Question Intel's current woes and the low end of the desktop CPU market

[DavidC1]

PTL is majorly internal nodes.

No it's not. The important GPU is on TSMC. Even if we count that, I can't see why they had to use N5 rather than Intel 3 for PCH, because that was their modus operandi since forever. Deprecated nodes for chipsets. Pat became CEO in early 2021. He could have had Pantherlake PCH on Intel 3. Heck even Intel 7 would have worked.

There's no excuse for using TSMC for cost, none at all. Because having their own fabs filled outweighs most of that, unless their internal process is substantially higher. Say it costs 50% more for Intel process, you are still filling your fabs! This is paramount, because real-world experience cannot be substituted by anything else, and yield-learning comes from volume.

That "unlimited budget" thing is somewhat overrated. R&D was increased, the test wafers were increased many folds over which was required.

He overestimated demand and the fab buildout, and few countries that Intel promised were going to have they had to rescind them, because Intel couldn't make them viable.

What about the whole "Oh we thought iGPU drivers would suffice for dGPU" nonsense? So they don't understand their marketbase at all. Great. It's amazing how incompetent and disconnected people with supposedly great education and leadership can be. It's like that's literally that's all in their brains, and nothing else.

Remember the Intel executive that said "my daughter wants two laptops" so PC demand = up and up? Like, what? Do they all live in Beverly hills and don't know a single person making under $300K?

 

[DrMrLordX]

The alternative is 18A + Intel 3. That should be cheaper and more profitable for Intel as it will fill fab space and that directly affects entire revenue not just margins.

That only applies if they have the spare wafer capacity. If not, they're cannibalizing some other product(s).

 

[regen1]

No it's not. The important GPU is on TSMC. Even if we count that, I can't see why they had to use N5 rather than Intel 3 for PCH, because that was their modus operandi since forever. Deprecated nodes for chipsets. Pat became CEO in early 2021. He could have had Pantherlake PCH on Intel 3. Heck even Intel 7 would have worked.

There's no excuse for using TSMC for cost, none at all. Because having their own fabs filled outweighs most of that, unless their internal process is substantially higher. Say it costs 50% more for Intel process, you are still filling your fabs! This is paramount, because real-world experience cannot be substituted by anything else, and yield-learning comes from volume.

One IGPU tile being on TSMC N3E(the other IGPU and all compute tiles are internal) does not change PTL being majorly internal nodes and TSMC N5 is not used anywhere in PTL, the PCD is on TSMC N6. N6 has been used is MTL/ARL/LNL for SoC/PCD/IO tiles. It's a case of cost-optimization plus TSMC N7/6 family is one of those that doesn't see price increase compared to N5 and above. IO/PCH being on older node is not something unique to Intel and it makes sense. Intel 3 is yet to hit peak ramp for GNR.
PTL relative to any of the three: MTL/ ARL/ LNL is very different from Internal vs external. Apart from base FOVEROS dies and MTL and ARL-U compute tiles everything on them was external be it compute tiles, SoC/PCD, iGPUs. PTL was start of reversal and NVL was to be even more so. NVL has SoC/Hub and iGPU also internal. If not for moving NVL-H compute(likely later) to N2/P it would have been way more internal apart from few tiles.

He overestimated demand and the fab buildout, and few countries that Intel promised were going to have they had to rescind them, because Intel couldn't make them viable.

The excessive fab-build out actually didn't happen to that extent, for eg. they did mark Germany and Poland but apart from some ground digging in Germany they didn't really proceed(Germany did promise high subsidization and State-support). Ohio is bit similar and complex as politics got involved. The CHIPS Act Grant was heavily delayed and it came out to be much lower than initially anticipated, TSMC got near similar main Grant for not even bringing out their latest node to US.
Not to forget 14A's fundamental research was done under Ann Kelleher's TD team.
Now the Board's December mess has brought us cancellation of mainstream Xeon(DMR-SP), 18A(-P)/foundry utilisation getting affected severely when DMR-SP is cancelled among many other mess. GNR was their best release in a while and when it seemed they were getting back to track in some ways, delays and cancelation have some back in Server roadmap.

