Gelsinger out as Intel CEO

[chithanh]

Itanium was a pretty competent design for the assumptions that were widespread in the industry at the time. Everyone who was an expert industry wide was saying that VLIW was the way to go, as compilers would optimize everything. It wasn't until much later the pendulum swung and the software world realized the compiler that would be necessary simply wasn't possible to write.

Yeah so Intel designed a CPU around a mythical compiler which didn't exist yet, was absolutely huge and expensive, and solved problems that almost nobody actually had.
At the time, JIT compiled software (Java etc.) was already a thing, and JIT and a highly complex compilation process don't mix well for obvious reasons.
Itanium was designed for two things, mop up the server CPU landscape (Alpha, MIPS, PA-RISC, POWER, SPARC) and run in HP servers.

the software world realized the compiler that would be necessary simply wasn't possible to write.

Not only the compiler didn't exist. The techniques to write (and more importantly, debug) software for Itanium didn't exist.
If you ever had the "pleasure" of debugging a software crash on Itanium, then the crash dump would tell you exactly nothing because you were inside some VLIW instruction where it was impossible to see what was going on. Later generation compilers introduced a debug mode which got rid of all the VLIW / instruction level parallelism stuff, and hopefully the crash would be reproducible there (or on other CPUs).

 

[Zarathustra[H]]

Raja - eh, we'll see - he's better than what they had before for sure, and I'm not sold on their GPU business...

I tend to think of Raja as somewhat of a fraud.

First he tanked AMD's Vega, and then he sold himself to Intel and fumbled that too.

Now he is apparently at some AI -focused software company. God help them.

 

[ChadD]

Can't blame him for canning projects when he inherited a company that was set up for plummeting revenues. He had to do what he had to do to cut costs.

He may have had an impact on the marketing, branding and posture of Core Ultra, but - again - CPU projects - from inception to launch - are complicated and really long lead projects. The beginnings of Core Ultra probably long pre-date Gelsinger as CEO. As did Arrow Lake and Raptor Lake. Their failures happened on his watch, but likely long predated him.

I'm not saying he was the right choice to lead the Intel turn-around. He most definitely was not. Intel needed a Change agent, and putting the old guard at the top was not the way to get change, but I think he did as good a job as anyone could have done given the shit sandwich he was handed. And now the board has turned him into the fall-guy for it.

I know you can't blame him for everything.
In the end, he should never have been named CEO. IMO and I'm not the only one Pats career at Intel was not the big success that it was sold as. He was a favorite of Andy Groves at one point, and in the end all of Intels issues happened in and around that 40 year ann point. Were there was obviously internal power struggles, and the second gen Intel management started to ingrain the wrong attitudes. Intel went from plucky startup fighting the big boys. To being the big boy. Guys like Pat came in when Intel was THE company changing the freaking world... and they believed they could do no wrong. I don't blame Pat for culling some of the odd tech Intel had picked up over the years and streamlining. He did make some terrible choices as well in regards to cutting.

In any event, this is good for Intel. They just got US gov sale protection, with the fab protections. I am sure the gov ask as no more Pat. Whoever comes in now to replace him HAS got to get the fabs rolling. The Chips act at this point is the music Intel is going to have to dance to.

 

[ChadD]

I tend to think of Raja as somewhat of a fraud.

First he tanked AMD's Vega, and then he sold himself to Intel and fumbled that too.

Now he is apparently at some AI -focused software company. God help them.

There are some players... that If Pat was responsible for driving them off may have been for the best. There were a few changes Fab side that probably needed to happen as well. Pat made some good and bad choices.

 

[LukeTbk]

, and solved problems that almost nobody actually had.

Considering in hindsight the giant success of GPU parallel compute not sure if this is a fair assessment, the power of ultra parallel computing was something people wanted a lot (even if they do not cared that it is parallel or not), the proposed solution for it was not the good (or lost to gpu for an other reason).

If you ever had the "pleasure" of debugging a software crash on Itanium

Having a crash inside if statement that were never true during your execution was a lot of fun to find out...

 

[Spirit_Retro]

Oh he failed. I was a strong defender of him, for certain - and I thought (think? Hope?) they're on the right path in the long run here, it takes time to build a new CPU design and get there. Jim Keller went back to AMD in 2012 - Zen released in 2017. 5 years for a new design to tape out and get to release. The arrow lake successor is what I figured we'd see as the first real "Gelsinger led" solution, although I'd hoped Arrow Lake would be better too and show improvement on the way. I got worried when Keller's Royal Core was canned, and ARL doesn't show any notable improvement yet... I still figured he had one more release to see if he'd been right on that choice. Now I fear they're fucked, and can only hope that what they've been working on pans out. I don't want a single vendor market with just AMD any more than I want a single vendor market with Intel.

Remember when I said Intel would sell off fabs? We would meet here in 6 months?

We are pretty much there. Intel will not be able to support in house manufacturing. They will have to sell.

This means that TSMC's fab(s) that are coming into the USA will be augmented by Intel exiting the chip manufacturing business.

Now the scary part is TSMC's situation with all the politics around China an Taiwan. Though I fear if that kicks off, with other activity around the world, we won't be too worried about it due to more pressing issues.

 

[LukeTbk]

In any event, this is good for Intel.

That a company like Intel did not had a CEO replacement ready for when they fire their CEO..... it was a short reign but still... that make it look a bit chaotic no ?

Would wait to see who will be the replacement before judging if it is good or not.

Maybe instead of focusing energy-money on foundry they could have went more like AMD instead and if they would have focused on AI who knows..., but the focus on the foundry level reinforced their America first, onshoring supply chain narrative and CHIPS money coming their way.

We are pretty much there. Intel will not be able to support in house manufacturing. They will have to sell.

Will the US let that happen ? To someone like TSMC to boot ? the giant CHIPS 8 billions check come with the condition that Intel stay the majority owner of its foundry business and we can imagine building fab capacity has well.

 

[Mega6]

FYI - By Taiwan Law, the TSMC cutting edge nodes will stay in Taiwan. There will probably never be cutting edge semi nodes in the USA ever again. Is this needed for national security? No. The military is perfectly happy designing on older nodes (at least for now).

