Discussion Intel current and future Lakes & Rapids thread

[Geddagod]

What is the board's opinion of Daniel Nenni (SemiWiki founder)? I know MLID is not popular (for understandable reasons) but he somehow gets these really credible guests on his podcasts that I think are worth a listen.

Anyway, his statements on Intel's foundry services is interesting. As a summary: He says Intel 4 & 3 are both already solid. He seems confident in 18A too, although claims it's success is dependent on the PDK they offer. He thinks IFS offering for Intel 3 is basically DOA against TSMC since they can't compete against their current ecosystem and PDK. Seems Intel's issue with MTL-S (presumably being cancelled) aren't necessarily related to the manufacturing side but rather first adoption of chiplets and early EMIB bugs.

He seems pretty cool. Was tired of studying so I decided to give part of the video a watch before I go to sleep. Ye he seems pretty confident about Intel 4's health, something interesting though I heard Tom mention was that there were 12 steppings? of MTL. Sounds crazy, but idk if Tom lifted that off SPR since apparently it also had 12 steppings. I mean I wouldn't be surprised if MTL required a bit more work, but idk if it was as bad as SPR.
He thinks TSMC 2nm is ahead of Intel 18a but are comparable, and those two are comparable to Samsung 3nm
He disagreed with Tom saying that Nvidia was going to cancel MCM and only focus on mid range of gaming chips lmao, so I'm glad I heard that
The part where he talked about how demanding Apple was for TSMC and how Apple, AMD, and Xylinx? I think it was, played the largest role in TSMC developing packaging tech was interesting
Finfet is much easier to design for than gate all around, so perhaps that's part of the reason Intel is able to push out ARL on TSMC 3nm faster than they can get ARL on 20A (second part is speculation on my end)
Intel 7 ultra is ridiculously expensive according to Tom, more expensive than TSMC 7nm but also more expensive than TSMC 5nm!
Daniel seems insistent that large, chiplets for expanding max die size (Intel possibly Nvidia) is worse than AMD's method of many small chiplets with better yields
Based Intel 18A, cringe Intel 3, Samsung sucks in general
Also MTL doesn't use EMIB afaik just foveros
Interesting vid.

 

[ashFTW]

Well that would certainly do it. Interesting that he lists "memory channels". Seems to imply that DMR will move up to 16c on the AP socket.

Could he be referring to DDR6 instead?

 

[Exist50]

something interesting though I heard Tom mention was that there were 12 steppings? of MTL

I doubt it. That many steppings takes way too long, as we say with SPR. Timeline alone doesn't match. And MTL might not be great, but it's hardly a SPR level dumpster fire.

Intel 7 ultra is ridiculously expensive according to Tom, more expensive than TSMC 7nm but also more expensive than TSMC 5nm!

Wouldn't be surprised. They're not offering it for IFS for a reason.

Could he be referring to DDR6 instead?

Possibly, but I think it's too soon for DDR6. Maybe the gen after, but I think DDR6 is pegged for late in the decade. We'll see though.

I guess another possibility is MRDIMM support, but I'm kind of hoping we don't have to wait for DMR for that.

 

[aigomorla]

Could he be referring to DDR6 instead?

I heard they are trying to incorporate HBM memory on CPU now.
I know Intel has a hybrid HBM memory chip.

Speed-Up Memory Bandwidth Bound Workloads with Intel® Xeon® Processors

Discover Intel® Xeon® processors, codenamed Sapphire Rapids with high bandwidth memory (HBM) and how it is solving industry memory issues—first with PMem and now with HBM.

www.intel.com

 

[A///]

samsung wants to deliver the first ddr6 chips by 2026. you're looking at 2027 before regular consumers get anything maybe early 28. it'll be around some time in 2026 for certain industries just like ddr5. in other news intel is allegedly working on bringing back avx512 to their consumer lineup. very cool stuff in the next few years.

 

[Exist50]

in other news intel is allegedly working on bringing back avx512 to their consumer lineup

According to whom?

 

[A///]

According to whom?

rumor on wctech referencing some aws ec2 engineer from amazon in Germany who says it's coming along slowly and some other commentary from roacho about 256 plus which makes no sense to me other than it being a similar double pump like amd's. i don't involve myself in these matters because it's of no importance to me but my basic understanding of amd's approach is that their double pump method avoids issues that intel's singular 512 unit would encounter and cause other issues. idk of the drawbacks of amd's design and there may not be that concern the regular user and this might be intel having their moment learning from a compeittor. the username on twitter is longorn at never_released. the name sounds familiar other than it being an ms code name for vista back in the day but I've seen him referenced in rumors before.

 

[Exist50]

rumor on wctech referencing some aws ec2 engineer from amazon in Germany who says it's coming along slowly and some other commentary from roacho about 256 plus which makes no sense to me other than it being a similar double pump like amd's.