What about the whole "Oh we thought iGPU drivers would suffice for dGPU" nonsense? So they don't understand their marketbase at all. Great. It's amazing how incompetent and disconnected people with supposedly great education and leadership can be. It's like that's literally that's all in their brains, and nothing else.

Remember the Intel executive that said "my daughter wants two laptops" so PC demand = up and up? Like, what? Do they all live in Beverly hills and don't know a single person making under $300K?

Well that's more on Raja Koduri and the team when they were develpoing the GPU line for so long. The team also did produce extremely complex duds like PVC and Rialto Bridge(thankfully Rialto got cancelled, once ChatGPT moment came through). Anyway their GPU IP is getting better and hopefully they keep investing there.
Also statements from executives esp. Sales and marketing sometimes can be quite out of touch with reality, you can see some of the statements by an Interim CEO when she was client head.

 

[adroc_thurston]

delays and cancelation have some back in Server roadmap.

DMR plain sucks, really.

Rialto Bridge

Rialto looked fine.
In any case, they didn't even try actually shipping anything after PVC.
PVC was announced in 2019 and since then AMD shipped 4 gens of CDNA. Intel shipped PVC alone.

 

[regen1]

DMR plain sucks, really.

It won't have been as bad but moreover now you have nothing to show for except GNR-SP. That's way worse for competing timeline.

Rialto looked fine.
In any case, they didn't even try actually shipping anything after PVC.
PVC was announced in 2019 and since then AMD shipped 4 gens of CDNA. Intel shipped PVC alone.

PVC had massive delays but still didn't have that many design wins and was too complex, so many tiles on Advanced packaging.
Had lots of B/W and latency issues as well. The few good things about PVC are that the Aurora/Argonne contract(which was actually pre-earned by Knights Hilll) helped them get inputs for software optimization and some learning for designing complex chiplets of advanced packaging, rest was a fail.

As for Rialto, even though Raja still projects it as some kind of missed opportunity it would have likely failed overall.
It was an optimized PVC on Xe2 with upgraded XeLinks and some other stuff.
It's AI performance in general was quite worse than Gaudi 3, that should speak volume.
That's a disaster. If Rialto was a much simpler design and more performing esp. in AI workloads Intel might not have to pivot to Gaudi 3 series that much. Gaudi could never adopt to SYCL/OneAPI among other issues.

 

[adroc_thurston]

It won't have been as bad but moreover now you have nothing to show for except GNR-SP

True but DMR sucked.

As for Rialto, even though Raja still projects it as some kind of missed opportunity it would have likely failed overall.
It was an optimized PVC on Xe2 with upgraded XeLinks and some other stuff.
It's AI performance in general was quite worse than Gaudi 3, that should speak volume.
That's a disaster. If Rialto was a much simpler design and more performing esp. in AI workloads Intel might not have to pivot to Gaudi 3 series that much. Gaudi could never adopt to SYCL/OneAPI among other issues

Rialto would've been an s/w ecosystem play. Alas.

 

[johnsonwax]

There's an old expression in marketing that people don't want to buy a ¼" drill, they want to buy a ¼" hole.

The problem with the enthusiast market is that they really like drills more than holes, while the rest of the world likes holes more than drills. I read though this thread as a fairly technical person who is not put off by learning technical things and my eyes glaze over. Why the f would anyone want to have to learn the details you are all talking about to buy a low-end PC? High end, sure. But you know what the fastest low-end PC is in the world of everything being an internet service? It's the iPhone and a shocking number of people have replaced their low-end PCs with iPhones, or Android phones or iPads or some device that doesn't require them to read a spec sheet to know if they aren't buying garbage.

Because of that, and because the capabilities of modern silicon is outstripping the need of basic consumer PC users (non-gamers) why would a constellation of low-end chips be needed when you can make something in the category of the M4 that will easily meet 90% of all users needs? That constellation was needed a decade+ ago when there were real costs attached to that GPU but there really aren't any longer. And that constellation held on because OEMs use those SKUs to differentiate their products. It's marketing, mostly. Pretty sure if you combined all of the best features of all of Intels low end iGPU chips, cut out the costs of designing and masking and supply chaining all of the others, and just turned on the economy of scale machine you'd wind up with a product that cost roughly the same and had no trade offs - like Apple has done with their base M series chip, like Microsoft asked Qualcomm to do with Elite X. Like, everything you guys are talking about is an anachronism, that perhaps Intel, in their financial straits have decided they can no longer entertain. The OEMs will need to find a different way to differentiate other than whether this CPU has deliberately crap networking or that CPU has deliberately crap GPU. AMD can choose to entertain them, or throw in the towel and simplify their lineup and move on since they no longer need to compete with Intel on that front.