 

[erek]

https://finance.yahoo.com/news/intel-announces-retirement-ceo-pat-133000708.html

Had been CEO since 2021 and with the company for decades before that.

Intel CEO admits 'I've bet the whole company on 18A'

Feb 28, 2024 — Pet Gelsinger resisted the idea last year, now he won't deny everything is riding on Intel's most advanced node yet. "I've bet the whole company”

(Pat’s signature 386)

 

[lopoetve]

I'm pretty much with you there. The first releases made off of strategic decisions are about 5 years out. My 10 year estimate is based on that things generally need to get out into the marketplace and mellow a bit, as well as product refreshes need to hit before you really see the impact the have.

I'm not suggesting that 16th is going to be amazing and it will redeem Gelsinger. I don't know one way or another. All I am suggesting is that based on product cycles it is too early to really judge the decisions he has made based on those products performance in the marketplace.

I honestly wouldn't be surprised if this was how the job was sold to him in the first place.

The board knew they needed to change, and that change was going to hurt a lot of feelings and result in a lot of poised off people, so they needed someone to come in, do the unpopular things, and then afterwards be able to blame them and fire them.

Find someone competent enough to take those actions, promise him a big golden parachute knowing he will eventually need it, and then blame him and fire him when the inevitable difficult decisions are made

I think Pat Gelsingers decisions have probably saved the company (at least for now). Intel was in a free-fall, and now that free-fall has been stopped. Sure, the outlook going forward is much more tempered than it was in the past, but that past outlook was completely unrealistic, and based on Intel past, when they could leverage the dominant best silicon fabs in the world to stay ahead. This has obviously not been the case since at least 2016 when they were forced off of "tick-tock" which we later found out was due to the early rumblings of the 10nm fiasco.

Pat Gelsinger's tenure really just forced everyone (board, shareholders, and fanboys alike) to face some harsh realities about the company, and that is never popular.

Bingo. 16th for me is the first one we can really judge on - the rest he could cancel, but you can't change something in tape-out other than cancelling it, and something close to... yeah, no.

You may be right that he was a sacrificial lamb to do the needful, as our friends overseas say - we may never know.

To be clear I'm not saying some of the things he cancelled didn't make sense to cancel. I just wonder how his canceling, pushing up, splitting CPU teams into those working for non Intel fabs and those still working toward Intel fabrication. One rumor that seems pretty well supported by recent products is that their is two groups of Intel CPU engineers, one group hates E cores another that supports them... this resulting in odd decisions like SMT being disabled. Also AVX512 being dropped at one point.

That's definitely the Royal Core mess - the P core folks were on Keller's side, while E was the other team. But it's also always been that way - US vs Israel dev shops, it just usually was that one designed the current gen, the other the next gen, and back and forth (historically). Gelsinger believed in all-E core designs, which I find... lacking in vision.
AVX is an oddity - I need to catch back up on where that's at now.

Publicly canned things, and to be fair some of these things needed to be culled perhaps. Remember Intel is a lot more then just CPUs.
Optane.

I worked closely with that team. Dead for a reason. Sucks, but the PMEM side wasn't working except for SAP, the storage side wasn't needed, it just didn't fit. I own dozens of the things.

Tofino network switch chip. Intel bought Barefoot networks in 2019. Pat canned the new division a few quarters later.
4g and 5g chipsets. Pat canned all Intels cellular wireless plays.

Both good points; the whole push for virtualized networking that barefoot was supposed to feed ended up being a dead end owned by two vendors (VMware NSX and Cisco). Lots of others invested in similar though - intel was just late to the market there (see also lightbits, tied to the above optane point)

The Data Center Solutions Group (Intels own server business) Pat sold it to MiTAC

Extremely small margins (competing against Super Micro, Quanta, and even Dell).
Extremely low volume (they couldn't sign whitebox deals outside of a couple with EMC and one or two major customers @ a couple of thousand nodes total)
Took a lot of R&D that wasn't paid back easily
Support costs ate limited margin

I worked a bit with this group (see the whitebox deals above) and it was not a good business line for them. Same as the NUC lineup - it was for homelab and geek enthusiasts, but not great for actual profit.

Blockscale Intels mining ASIC pat canned that.

eh

He canned the division that resulted from Intels purchase of Replay Tech.... I don't know that one might have deserved it and Pat wasn't their for the purchase of Replay. Intel making software for sports replays does seem odd.
The Intel drone division... ok that also may have been warranted as beyond the scope.

ROFL

The non public things are things like Royal Core. The GPU division. Not sure what is going on with GPU, we know battle mage looks like it was scaled back but I don't know if anyone knows work is on going for the battle mage follow up or not. Some reports say its shit canned others say its still in progress?

I agree on these for sure. GPU tech is required at this level. Even if not going into top-end GPUs.

Remember when I said Intel would sell off fabs? We would meet here in 6 months?

We are pretty much there. Intel will not be able to support in house manufacturing. They will have to sell.

This means that TSMC's fab(s) that are coming into the USA will be augmented by Intel exiting the chip manufacturing business.

Now the scary part is TSMC's situation with all the politics around China an Taiwan. Though I fear if that kicks off, with other activity around the world, we won't be too worried about it due to more pressing issues.

The CHIPS act keeps them from selling the fabs, pretty sure.

 

[Lakados]

Intel's downfall started when their fabs fell behind.
Intel was too proud to ask for help fixing their fabs when it counted and now they are having to burn money in an attempt to make up ground on TSMC.
Intel had to "get creative" with their CPU's to make up for their fabrication failures and that "creativity" is circling back around to kick them in the teeth.

Intel spent decades promoting Marketing personnel to management overseeing engineers, which lead to generational failures in leadership, direction, and ultimately blead talent dry.
Intel needed to put an Engineer back at the helm for sure, but Pat wasn't that engineer.
Intel needed a hatchet man, who could tear down the management structure and reengineer Intel division by division, and tighten up the whole ship.

Pat had great plans but failed on their execution.