Raichu knows what he's talking about. He's the one to pay attention to.

my basic understanding of amd's approach is that their double pump method avoids issues that intel's singular 512 unit would encounter and cause other issues

The problem with AVX512 isn't terribly complicated. Current Atom uarchs have a 128b datapath. "Cracking" a 512b op into two 256b ops (as AMD does) has its challenges, but is very doable. But cracking it into four 128b components scales the overhead significantly. It's non-linear. So to support AVX512, Atom has to either move to a 256b datapath, or spend a lot of effort and overhead to make that 4:1 cracking work.

That 2:1 "cracking" is what Atom does today for 256b AVX ops, btw, so I'm not sure why people act like AMD invented the concept. It works well for their purposes, but it's hardly a new idea, much less to Intel. Nor is it faster or more efficient than native width execution.

 

[A///]

Raichu knows what he's talking about. He's the one to pay attention to.

The problem with AVX512 isn't terribly complicated. Current Atom uarchs have a 128b datapath. "Cracking" a 512b op into two 256b ops (as AMD does) has its challenges, but is very doable. But cracking it into four 128b components scales the overhead significantly. It's non-linear. So to support AVX512, Atom has to either move to a 256b datapath, or spend a lot of effort and overhead to make that 4:1 cracking work.

That 2:1 "cracking" is what Atom does today for 256b AVX ops, btw, so I'm not sure why people act like AMD invented the concept. It works well for their purposes, but it's hardly a new idea, much less to Intel. Nor is it faster or more efficient than native width execution.

to a degree he's made some bs before but ignoring that for now.

no one here claimed amd invented anything. hard to invent something intel already did no? the way i understand amd's implement is they have 4x 256b units, how the processor handles 1 to all is something I'd have to brush up on because x86 is a bore to me. but my understanding is two units form and process the split data and another portion of data goes to the other two units. the double pump effect. from my very likely dated understanding now is that amd's implement is better than intel's because they manage to keep clocks higher and the temps lower when processing this data that calls for avx512 execution. or I'm terribly wrong and need to brush up on writeups from the last 8 months.

from longhorns deviously sly comment reply on the twitter and raocho's whatever's my understanding is based on raicho's is that intel has taken what amd has done and found an even better implement. which II wouldn't put it past intel to do. if you recall in the 90s and early 2000s their cat and mouse game of one upmanship was a marvel and a wallet drain for some of us.

now here's where I inject my personal opinion. what for? what the hell is the point? Up until rocket lake the only time i'd heard of avx512 was for machine learning and you would be better off with gpus for that. so now all of a sudden every jack terry and nigel want avx512 in their system, yes and I know I've been pushing for intel to bring it back but mostly because I'm tired of reading the whining about it or the taunting from amd folks. I don't know what benefit it gives to normal joes and janes who may only utilise it for old console emulation or very niche software they won't be buying in the first place.

this move seems like a first step of intel implementing altera ip into their mass general chips as amd plans on implementing xilinx ip beginning with zen 5 or 6. this reminds me of how people just had to have windows xp 64 bit back in the day to run more ram because they were "power users" but would whine none stop when things didn't behave correctly because wind xp 64 bit was a half baked rotten potato to begin with.

 

[igor_kavinski]

this reminds me of how people just had to have windows xp 64 bit back in the day to run more ram because they were "power users" but would whine none stop when things didn't behave correctly because wind xp 64 bit was a half baked rotten potato to begin with.

But 64- bit did become a necessity. People interested in AVX-512 want the same. They want developers to explore it more and make our computations faster. If AVX-512 is useless, AMD shouldn't have implemented it. They did so it means there is at least SOME benefit to it, regardless of the current limited use cases. More than that, it might be the future where software utilizing AI computations may not work without AVX-512 or work in really horrible compatibility mode at a snail's pace.

 

[A///]

But 64- bit did become a necessity. People interested in AVX-512 want the same. They want developers to explore it more and make our computations faster. If AVX-512 is useless, AMD shouldn't have implemented it. They did so it means there is at least SOME benefit to it, regardless of the current limited use cases. More than that, it might be the future where software utilizing AI computations may not work without AVX-512 or work in really horrible compatibility mode at a snail's pace.

it did? when igor? in 2009 when windows 7 came out? xp x64 came out in 2005 and earlier if you had a ms subscription. it was abysmal, windows vista 64 was nothing to write home about either. windows 7 is an improved vista and even then in 2009 most computers were still coming with 4 gb of RAM and only did 8 gb become the common 2 years later with sandy bridge.

you say this but what benefit to the masses? who in the general public is using software that is using that instructional set? I want to know. avx512 is so niche and under utilised by general software that it hasn't gone much of anywhere ever. avx512 hardware on a cpu is nice but purpose built accelerators will outpace it while sipping on less power. the cons of avx512 and latency are why its main competitor a green company run by an egomaniac with an obsession for leather jackets is outpacing cpu use in deep learning. there's more efficient ways to process parallel data.

