it was on 10nm and was an 8+8 die and a 6+0 die.

Correct me if I am wrong but Raptor Lake thus far has only had one 8+16 die and no others and the 13th Gen rebrands some were 8+16 dies with cores disabled and some were 8+8 Alder Lake dies with parts disabled so Intel was kind of lying by branding some non K 13 series parts as Raptor when they were really better binned Alder Lake parts? So maybe just a bad die and a new stepping needed which Bartlett Lake will fix hopefully.

I would love a 12 + 0 die on Intel's latest process with the latest architecture. But if you have to wait at least another year for 12 + 0 Bartlett Lake, do you really want a 1 gen plus another half gen (1 year closer to Zen 6) trailing architecture on an ancient node? Is the ring bus going to actually make up for that?

 

[carancho]

From the Geekbench 6 average score:

3950X 1717
5950X 2190 +27.54%
7950X 2941 +34.29%
9950X 3477 +18.22%

M1 2336
M2 2595 +11.08%
M3 3065 +18.11%
M4 3675 +19.90%

Thanks - that's a 16.3% CAGR for AMD and 12.0% for Apple. That's a doubling every 5 and a half years @14% CAGR. It probably is too slow of a rate to unlock new use cases in less than a decade. You can only make Excel so much faster before people stop noticing. And that rate of improvement seems insufficient for heavier loads like local mid-sized LLM inferencing.

Any new usecase will need to be powered by significantly different technologies (packaging, accelerators, photonics, more and faster ram, etc) rather than a continued improvement won't current trends. I'm not saying anything new. It's just sad to see how slow progress has been recently.

 

[Nothingness]

Thanks - that's a 16.3% CAGR for AMD and 12.0% for Apple. That's a doubling every 5 and a half years @14% CAGR. It probably is too slow of a rate to unlock new use cases in less than a decade.

M1 was released at about the same time as 5950x IIRC. Did you take that into account (I feel too lazy to check at the moment 😀)?

You can only make Excel so much faster before people stop noticing. And that rate of improvement seems insufficient for heavier loads like local mid-sized LLM inferencing.

Any new usecase will need to be powered by significantly different technologies (packaging, accelerators, photonics, more and faster ram, etc) rather than a continued improvement won't current trends. I'm not saying anything new. It's just sad to see how slow progress has been recently.

Yes, some tasks are better left to dedicated units. Only Intel could believe they could make a GPU with a CPU and some units attached to it. The same applies to LLM, you want to do it on a dedicated which will be much more efficient.

For the rest, I'm afraid everyone is more or less converging to the same performance which means we are dangerously close to the max perf achievable with OoOE (I mean with reasonable amounts of power and area). Some paradigm shift might be needed, the problem being that it has to happen both at programming language level and at CPU level; that's a vicious circle.

 

[poke01]

From the Geekbench 6 average score:

3950X 1717
5950X 2190 +27.54%
7950X 2941 +34.29%
9950X 3477 +18.22%

M1 2336
M2 2595 +11.08%
M3 3065 +18.11%
M4 3675 +19.90%

I think it’s better to use GB5. It will mostly be the same but no SME or AVX-512 will buffer the scores

 

[igor_kavinski]

But if you have to wait at least another year for 12 + 0 Bartlett Lake, do you really want a 1 gen plus another half gen (1 year closer to Zen 6) trailing architecture on an ancient node?

Everyone who invested in LGA1700 deserves a proper refresh which Bartlett Lake should be. Yeah, it may be older architecture (hopefully redwood cove minus the tile latency) but it would be better than a 14900K if you don't want to get rid of your LGA1700 mobo. It should be less than a year. About 8 months by my reckoning.

 

[igor_kavinski]

I think it’s better to use GB5.

Be my guest

 

[igor_kavinski]

Some paradigm shift might be needed, the problem being that it has to happen both at programming language level and at CPU level; that's a vicious circle.

We'll let AI figure it out for itself. Imagine a future where we get faster and faster architectures but we don't understand completely how AI is making them faster because just when we are trying to wrap our heads around the current architecture, AI will bring out a new one. Ad infinitum.

 

[Nothingness]

We'll let AI figure it out for itself. Imagine a future where we get faster and faster architectures but we don't understand completely how AI is making them faster because just when we are trying to wrap our heads around the current architecture, AI will bring out a new one. Ad infinitum.

I hope this will happen after I retire 😁

 

[poke01]

I think the bigger problem is Intel gettinng only a 5% improvement going from Intel 7 to N3B

 

[adamge]

I think the bigger problem is Intel gettinng only a 5% improvement going from Intel 7 to N3B

It's preferable to the negative improvement when they went from 14nm++++ to 10nm.