 

[DrMrLordX]

Why the f would anyone want to have to learn the details you are all talking about to buy a low-end PC?

In a perfect world, everyone would be curious about such things. But, let's be realistic: this thread is not so much about what the consumer will care about as it is about what Intel (in particular) can do to produce attractive products for OEMs to stick in their next generation of consumer low-end PCs.

But you know what the fastest low-end PC is in the world of everything being an internet service? It's the iPhone and a shocking number of people have replaced their low-end PCs with iPhones, or Android phones or iPads or some device that doesn't require them to read a spec sheet to know if they aren't buying garbage.

There's still a PC market. It has far less growth potential than in the past, but it still exists, and is in need of products to service that need.

In summary:

1). The low-end PC market is still here and there is still demand, to be filled by OEMs
2). Intel may not have much of anything new or attractive to offer OEMs that they might want to offer to budget-minded consumers.

Maybe @mikeymikec did not explicitly intend to establish this precise relationship as a part of his initial analysis, but ultimately any serious conversation about low-end CPUs should head in this direction.

 

[511]

No it's not. The important GPU is on TSMC. Even if we count that, I can't see why they had to use N5 rather than Intel 3 for PCH, because that was their modus operandi since forever. Deprecated nodes for chipsets. Pat became CEO in early 2021. He could have had Pantherlake PCH on Intel 3. Heck even Intel 7 would have worked

For us enthusiastic maybe the volume will be the 4+0+4+4 die not the 4+8+4+12 die and N6 is very cheap node and I can bet they must have unused capacity from the OG Arc Failure

True but DMR sucked

DMR is a 2X upgrade over GNR .....

Rialto would've been an s/w ecosystem play. Alas.

Gaudi had much better perf/watt than Rialto but falcon shores shouldn't have been dropped it is much better than Gaudi 3

 

[regen1]

Gaudi had much better perf/watt than Rialto but falcon shores shouldn't have been dropped it is much better than Gaudi 3

They compared Falcon Shores to Nvidia's solutions and thought
1. not competitive in Perf/Watt, 2. isn't anywhere near a Rack-Scale solution as Nvidia's(so was little use for training), 3. software-ecosystem still not where they want it to be. Then decided to cancel and use some for "internal usage".
Perhaps if they could try making deals with some medium/big companies for Falcon Shores and got input for software optimization that could be good but that seems off the table(?) after cancellation. Regardless it wasn't that straightforward a scenario.
Anyway I hope the path is clearer after they get their software-ecosystem to a good state with Project Battlematrix and other efforts and launch Jaguar Shores now that Gaudi series should be out of the way for future stuff. But Sambanova could be a question mark(?), hopefully not.

 

[adroc_thurston]

They compared Falcon Shores to Nvidia's solutions and thought
1. not competitive in Perf/Watt, 2. isn't anywhere near a Rack-Scale solution as Nvidia's(so was little use for training), 3. software-ecosystem still not where they want it to be. Then decided to cancel and use some for "internal usage".
Perhaps if they could try making deals with some medium/big companies for Falcon Shores and got input for software optimization that could be good but that seems off the table(?) after cancellation. Regardless it wasn't that straightforward a scenario.
Anyway I hope the path is clearer after they get their software-ecosystem to a good state with Project Battlematrix and other efforts and launch Jaguar Shores now that Gaudi series should be out of the way for future stuff. But Sambanova could be a question mark(?), hopefully not.

They're not shipping so they're dead.
Rule #1 of the Great GPU Fight: you gotta show up.

AMD and NV show up every year and slug it out in the cage.
Intel did that once, 2 years overdue, then gave up fighting.

 

[regen1]

They're not shipping so they're dead.
Rule #1 of the Great GPU Fight: you gotta show up.

AMD and NV show up every year and slug it out in the cage.
Intel did that once, 2 years overdue, then gave up fighting.