O highly doubt Pat woke up one morning and said "F-This I'm out", guaranteed the board called him up and said "you can go peacefully or forcefully but you are going", and he chose to take the path of least resistance.

 

[ChadD]

FYI - By Taiwan Law, the TSMC cutting edge nodes will stay in Taiwan. There will probably never be cutting edge semi nodes in the USA ever again. Is this needed for national security? No. The military is perfectly happy designing on older nodes (at least for now).

For anything hardened with redundant ECC logic designs they don't need cutting edge nodes anyway. Though you can probably make a case for wanting to be in control of the chips powering their super computing needs. For weapon systems the latest nodes are probably not ideal.

 

[ChadD]

O highly doubt Pat woke up one morning and said "F-This I'm out", guaranteed the board called him up and said "you can go peacefully or forcefully but you are going", and he chose to take the path of least resistance.

I'm sure the call went more; the gov says we can have our chips money before the end of the year. The catch is...

 

[Aurelius]

One thing's for sure: Gelsinger went out sooner than he planned. You don't announce your retirement a day after it happened, especially not when your socials suggested it was business as usual the week before.

 

[ChadD]

That's definitely the Royal Core mess - the P core folks were on Keller's side, while E was the other team. But it's also always been that way - US vs Israel dev shops, it just usually was that one designed the current gen, the other the next gen, and back and forth (historically). Gelsinger believed in all-E core designs, which I find... lacking in vision.
AVX is an oddity - I need to catch back up on where that's at now.

I agree on these for sure. GPU tech is required at this level. Even if not going into top-end GPUs.

The CHIPS act keeps them from selling the fabs, pretty sure.

Where AVX sits now... is AMD has Zen5 running a double pumped AVX512 path, that eats Intels comparable server solutions for breakfast. AVX512 is actually starting to get more in more use in AI datacenter workloads. In stuff that leverages it the performance difference between Zen5 server vs Intel is massive. Like 4-5x faster. I understand why Intel dropped it off their consumer CPUs, it was a lot of silicon and Intel basically being a fab process behind there was logic to dropping it. It seems like its going to be an important feature going forward data center wise. Consumer end, probably not that important unless your encoding a lot of video... apparently a few of the popular encoders have had massive avx gains recently.
https://www.tomshardware.com/pc-com...mplementing-handwritten-avx-512-assembly-code

All the little divisions I mentioned that Pat canned are funny reading through. I mean its hard to fault killing most of it. I think they should have stayed in the Intel server business. I know it wasn't a big profit mover. I think it could have helped Intel shore up some KEY big players that have switched to AMD over the last few years. It was one more tool IF and it is if they managed it properly and used it that way. Your likely 100% correct it was probably never and would never have been managed properly so axing it probably was the right call.

On the chips act it sounds like... to release the money, they added some new stipulations. The Fabs can no longer be split off from Intel proper which was Pats plan. I wonder if Pat was in on the negotiation with the gov on that, I mean he must have been right. The gov essentially called out his plans on that end and said NO. Maybe with the gov putting an end to his plans there he figured it was time to go. I mean I'm sure his departure was what they were looking for in general. No more talk of splitting Intel, no more talk of selling Intel off in bits. The Gov frankly seems to have saved Intel. They made Intel to much of a mess for anyone that would consider a full purchase or worse a merger.

 

[Zarathustra[H]]

That's definitely the Royal Core mess - the P core folks were on Keller's side, while E was the other team. But it's also always been that way - US vs Israel dev shops, it just usually was that one designed the current gen, the other the next gen, and back and forth (historically). Gelsinger believed in all-E core designs, which I find... lacking in vision.

I think both are good.

P cores serve their purpose and E cores serve their purpose.

For a machine used for games, I'd probably lean towards and all P-core design. For a lightweight server or a basic office/web browsing design an all E core design (like Alder Lake-N) is fantastic.

For a more general purpose machine that does both, having a few P-cores to power through the heavy stuff, and a few e-cores to offload light background tasks during load (or when the machine is idle to save power) is a great idea. ARM BIG.little (or was it big.LITTLE) has been doing this well for over a decade.

You just have to have a scheduler that can tell the difference and is good at prioritizing things on the right core, and Microsoft has been really bad at implementing this well. Even worse when you throw SMT on those P cores.

To be fair, it gets complicated. Lets say you have 16 cores, 8 P and 8 E (like an Alder Lake i7) and you are running a game that loads 6-7 of them, and several background tasks.

If you have 16 threads going at once, that is easy. Sort them in order of heavy to light, but the heaviest threads on the P-cores, and the lightest threads on the E-cores.

Along comes that 17th thread. What do you do? If it is really small, it probably makes sense to make the two smallest background tasks share an E-Core. But what if it is heavier (but not quite as heavy as the main game cores)? Is it better to consolidate the two lightest E-core loads, and give it a dedicated E-Core, OR do you queue it up on a secondary logical cores on one of the SMT enabled P-cores?

Then factor in heat generated, and how it likely reduces the max boost clock on the P cores. Is it more performant to consolidate a lot of less important tasks on the E cores in order to keep heat low, leave a couple of P cores idle and get the most out of the best P-cores boost clocks?

And no one tells you which threads are important and which don't matter. At least not consistently. So you have to guess. And you are doing all of this in real time without th ebenefit of seeking input online or elsewhere due to the insane latency that would have compared to intra-CPU latency.

It gets really goddamn complicated in a hurry. So many dynamic clock speeds, dynamic voltages, core temperatures, fan speeds, ambient temperatures, P-cores, SMT, E-cores, etc. etc. And that's only in this theoretical example with <20 threads. Now consider that even an "idle" windows box has like 1200 threads running at any given time, and the complexity gets turned up to like 11.

AVX is an oddity - I need to catch back up on where that's at now.

...and there's another variable. On a compatible system, does it make more sense to run something in AVX-512, while artificually limiting the CPU clocks and generating more heat, or is a lesser impact to just run it with more traditional instructions?

Ideally this is something the scheduler could figure out on the fly, but I don't think that's how it has been implemented. The developer - i believe - has to either enable it or disable it statically in their code/compiler.