 

[Wolverine2349]

I would love a 12 + 0 die on Intel's latest process with the latest architecture. But if you have to wait at least another year for 12 + 0 Bartlett Lake, do you really want a 1 gen plus another half gen (1 year closer to Zen 6) trailing architecture on an ancient node? Is the ring bus going to actually make up for that?

Yes good point I agree with you there. Almost makes release date bad for it unless CPU progress stagnates badly in 2025 until at least 2027 or 2028 in IPC gains. And that is not likely to happen so because of waiting 1 year it almost is not a good buy in 2025. Besides who knows if it really fixes stability and degradation issues of Raptor Lake and Raptor Cove

Though still does not make me at all interested in Sapphire Rapids Xeon Workstation with mandatory Enterprise features detrimental to gaming like ECC required RAM and such. And its IPC absolutely sucks compared to client Golden Cove besides just being on a mesh. Skylake X did not have worse IPC compared to client Skylake despite being on a mesh, though it had much worse latency. SPR has worse both compared to client Golden Cove.

 

[DAPUNISHER]

I would love a 12 + 0 die on Intel's latest process with the latest architecture. But if you have to wait at least another year for 12 + 0 Bartlett Lake, do you really want a 1 gen plus another half gen (1 year closer to Zen 6) trailing architecture on an ancient node? Is the ring bus going to actually make up for that?

If it doesn't have any of the flaws of raptor, is priced right, and games well, I think it could be popular with users already on the platform. You need look no further than AM4 for precedence of how well old tech can sell if the bang for buck is there.

 

[Wolverine2349]

If it doesn't have any of the flaws of raptor, is priced right, and games well, I think it could be popular with users already on the platform. You need look no further than AM4 for precedence of how well old tech can sell if the bang for buck is there.

I would argue even better. Will work with WIN10, all cores on single die. If rumors that Zen 6 is also going to max out at 8 cores per CCD as well, the 12 + 0 die may be best for heavily threaded games assuming Intel never releases a 12 P core Arrow Lake.

If Intel releases 12 P core Arrow Lake then yes the 12 + 0 die as long as it does not inherit Raptor Lake flaws is good bang for buck for those on LGA 1700 mobos still. But if Intel does not, an older lower IPC may be better off short term for some games until all the Big.Little and cross CCD-CCX latency7 scheduling quirks get fixed perfectly which may be long ways away given the laziness of some game devs.

Though CPU progress may really be stagnating so as long as the 12 + 0 die does not inherit Raptor Lake stability and degradation flaws, it may age well even with 2025 release if trends such as in this video below keep up.

I know its only 2 weeks but still:

With reports of Zen 5 being underwhelming IPC gains and Lion Cove only 14% IPC uplift it seems CPU IPC and raw performance progress on IPC front is stalling and slowing down from both Intel and AMD.

So if that is indeed true and keeps going that way, that 12 + 0 P core only die as long as it is stable and does not degrade may age very well even with a mid or later 2025 release.

Intel AMD and others seem so obsessed with AI and NPU these days. Though so is NVIDIA yet their GPU raw power of RTX 5090 looks like amassive leap over RTX 4090 in raw performance unlike Zen 5 is over Zen 4 and Arrow Lake will be over Raptor Lake other than fixing the stabiliutry degradation mess. So CPU progress from both AMD and Intel companies is struggling to move much forward? What do you think?

 

[DigDog]

obviously my previous comments do not include the last generation of intel processors

 

[igor_kavinski]

obviously my previous comments do not include the last generation of intel processors

These comments? https://forums.anandtech.com/thread...m-cpu-market-stagnation.2616907/post-41135729

Please elaborate how the last gen Intel CPUs are not covered by your comments.

 

[igor_kavinski]

So CPU progress from both AMD and Intel companies is struggling to move much forward? What do you think?

I think only Intel is struggling. AMD is just keeping their cards close to their chest. They could've blown Intel off the x86 landscape with a much higher performing Zen 5 design but chose not to, just as they don't feel any pressure to increase desktop CPU threads above 32. Think about it. Can there really be any serious technical reason why we don't have a 9990X with Zen 5 fat CCD and Zen 5c 16-core CCD, for a total of 48 threads? Heck, OK so they don't want their TR sales getting impacted. How about they just disable some cores on the Zen5c CCD for a total of 40 threads?

Nothing anyone says is going to convince me that AMD can't do it.

Their 9950X ES is already eating 300W@Unlimited power. That should be more than enough for 40 threads and possibly even for 48 threads since the Zen5c CCD isn't going to eat power like the fat CCD.

If Pat weren't such a feeble old CEO, we would've seen a 40 core K or KS CPU from Intel by now because a CEO with enough testosterone would ABSOLUTELY DEMAND it!