They should/could have with Falcon Shores but by then Board and Management was in turmoil.
But not Rialto Bridge like stuff when it fails at relevant parameters relative to competition and are also very costly to manufacture. Shipping duds that don't sell(like PVC) doesn't do anything for their cause because it doesn't get adopted as much while being expensive to produce.
At least Falcon Shores had some bright spots despite other major issues relative to competition.
What Intel(esp. GPU/accelerator team) acted liked throughout PVC, Rialto Bridge was they could make anything just because they have fancy advanced packaging technology without sorting out performance, efficiency and software-stack. That's the bigger issue.

 

[adroc_thurston]

They should/could have with Falcon Shores but by then Board and Management was in turmoil.
But not Rialto Bridge like stuff when it fails at relevant parameters relative to competition and are also very costly to manufacture. Shipping duds that don't sell(like PVC) doesn't do anything for their cause because it doesn't get adopted as much while being expensive to produce.
At least Falcon Shores had some bright spots despite other major issues relative to competition.
What Intel(esp. GPU/accelerator team) acted liked throughout PVC, Rialto Bridge was they could make anything just because they have fancy advanced packaging technology without sorting out performance, efficiency and software-stack. That's the bigger issue.

Again, rule #1: you either ship or you're dead.
Was Vega20 a meme? Hell yeah. MI100 was also that. But AMD shipped it.

Intel had exactly zero reasons to acquire Habana and waste time and effort keeping that roadmap and s/w stack on life support, only to ditch it in the end for FCS.

 

[jpiniero]

Again, rule #1: you either ship or you're dead.
Was Vega20 a meme? Hell yeah. MI100 was also that. But AMD shipped it.

Gotta have customers. It feels like anyone not buying nVidia (for whatever reason) are buying AMD or the Cloud guys rolling their own. Don't have any real reason to consider Intel.

 

[adroc_thurston]

Gotta have customers

It's less that and more like by shipping frequently and iterating your s/w stack you'll get somewhere.

Don't have any real reason to consider Intel.

Intel doesn't have anything to sell you! They don't ship new GPGPU parts!
It's not even about being bad; Intel outright does not exist in GPGPU right now.

 

[jpiniero]

Intel doesn't have anything to sell you! They don't ship new GPGPU parts!
It's not even about being bad; Intel outright does not exist in GPGPU right now.

There's no room.

They claimed Gaudi was competitive. Nobody bought it.

 

[adroc_thurston]

There's no room.

There was no room for AMD.

They claimed Gaudi was competitive. Nobody bought it.

Gaudi is a dead part tied to a dead roadmap and an even more dead s/w stack.
Intel is trying to commit to having a GPGPU roadmap without *actually* having a GPGPU roadmap. They don't ship anything! (Literally. PVC went EOL with no replacement)

 

[regen1]

Again, rule #1: you either ship or you're dead.
Was Vega20 a meme? Hell yeah. MI100 was also that. But AMD shipped it.

That doesn't work everywhere.
Intel shipped PVC, didn't get them anywhere. Shipping schedule and consistency is important when there is something viable which PVC and esp. Rialto Bridge were not for their timing. Remember they had to go to Gaudi 3 as prime thing because Rialto was bad that again changed the hardware/software track(no SYCL/OneAPI). Didn't succeed as much as even Gaudi 3 wasn't that great relative to competition and had all the software plus other issues.
At least they should start shipping with inference Accelerators(Crescent Island and successors) and Jaguar Shores now as roadmap is simpler now.

 

[jpiniero]

There was no room for AMD.

Course there is. There's the "don't want to deal with nVidia" crowd for one.

(and what does this have to do with low end desktop CPUs, anyway)

 

[adroc_thurston]

That doesn't work everywhere.

YEAH IT DOES.

Intel shipped PVC, didn't get them anywhere

Vega20 went nowhere.
MI100 basically didn't exist. So they continued.

Rialto Bridge were not for their timing

Rialto had enough labs as customers to keep the s/w stack iterating.

At least they should start shipping with inference Accelerators(Crescent Island and successors) and Jaguar Shores now as roadmap is simpler now.

You do understand that no one's gonna dip their fingers into the cookie jar that torched their roadmap for ML big iron like 4 different times?