Anyway, you'd have to have run some serious benchmarking on how all of these things interact in different circumstances, and with all of the variables and interactions this gets really difficult. I'd imagine some sort of Design of Experiments methodology combined with AI might be the only way to gather and analyze this data. And for all Microsofts pushing of Co-Pilot AI for end user stupid shit where we would all be better off without it, they can't seem to use AI properly for something where it would really make sense.

For whatever reason, the Linux kernel does a better job with all of these complex variables, to the point where many CPU tasks where there is resource constriction actually run faster in a Windows VM in KVM on a linux host than they do in Windows themselves, as the linux Kernel is pretty good at sorting threads on one half of SMT enabled cores instead of bombarding both logical cores. I don't think that can be credited to DoE and AI though. Probably just some well through out algorithms. But this was back before the unholy mess of combination of Dynamic Clock and Voltage P-Cores with SMT and E-Cores without. I have no idea how the Linux and Windows schedulers compare on these highly complicated 12th, 13th and 14th gen multi-mode CPU's.

I think Intel expected Microsoft to step up to the plate and optimize the shit out of their scheduler, and they just didn't. And this is why Intel decided to move away from SMT to make the problem easier to solve. This is probably a missed opportunity IMHO, and it is more based on Microsoft dropping the ball than on INtel themselves (but to be fair, the best hardware in the world is worthless if you don't have software hat can adequately run on it)

 

[Lakados]

Where AVX sits now... is AMD has Zen5 running a double pumped AVX512 path, that eats Intels comparable server solutions for breakfast. AVX512 is actually starting to get more in more use in AI datacenter workloads. In stuff that leverages it the performance difference between Zen5 server vs Intel is massive. Like 4-5x faster. I understand why Intel dropped it off their consumer CPUs, it was a lot of silicon and Intel basically being a fab process behind there was logic to dropping it. It seems like its going to be an important feature going forward data center wise. Consumer end, probably not that important unless your encoding a lot of video... apparently a few of the popular encoders have had massive avx gains recently.
https://www.tomshardware.com/pc-com...mplementing-handwritten-avx-512-assembly-code

All the little divisions I mentioned that Pat canned are funny reading through. I mean its hard to fault killing most of it. I think they should have stayed in the Intel server business. I know it wasn't a big profit mover. I think it could have helped Intel shore up some KEY big players that have switched to AMD over the last few years. It was one more tool IF and it is if they managed it properly and used it that way. Your likely 100% correct it was probably never and would never have been managed properly so axing it probably was the right call.

On the chips act it sounds like... to release the money, they added some new stipulations. The Fabs can no longer be split off from Intel proper which was Pats plan. I wonder if Pat was in on the negotiation with the gov on that, I mean he must have been right. The gov essentially called out his plans on that end and said NO. Maybe with the gov putting an end to his plans there he figured it was time to go. I mean I'm sure his departure was what they were looking for in general. No more talk of splitting Intel, no more talk of selling Intel off in bits. The Gov frankly seems to have saved Intel. They made Intel to much of a mess for anyone that would consider a full purchase or worse a merger.

AVX 512 doesn't work on mixed core CPUs, Intel developed AVX10 to circumvent that.
AVX10 works with the full mixed core stuff and uses all the same code as AVX512 but compiles out to a different DLL, so you can't just use the same DLL but you can use the same code.
AVX10 also fixes some issues with 512, but I am pretty sure AMD technically already supports AVX10 on the Zen 4 and 5 parts

 

[Sir Beregond]

I tend to think of Raja as somewhat of a fraud.

First he tanked AMD's Vega, and then he sold himself to Intel and fumbled that too.

Now he is apparently at some AI -focused software company. God help them.

100% agree. He wasn't good for RTG and I think Pat probably saw he wasn't going to be good for Intel GPU. Raja oversaw what...Polaris, Vega, maybe Navi? Radeon is still floundering in my opinion. I think the only good generation Radeon has had in recent memory is RDNA2.

 

[Zarathustra[H]]

The CHIPS act keeps them from selling the fabs, pretty sure.

I think this hurts intel shareholders short term, looking for a quick restructure to boost stock prices, but long term, I think it may actually result in a stronger Intel, if the federal government holds their feet to the fire, silencing the quarterly earnings call gamblers screams andbout needing to see a higher P/E ratio erach quarter, and allows them to ignore the fickle shareholders for a little bit and fix their in housefabs.

Imagine if AMD 15 years ago got billions of dollars from the federal government to fix their fabs and still had them in house?

I'm not a fan of the federal government throwing taxpayer money in the pockets of several hundred billion dollar corporations, but I think there is no doubt that AMD would be a stronger company today if they had their own fabs brought up to modern standards with taxpayer dollars.

GlobalFoundtries kind of made a mess of it when they decided "sub 12nm is hard, let's just not", but in this alternate universae where AMD got to not only keep their fabs, but also upgrade them on the government dime, they would likely be a much stronger company today.

(Lol, I just remembered that for a short period there, right after Gelsinger took the helm in 2021, Intel was seriously proposing buying out GlobalFoundries)

 

[lopoetve]

Where AVX sits now... is AMD has Zen5 running a double pumped AVX512 path, that eats Intels comparable server solutions for breakfast. AVX512 is actually starting to get more in more use in AI datacenter workloads. In stuff that leverages it the performance difference between Zen5 server vs Intel is massive. Like 4-5x faster. I understand why Intel dropped it off their consumer CPUs, it was a lot of silicon and Intel basically being a fab process behind there was logic to dropping it. It seems like its going to be an important feature going forward data center wise. Consumer end, probably not that important unless your encoding a lot of video... apparently a few of the popular encoders have had massive avx gains recently.
https://www.tomshardware.com/pc-com...mplementing-handwritten-avx-512-assembly-code

Yeah, was definitely thinking consumer. DC side I've moved into security, so I haven't kept up as close with what's going on there (and especially not the AI hype train - we've still got the crash and then real-world to deal with there).