 

[igor_kavinski]

you say this but what benefit to the masses?

Excel. Probably the most used Office software in the world. I bet it could help speed up processing huge amounts of data.

 

[A///]

Excel. Probably the most used Office software in the world. I bet it could help speed up processing huge amounts of data.

excel already uses language models that don't need avx512 to speed up workflows. you asked me about this weeks ago.

 

[igor_kavinski]

windows vista 64 was nothing to write home about either.

My first brush with 8GB RAM was with Vista x64. It felt so good to be released from the shackles of low RAM availability. Opened so many browser tabs on that Core 2 Quad PC. In 2008, that was heaven for me.

 

[igor_kavinski]

excel already uses language models that don't need avx512 to speed up workflows. you asked me about this weeks ago.

Yes but I meant the user being oblivious to AVX512 and Excel using it automatically and speeding up the slow formulas. It would be like a fresh breeze. And then people getting curious. Hey, my laptop isn't this fast. What's special about this plain jane office PC? Oh, AVX512. Wow. I must get one for my home too!

 

[A///]

My first brush with 8GB RAM was with Vista x64. It felt so good to be released from the shackles of low RAM availability. Opened so many browser tabs on that Core 2 Quad PC. In 2008, that was heaven for me.

in 2011 I was given a work laptop with 1.5 GB....

 

[igor_kavinski]

in 2011 I was given a work laptop with 1.5 GB....

How long did you suffer with it?

 

[A///]

Yes but I meant the user being oblivious to AVX512 and Excel using it automatically and speeding up the slow formulas. It would be like a fresh breeze.

to repeat myself, the new models don't use avx512. they're plenty fast and crunch the data and do the work it would take a half hour or more to do in minutes. Microsoft rewrote portions of the excel engine to make better use of current longstanding older instruction sets. this was microsoft being smart. to make avx512 be a thing for excel it would need a complete rebuild. what's the point?

 

[A///]

How long did you suffer with it?

About 2 or 3 years. It ran Windows 7 Professional and you could type several sentences fast before they registered and appeared on the screen due to how slow the 5400 rpm drive was. My air gapped system was a state of the art computer for the time, that otoh was a pile of junk.

 

[igor_kavinski]

what's the point?

It's not just data analysis. Simple formulas like vlookup can be pretty slow on large data sets. They should see an improvement with the ability to crunch more data in fewer clock cycles. I'll shut up if you have a technical reason for why that is not feasible. So far, it just seems laziness on Microsoft's part.

 

[igor_kavinski]

About 2 or 3 years.

I would have bought my own laptop to stay sane if my requests for increased RAM had been denied/ignored.

 

[A///]

I would have bought my own laptop to stay sane if my requests for increased RAM had been denied/ignored.

No personal devices allowed. The workstation was a dual processor very fast Dell not the most cores but speed was the essential factor. 3.7 or 3.8 ghz. The entire machines each cost in the mac pro of then territory. You either got that or these high end fast all in ones that weighed a ton but were very easy to work on if the Dell guy couldn't show up to change or fix something. Looking back they were outclassed by 5th or 6th gen intel mainstream processors.

 

[eek2121]

Sounds like Intel is still having challenges with Chiplets. EMR is said to contain just 2: https://www.semianalysis.com/p/intel-emerald-rapids-backtracks-on

I find this to be interesting. It would explain a LOT of the rumors of the troubles surrounding Sapphire Rapids prior to launch and also the current rumors with Meteor Lake difficulties.

I guess I'm just curious why Intel is having such a hard time doing something that AMD has seemingly done so easily.

 

[Ajay]

Intel said clearwater forest in 2025 is the lead product for 18A so it can't be used for lunar lake and with intel 3 being used only in server products I think it's safe to assume lunar lake is going to be on TSMC N3.

I thought Intel 4 was MTL only and Intel 3 was full stack. In which case Intel would be manufacturing client CPU(SOCs) on their N3 node. Maybe we should start using IN3 or something to make it easier to distinguish from TSMC .

 

[Ajay]

it did? when igor? in 2009 when windows 7 came out? xp x64 came out in 2005 and earlier if you had a ms subscription. it was abysmal, windows vista 64 was nothing to write home about either. windows 7 is an improved vista and even then in 2009 most computers were still coming with 4 gb of RAM and only did 8 gb become the common 2 years later with sandy bridge.

XP x64 had a solid base (taken from server 2003 R2, IIRC). But, it was a nightmare in terms of driver support. The initial release of Vista was pretty bad - too many changes under the hood and with the UI. It was much better after SP1 and pretty spot on after SP2 - but it was a memory hog.