Course there is. There's the "don't want to deal with nVidia" crowd for one.

(and what does this have to do with low end desktop CPUs, anyway)

Everyone wanted and wants Nvidia.
AMD winning Frontier contract in 2019 felt like a bad fever dream and look where they are now.

 

[regen1]

YEAH IT DOES.

No, it doesn't, not always anyway. Intel and AMD can ship notebook dGPUs now doesn't mean it will necessary work, it hasn't worked for AMD(or Intel with Alchemist) for a while.

Rialto had enough labs as customers to keep the s/w stack iterating.

Don't be that sure. It could have had irrelevant adoption given the trends. PVC's biggest adoption(Aurora/Argonne) came through pre-earned contract for Knight's Hill which is what gave them some software feedback.
Compare Nvidia's H series accelerators to Gaudi 3 and now imagine something way worse in AI perf than Gaudi 3, that's Rialto.
Don't forget that some of the AMD's MI series had Govt. funds helping them. Kudos to AMD for their progress throughout the years.
AMD's roadmap was still simpler relatively with only GPU based accelerators and steady progress unlike Intel where Gaudi 3 destroyed Rialto and had better Scale Out.

They should have tried more with Falcon Shores though.

You do understand that no one's gonna dip their fingers into the cookie jar that torched their roadmap for ML big iron like 4 different times?

This is their reality, some bad products and bad decisions, acquisitions that they didn't deal well with, roadmap slips. Still doesn't justify shipping Rialto with its timing.
They messed up bad with PVC, Rialto, Habana, Nervana, etc. but that's past, nothing can be done about it.
At least they can try now with focus starting with simpler designs for Inference and other Accelerators like Jaguar Shores. They can still mess up by acquiring some AI start ups and then not being able to deal with 2 different product lines and their software integration.

 

[511]

Don't be that sure. It could have had irrelevant adoption given the trends. PVC's biggest adoption(Aurora/Argonne) came through pre-earned contract for Knight's Hill which is what gave them some software feedback.
Compare Nvidia's H series accelerators to Gaudi 3 and now imagine something way worse in AI perf than Gaudi 3, that's Rialto.
Don't forget that some of the AMD's MI series had Govt. funds helping them. Kudos to AMD for their progress throughout the years.
AMD's roadmap was still simpler relatively with only GPU based accelerators and steady progress unlike Intel where Gaudi 3 destroyed Rialto and had better Scale Out.

They should have tried more with Falcon Shores though.

Yeah Gaudi 2 blew Rialto in ppw and perform so going Gaudi was no brainier decision but skipping Falcon Shores was stupid decision

 

[adroc_thurston]

Intel and AMD can ship notebook dGPUs now doesn't mean it will necessary work, it hasn't worked for AMD(or Intel with Alchemist) for a while.

Oh yes it does and it was working for AMD before RDNA3.

Don't be that sure. It could have had irrelevant adoption given the trends. PVC's biggest adoption(Aurora/Argonne) came through pre-earned contract for Knight's Hill which is what gave them some software feedback.
Compare Nvidia's H series accelerators to Gaudi 3 and now imagine something way worse in AI perf than Gaudi 3, that's Rialto.
Don't forget that some of the AMD's MI series had Govt. funds helping them. Kudos to AMD for their progress throughout the years.
AMD's roadmap was still simpler relatively with only GPU based accelerators and steady progress unlike Intel where Gaudi 3 destroyed Rialto and had better Scale Out.

Perf does not matter when you're trying to run a roadmap and cultivate a software ecosystem.

This is their reality, some bad products and bad decisions, acquisitions that they didn't deal well with, roadmap slips. Still doesn't justify shipping Rialto with its timing.

Yeah it does. Again, Rialto had labs as customers.

 

[regen1]

Oh yes it does and it was working for AMD before RDNA3.

No, if it worked that well they won't have negligible adoption for RDNA 3 notebook SKUs nor would they abandon RDNA 4 notebook line up for long. Just showing up doesn't always work, isn't enough. Nvidia have practically cornered the notebook dGPU segment. Others didn't do as well on things that mattered(volume, brand image, perf, perf/W, features, media engine, ISV/IHV related works, etc.) consistently to be in that game.