All the little divisions I mentioned that Pat canned are funny reading through. I mean its hard to fault killing most of it. I think they should have stayed in the Intel server business. I know it wasn't a big profit mover. I think it could have helped Intel shore up some KEY big players that have switched to AMD over the last few years. It was one more tool IF and it is if they managed it properly and used it that way. Your likely 100% correct it was probably never and would never have been managed properly so axing it probably was the right call.

Ah, so the server business didn't do that. They literally sold competing products against Dell, HP, Lenovo, etc - and they weren't particularly good (there's a reason you don't see them around). Think a slimmed down SuperMicro, but for more money. Few options, few configurations available, etc - they dreamed of making it as an OEM for things like storage arrays, but Quanta/Flextronics/SuperMicro have that market utterly nailed. They also weren't great at support, limited parts depot network, etc.

Honestly I'm not sure how to shore up that business right now, not with Epyc doing so well.

On the chips act it sounds like... to release the money, they added some new stipulations. The Fabs can no longer be split off from Intel proper which was Pats plan. I wonder if Pat was in on the negotiation with the gov on that, I mean he must have been right. The gov essentially called out his plans on that end and said NO. Maybe with the gov putting an end to his plans there he figured it was time to go. I mean I'm sure his departure was what they were looking for in general. No more talk of splitting Intel, no more talk of selling Intel off in bits. The Gov frankly seems to have saved Intel. They made Intel to much of a mess for anyone that would consider a full purchase or worse a merger.

Once split, much easier for a foreign company to buy them. Agreed entirely.

I think both are good.

P cores serve their purpose and E cores serve their purpose.

For a machine used for games, I'd probably lean towards and all P-core design. For a lightweight server or a basic office/web browsing design an all E core design (like Alder Lake-N) is fantastic.

For a more general purpose machine that does both, having a few P-cores to power through the heavy stuff, and a few e-cores to offload light background tasks during load (or when the machine is idle to save power) is a great idea. ARM BIG.little (or was it big.LITTLE) has been doing this well for over a decade.

You just have to have a scheduler that can tell the difference and is good at prioritizing things on the right core, and Microsoft has been really bad at implementing this well. Even worse when you throw SMT on those P cores.

To be fair, it gets complicated. Lets say you have 16 cores, 8 P and 8 E (like an Alder Lake i7) and you are running a game that loads 6-7 of them, and several background tasks.

If you have 16 threads going at once, that is easy. Sort them in order of heavy to light, but the heaviest threads on the P-cores, and the lightest threads on the E-cores.

Along comes that 17th thread. What do you do? If it is really small, it probably makes sense to make the two smallest background tasks share an E-Core. But what if it is heavier (but not quite as heavy as the main game cores)? Is it better to consolidate the two lightest E-core loads, and give it a dedicated E-Core, OR do you queue it up on a secondary logical cores on one of the SMT enabled P-cores?

Then factor in heat generated, and how it likely reduces the max boost clock on the P cores. Is it more performant to consolidate a lot of less important tasks on the E cores in order to keep heat low, leave a couple of P cores idle and get the most out of the best P-cores boost clocks?

And no one tells you which threads are important and which don't matter. At least not consistently. So you have to guess. And you are doing all of this in real time without th ebenefit of seeking input online or elsewhere due to the insane latency that would have compared to intra-CPU latency.

It gets really goddamn complicated in a hurry. So many dynamic clock speeds, dynamic voltages, core temperatures, fan speeds, ambient temperatures, P-cores, SMT, E-cores, etc. etc. And that's only in this theoretical example with <20 threads. Now consider that even an "idle" windows box has like 1200 threads running at any given time, and the complexity gets turned up to like 11.

Totally agreed. Gelsinger's belief (that he came out a few months ago with) was that the future of Datacenter workloads was all E cores and GPUs. That works for very specific solutions, of course (AI among them), but is gonna SUCK for big DB/SAP/etc, which still back all those thousands of app servers on the front end. Software isn't that massively parallel yet, and containerization hasn't taken off enough to make it efficient yet except at the hyperscaler level, and that's abandoning on-prem traditional workloads (which aren't going away, and are often coming back from cloud) to AMD.

...and there's another variable. On a compatible system, does it make more sense to run something in AVX-512, while artificually limiting the CPU clocks and generating more heat, or is a lesser impact to just run it with more traditional instructions?

Ideally this is something the scheduler could figure out on the fly, but I don't think that's how it has been implemented. The developer - i believe - has to either enable it or disable it statically in their code/compiler.

Anyway, you'd have to have run some serious benchmarking on how all of these things interact in different circumstances, and with all of the variables and interactions this gets really difficult. I'd imagine some sort of Design of Experiments methodology combined with AI might be the only way to gather and analyze this data. And for all Microsofts pushing of Co-Pilot AI for end user stupid shit where we would all be better off without it, they can't seem to use AI properly for something where it would really make sense.

For whatever reason, the Linux kernel does a better job with all of these complex variables, to the point where many CPU tasks where there is resource constriction actually run faster in a Windows VM in KVM on a linux host than they do in Windows themselves, as the linux Kernel is pretty good at sorting threads on one half of SMT enabled cores instead of bombarding both logical cores. I don't think that can be credited to DoE and AI though. Probably just some well through out algorithms. But this was back before the unholy mess of combination of Dynamic Clock and Voltage P-Cores with SMT and E-Cores without. I have no idea how the Linux and Windows schedulers compare on these highly complicated 12th, 13th and 14th gen multi-mode CPU's.

I think Intel expected Microsoft to step up to the plate and optimize the shit out of their scheduler, and they just didn't. And this is why Intel decided to move away from SMT to make the problem easier to solve. This is probably a missed opportunity IMHO, and it is more based on Microsoft dropping the ball than on INtel themselves (but to be fair, the best hardware in the world is worthless if you don't have software hat can adequately run on it)

Yup. And I'm betting it's that the Linux guys had a better basis for design with where they started and all the oddball support, but that's my guess alone.

 

[Mega6]

Where AVX sits now... is AMD has Zen5 running a double pumped AVX512 path, that eats Intels comparable server solutions for breakfast.

Zen4 is AVX256 ie double pumped AVX512.
Zen5 is native 512AVX - which is where a lot of the gains over Zen4 came from.

 

[Spirit_Retro]

The CHIPS act keeps them from selling the fabs, pretty sure.

Nah- they can sell to another American company with government oversight.

Spring I think- Intel will begin to divest itself.

 

[Zarathustra[H]]

100% agree. He wasn't good for RTG and I think Pat probably saw he wasn't going to be good for Intel GPU. Raja oversaw what...Polaris, Vega, maybe Navi? Radeon is still floundering in my opinion. I think the only good generation Radeon has had in recent memory is RDNA2.

I think AMD was about to catch up with the Radeon RX 6000 series (RDNA 2)

It would have been a smashing success. Except for halfway through its development cycle Nvidia absolutely blindsided them with RayTracing, and sent them back to the drawing board.

My RX 6900xt (granted a hefty binned overclock release with a highly binned XTX die, custom power stage and a waterblock, the XFX Speedster Zero EKWB Edition) had the performance crown among contemporary GPU's when it came to traditional raster graphics (which when it was being developed was all there was). Except, if you turned on even the slightest amount of RT it turned into a 20fps slide show.

I don't know whether to praise or criticize Nvidia for that one. They certainly completely overturned the marketplace with the RTX launch, and successfully kept it relatively quiet to the point where it surprised almost everyone, especially AMD.

I tend to think we didn't need RTX. More so than end users it is really a huge gift to developers. Rather than having digital artists pain stakingly work on light and shadow maps they can just throw a camera and some light sources in there and have the scene be rendered. The user winds up paying the bill from both a "they need a 4x more powerful GPU than before" perspective, and an electric bill perspective.

Raster randering was within 95%-98% of what RT could do, when RT launched. After RT launched, Raster quality in new titles dropped (due to developers not spending as much time/money on it) so if you wanted the best looking game, you needed RT, and that meant you needed nvidia.

This absolutely killed AMD's Radeon performance, and thus sales.

It really sent AMD back to the drawing board.

The 7000 series improved a little upon the 6000 series on the RT front, but not enough to close the gap. There is only so much you can do with an existing design. You really need a groun up design to implement RT well. The 8000 series supposedly improves RT performance even further (if rumors are to be believed, but you never know with those) but at the same yime AMD has decided to abandon the high end. At least this gen. Figures.

RDNA2 would ahve been the crowning achievement of AMD's effort to "catch back up" in GPU's, if Nvidia hadn't upended the market with RTX.

From the perspective of the last truly good Radeon launch? I'd argue it was GCN (Graphics Core Next) in the Radeon HD7970 and subsequent r9 290x. Those cards held their own against Nvidia's high end, even the 680 that launched after it. (at least if you exclude the Titan that launched in 2013, but that was a whole new $1000 ball game, insane at the time. Of course I bought one)

 

[Mega6]

Nah- they can sell to another American company with government oversight.

That is the only way - to sell less than half of the fab interest to an American company and still profit from the CHIPS act.

 

[Zarathustra[H]]

Totally agreed. Gelsinger's belief (that he came out a few months ago with) was that the future of Datacenter workloads was all E cores and GPUs. That works for very specific solutions, of course (AI among them), but is gonna SUCK for big DB/SAP/etc, which still back all those thousands of app servers on the front end. Software isn't that massively parallel yet, and containerization hasn't taken off enough to make it efficient yet except at the hyperscaler level, and that's abandoning on-prem traditional workloads (which aren't going away, and are often coming back from cloud) to AMD.

People in business often get tunnel vision when it comes to the latest and greatest trendy thing. Right now that is AI.

The question is, was he just talking the game in order to pump up investors and make them confident that Intel was pursuing a slice of the AI money pie, or had he actually convinced himself that this was the only way to go, forgoing all other options in Enterprise?

For my personal home lab (not AI), an all E-Core design would actually be pretty great, but yeah, my home lab is not the market as a whole.

 

[Zarathustra[H]]

That is the only way - to sell less than half of the fab interest to an American company and still profit from the CHIPS act.

There are vanishingly few American companies which I think have both the means and the desire to pull that off.

Qualcomm was one, but they apparently decided the additional government oversight wasn't worth it.

Apple or Google might be interested. Apple certainly has the volume to benefit from that. Not sure if Google does. They have custom silicon (phone SoC's for Pixel phones, Coral TPU's and a bunch of stuff for their own datacenters) but I'm not sure the volume is there. Nvidia maybe? Would be really hilarious if AMD gobbled up Intels fabs


View: https://www.youtube.com/watch?v=XFagogEOZz8

Though I question if they have the means, maybe in a stock trade deal of some sort.

That said, AMD does reportedly have over $6B in cash on hand. I wonder how much it would take to buy Intel's fabs. A mixed cash/stock option might actually do it.

Who else would do it? One of the major defense companies (RTX (formerly Raytheon)? Lockheed Martin? Northrup Grumman?), to secure a source of silicon for chips for radar systems, jet fighters and missiles?

Those are really the only players I can think of.

 

[zerogg]

I'm really disappointed in how their new GPU efforts went. I thought they might actually have a change to become a viable option

 

[lopoetve]

Nah- they can sell to another American company with government oversight.

Spring I think- Intel will begin to divest itself.

We'll see.

People in business often get tunnel vision when it comes to the latest and greatest trendy thing. Right now that is AI.

The question is, was he just talking the game in order to pump up investors and make them confident that Intel was pursuing a slice of the AI money pie, or had he actually convinced himself that this was the only way to go, forgoing all other options in Enterprise?

For my personal home lab (not AI), an all E-Core design would actually be pretty great, but yeah, my home lab is not the market as a whole.

Agreed on all counts - except the way he described it was definitely the only way to go in the long run was piles of low-power E cores... which is just nuts.

 

[Sir Beregond]

I think AMD was about to catch up with the Radeon RX 6000 series (RDNA 2)

It would have been a smashing success. Except for halfway through its development cycle Nvidia absolutely blindsided them with RayTracing, and sent them back to the drawing board.

My RX 6900xt (granted a hefty binned overclock release with a highly binned XTX die, custom power stage and a waterblock, the XFX Speedster Zero EKWB Edition) had the performance crown among contemporary GPU's when it came to traditional raster graphics (which when it was being developed was all there was). Except, if you turned on even the slightest amount of RT it turned into a 20fps slide show.