Perf does not matter when you're trying to run a roadmap and cultivate a software ecosystem.

Performance viability plus timeline matters for adoption and very low adoption means not much movement in software feedback.
PVC and Rialto were never primarily designed as AI accelerators, were more HPC oriented. PVC(and thus Rialto) without delays could have fared better but Rialto on its timeline was not going to, could have been just a money-loser for not much in return. At least via Knights Hill, PVC had a big customer in Aurora/Argonne(and that came up way too late) which gives them software input, can't be certain but unlikely Rialto would have got anything similar.

Yeah it does. Again, Rialto had labs as customers.

Even if we consider that, PVC too had them(eg. "Dawn" SC at Cambridge Zettascale Lab) and few things like TACC's Stampede 3.
Rialto being a sort of improved PVC had its writing on the wall post-ChatGPT moment(and being so late for so little perf).
Things were so bad that PVC had to be discontinued just 2 years after release: https://www.servethehome.com/intel-ponte-vecchio-spaceship-gpu-no-longer-hunting-new-clusters/
Intel had unsold inventory of PVC that they were trying to dispose of by giving it at discounted rates to people like George Hotz and some had to be donated away.
They managed some customers and deals with Gaudi 3/2 as well(IBM, Stability AI, etc.) but that didn't do much for the line cause of H/W and S/W limitations.
Intel had cancelled/abandoned many products/line-ups(Omnipath, Knights series, PVC, Rialto, NNP-I and many others) but it's not like all of them were worthy of release.

but skipping Falcon Shores was stupid decision

Despite shortcomings(perf/W, rack-scale, software stack readiness) of Falcon Shores they could have released it and/or tried deals similar to what they did with Gaudi 2

Stability AI’s Intel fundraise came with hefty hardware purchase commitments, sources say

Stability AI was reported to have raised nearly $50m from Intel in late 2023, but the round came with significant strings attached

sifted.eu

Intel and Dell to Collaborate on AI; Intel to Build 'Most Powerful AI Supercomputer in Europe' for Stability AI - Gelsinger

This story was updated on September 21 based on input from Intel: With the goal of building the most powerful AI supercomputer in Europe, Intel has formed an alliance with Stability AI, a London-based artificial intelligence- startup. Announced at this week’s Intel Innovation 2023 conference in...

insidehpc.com

 

[adroc_thurston]

No, if it worked that well they won't have negligible adoption for RDNA 3 notebook SKUs

They screwed up!

nor would they abandon RDNA 4 notebook line up for long

APUs are the name of the game now.
Gorgon-Halo is next.
Then, we have a new stick.

Others didn't do as well on things that mattered(volume, brand image, perf, perf/W, features, media engine, ISV/IHV related works, etc.) consistently to be in that game.

Oh noes, RDNA2 was ramping dang well.

Performance viability plus timeline matters for adoption and very low adoption means not much movement in software feedback.
PVC and Rialto were never primarily designed as AI accelerators, were more HPC oriented. PVC(and thus Rialto) without delays could have fared better but Rialto on its timeline was not going to, could have been just a money-loser for not much in return. At least via Knights Hill, PVC had a big customer in Aurora/Argonne(and that came up way too late) which gives them software input, can't be certain but unlikely Rialto would have got anything similar.

Words words words.
They have to ship.

Even if we consider that, PVC too had them(eg. "Dawn" SC at Cambridge Zettascale Lab) and few things like TACC's Stampede 3.
Rialto being a sort of improved PVC had its writing on the wall post-ChatGPT moment(and being so late for so little perf).
Things were so bad that PVC had to be discontinued just 2 years after release: https://www.servethehome.com/intel-ponte-vecchio-spaceship-gpu-no-longer-hunting-new-clusters/
Intel had unsold inventory of PVC that they were trying to dispose of by giving it at discounted rates to people like George Hotz and some had to be donated away.
They managed some customers and deals with Gaudi 3/2 as well(IBM, Stability AI, etc.) but that didn't do much for the line cause of H/W and S/W limitations.
Intel had cancelled/abandoned many products/line-ups(Omnipath, Knights series, PVC, Rialto, NNP-I and many others) but it's not like all of them were worthy of release.

Words.
You know you can just ship.