I don't know whether to praise or criticize Nvidia for that one. They certainly completely overturned the marketplace with the RTX launch, and successfully kept it relatively quiet to the point where it surprised almost everyone, especially AMD.

I tend to think we didn't need RTX. More so than end users it is really a huge gift to developers. Rather than having digital artists pain stakingly work on light and shadow maps they can just throw a camera and some light sources in there and have the scene be rendered. The user winds up paying the bill from both a "they need a 4x more powerful GPU than before" perspective, and an electric bill perspective.

Raster randering was within 95%-98% of what RT could do, when RT launched. After RT launched, Raster quality in new titles dropped (due to developers not spending as much time/money on it) so if you wanted the best looking game, you needed RT, and that meant you needed nvidia.

This absolutely killed AMD's Radeon performance, and thus sales.

It really sent AMD back to the drawing board.

The 7000 series improved a little upon the 6000 series on the RT front, but not enough to close the gap. There is only so much you can do with an existing design. You really need a groun up design to implement RT well. The 8000 series supposedly improves RT performance even further (if rumors are to be believed, but you never know with those) but at the same yime AMD has decided to abandon the high end. At least this gen. Figures.

RDNA2 would ahve been the crowning achievement of AMD's effort to "catch back up" in GPU's, if Nvidia hadn't upended the market with RTX.

From the perspective of the last truly good Radeon launch? I'd argue it was GCN (Graphics Core Next) in the Radeon HD7970 and subsequent r9 290x. Those cards held their own against Nvidia's high end, even the 680 that launched after it. (at least if you exclude the Titan that launched in 2013, but that was a whole new $1000 ball game, insane at the time. Of course I bought one)

Considering the GTX 680 was midrange GK104 silicon (294mm2 if I recall), it wasn't surprising that GCN aged far better.

RT pisses me off. Considering it first released in 2018, almost a whole 7 years ago, you'd think that mainstream cards of today would be able to have a good RT experience, but do you really get one if you spend under $400 for a card? The answer is no. Just about 4 gens into RTX, that really just kind of sucks. Meanwhile the entire market just keeps going up. Sure there are other factors at work between inflation, wafer cost increases, etc. but Nvidia's genius marketing of ray tracing IS RTX and completely dominating the market isn't helping. As far as I am concerned, RT is still a rich man's game and it's come at the expense of everyone else in the gaming market who used to occupy the mainstream segments. Nvidia may have the best tech for RT, upscaling, and frame gen, but the dark side is that its absolutely led to lazier game devs, worst looking raster, and overall the market is worse for it in my opinion.

Anyway, to bring this back to Pat Gelsinger, Intel Arc was too little too late to have any real lasting impact. Probably a rope Intel pissed up and are regretting spending so much money on it. I really do hope Battlemage ends up being good for them and us, but I doubt it will be if it's just about to be 2025 and their performance targets are cards that released in 2022 and 2023. I suppose if the price is right and drivers are much much better, maybe Battlemage has a better chance. Especially if Nvidia just starts price gouging again like they did for the 40-series while leaving the under $500 mainstream market in a stagnate state.

 

[ChadD]

Nah- they can sell to another American company with government oversight.

Spring I think- Intel will begin to divest itself.

No one wants Intels fab business. Intel with this latest cash up from Chips will not be spinning the fab off into their own company which was Pats plan. The gov saved Intel here. Once Pat spun the fabs off... someone would have made an offer to Intels share holders for everything else.

 

[ChadD]

I'm really disappointed in how their new GPU efforts went. I thought they might actually have a change to become a viable option

I guess it might be telling that... Pat announces his retirement the day before a battle mage release. :/ I mean Intel dropping a BM on the market... wouldn't have saved Pat.

 

[Zarathustra[H]]

Considering the GTX 680 was midrange GK104 silicon (294mm2 if I recall), it wasn't surprising that GCN aged far better.

RT pisses me off. Considering it first released in 2018, almost a whole 7 years ago, you'd think that mainstream cards of today would be able to have a good RT experience, but do you really get one if you spend under $400 for a card? The answer is no. Just about 4 gens into RTX, that really just kind of sucks. Meanwhile the entire market just keeps going up. Sure there are other factors at work between inflation, wafer cost increases, etc. but Nvidia's genius marketing of ray tracing IS RTX and completely dominating the market isn't helping. As far as I am concerned, RT is still a rich man's game and it's come at the expense of everyone else in the gaming market who used to occupy the mainstream segments. Nvidia may have the best tech for RT, upscaling, and frame gen, but the dark side is that its absolutely led to lazier game devs, worst looking raster, and overall the market is worse for it in my opinion.

Anyway, to bring this back to Pat Gelsinger, Intel Arc was too little too late to have any real lasting impact. Probably a rope the pissed up and are regretting spending so much money on it. I really do hope Battlemage ends up being good for them and us, but I doubt it will be.

Yeah, I tend to think it was shitty of Nvidia, but you have to give them credit for reading the market.

They knew Devs would embrace RT because it reduces their workload and costs, so if they launched the first feasible RT product, they knew it would be embraced by devs, and it would lock out AMD's resurgence as they wouldn't have the tech.

They also made thousands upon thousands of people who otherwise would have been happy to spend $200 on a low to mid end GPU now feel like they weren't getting the full experience if they didn't get an RT capable card, and made them willing to move up in the market to more expensive models.

They might lose a handful of people in the process, sure, but they knew that there were enough die-hard "I will spend whatever it takes on gaming out of FOMO" types to make this happen.

They signle handedly moved the entire GPU market upscale while locking out the competition for at least a couple of generations.

Win-win (for Nvidia) right?

Oh, and as far as playability goes. Didn't you hear? You are supposed to embrace fake pixels (upscaling) and fake frames (frame gen) now. Come on! All the cool kids are doing it!

Pay more, get less. It's that simple!

 

[Sir Beregond]

Yeah, I tend to think it was shitty of Nvidia, but you have to give them credit for reading the market.

They knew Devs would embrace RT because it reduces their workload and costs, so if they launched the first feasible RT product, they knew it would be embraced by devs, and it would lock out AMD's resurgence as they wouldn't have the tech.

They also made thousands upon thousands of people who otherwise would have been happy to spend $200 on a low to mid end GPU now feel like they weren't getting the full experience if they didn't get an RT capable card, and made them willing to move up in the market to more expensive models.

They might lose a handful of people in the process, sure, but they knew that there were enough die-hard "I will spend whatever it takes on gaming out of FOMO" types to make this happen.

They signle handedly moved the entire GPU market upscale while locking out the competition for at least a couple of generations.

Win-win (for Nvidia) right?

Oh, and as far as playability goes. Didn't you hear? You are supposed to embrace fake pixels (upscaling) and fake frames (frame gen) now. Come on! All the cool kids are doing it!

Pay more, get less. It's that simple!

Oh absolute genius move by Nvidia. The fact they could find dual purpose use for Tensor in AI and pro markets and being able to still use that in the consumer gaming space? Genius. And then raising the cost of entry? Genius. Marketing in a way that RTX is ray tracing, and you have to have it? Genius. Timing perfectly with crypto and then pivoting immediately to AI when crypto dropped? Another genius move. Getting all the dev buy-in to the tech? Genius.

Everything you said is correct. Yet I'd say the gaming market is worse than ever, and so is the consumer PC graphics space (for price anyway), taking COVID/crypto out of the equation.

 

[chithanh]

Considering in hindsight the giant success of GPU parallel compute not sure if this is a fair assessment, the power of ultra parallel computing was something people wanted a lot (even if they do not cared that it is parallel or not), the proposed solution for it was not the good (or lost to gpu for an other reason).

Kinda, but not really?

In HPC, Itanium was never a big success. Intel managed to claim the #2 spot in one Top500 list before GPU compute became a thing. But other systems were cheaper, faster, and more efficient, and so Itanium was short-lived even there.
Itanium did manage to carve out a niche in mission-critical high-reliability systems (running OpenVMS). But as soon as Intel's x86 CPUs adopted the reliability features then it was game over and everybody just used Xeons.

 

[zerogg]

Kinda, but not really?

In HPC, Itanium was never a big success. Intel managed to claim the #2 spot in one Top500 list before GPU compute became a thing. But other systems were cheaper, faster, and more efficient, and so Itanium was short-lived even there.
Itanium did manage to carve out a niche in mission-critical high-reliability systems (running OpenVMS). But as soon as Intel's x86 CPUs adopted the reliability features then it was game over and everybody just used Xeons.

That reminds me of another CEO epic fail, Carly dumping Alpha processors and going all in on Itanium

 

[chameleoneel]

Holy cow. I didn't see that coming at all

I was actually really surprised that he seemed to survive this year.

 

[HockeyJon]

Well lets see what people say now that he is gone. Pat is about to get blamed for the shit show the next CEO inherits.

Intels downfall has been a long time coming no doubt it long predates Pat as CEO. In fact I would put the real beginning of the end back to his time as CTO. He is the guy that set 10nm fab targets. He is not the only player involved but he was the head of tech when intel set 10nm targets. I will grant that was 2 gens ahead of where they were. Still he was the guy in charge when the targets were made. It might not be his fault that Intel didn't correct earlier after he left.... still he was a bit part of them setting a stupid aggressive path that obviously didn't pan out. IMO though the issue was not Pat it was a deep ingrained issue. Intel in the early 2000s seems to have had a culture shift. My feeling is all the old guard left. The early 60-70s engineers retired or moved on. The new mid Intel lifespan guys like Pat that started in '79... and was a part of the hey days of 386 486. Intel dominating all... started to believe the hype Intel created for themselves. Pat still talks like that. When he was made CEO with his first public words I knew how it was going to go, "INTEL will win be cause Intel is Intel and Intel engineers can do antyhing. INTELLLLLL" Paraphrasing. lol I am pretty sure Pat was passed over for CEO the first time (which lead to him leaving Intel) as there may still have been a few old guard on the board who knew better then make a Intel Zellot CEO. I really believe Pat can't (or at least couldn't) forsee any scenario where Intel wouldn't WIN at anything they did. The idea that the competition might kick you in the ass never crossed his mind... at least until the last year or so.

On that note, remember this?

Intel boss Pat Gelsinger calls Arm's PC threat "insignificant"

https://www.techspot.com/news/100643-intel-boss-pat-gelsinger-calls-arm-pc-threat.html

 

[Lakados]

On that note, remember this?

Intel boss Pat Gelsinger calls Arm's PC threat "insignificant"

https://www.techspot.com/news/100643-intel-boss-pat-gelsinger-calls-arm-pc-threat.html

He's not wrong... Yet.
Nothing a threat until it is.

 

[LukeTbk]

Qualcomm arm pc can well be insignificant, if Nvidia-AMD start making them.... that's become quite the stronger statement and that what the news was about.

 

[ChadD]

On that note, remember this?

Intel boss Pat Gelsinger calls Arm's PC threat "insignificant"

https://www.techspot.com/news/100643-intel-boss-pat-gelsinger-calls-arm-pc-threat.html

You know what is hillairous.... there are quotes of him from his time as CTO. Where he talked about x86 mobile.
If Bill gates hadn't once said 640k was enough for everyone. Pats saying in the early 2000s "One day everyone will be on the internet 24/7, but that day is long from now" might go down as one of the most best tech quotes ever. He was talking about x86 mobile and said essentially oh ya we will have that market, Intel is the future... oh the future of everyone having an internet computer in their pocket ya that is years off I tell ya. Then like 2 years later Jobs releases the iphone... which was probably in development when Pat was blathering about x86 mobile and the FAR future. lol
Pat has some comical quotes if you dig.

The problem with bringing Pat on as CEO, really was his history with Intel. Pat didn't have a great track record. Itanium, Larabee, Xscale, Intels mobile missteps. Pat was part of the problem at Intel before he left, I guess it shows how incompetent their board is that they believed he would somehow be the solution.