Intel Might Drop 10nm Node for Desktop Processors, 14nm Until 2022?

[The Benjamins]

3 minutes ago, porina said:

Yes, but we have to consider architecture and process as two parts. Intel 10nm sucks, but what would that same architecture do on 14nm?

isn't it a lot of work to get a arch on a new node, AFAIK its not some copy pasta.

 

would it be worth the cost to re-engineer ice lake arch to 14nm for only a 1-2 year sales window?

 

IDK what they are going to do, all I know is they will not like it.

[S w a t s o n]

4 minutes ago, porina said:

*raises hand* I bought one. Well, I got the 5675C when it was current, and only more recently did I manage to find a cheap 5775C to go with it. The initial transition to 14nm was also a problem for Intel, very different from the 14nm we have today. Having said that, the 5675C and 5775C were odd balls in that they were more or less the mobile parts done on desktop. They had the L4 cache of Iris Pro, and this made them very interesting CPUs. Skylake was better on architecture, but that cache meant Broadwell desktop parts (not HEDT) punched far above their weight in memory intensive compute applications.

 

I actually would like to see that. Take Ice Lake, put it on a desktop package and let me overclock it. 

Yea I doubt there's anything close to the quality of the 5775C silicon yielding yet. If 5775C was truly mobile grade silicon then we'd be looking for the H series laptop parts to be around and that's not happening yet. Just U for now and everyone keeps saying it's not going to scale much further other than for server. The real oddball to me was that no one seemed to question that those parts and (R variants) were basically all of 5th gen desktop, no K skus or nothing. And nearly no one could get one. You are the first person I even sort of know that has one, let alone even knows about 5775C and brethren. Almost no one even stopped to think about it, including hardware outlets. And then intel quickly rolled out 6th gen.

[RejZoR]

43 minutes ago, jstudrawa said:

Intel's 14nm chips are doing very well, even against 7nm.  

 

I'm curious to see what benefits 7nm will give Intel, if they can continue to refine these processes so well or if there is a diminishing returns as we get smaller on current materials.

 

AMD and Intel are going to bump up against a hard limit soon, making them basically equal.  Who finds the next perfect material first?

Either material or whole new approach to chip design. Just look what GPU designers did when they were forced to use 28nm. Instead of just shrinking things and cramming more shit into the chip, they actually made chip's processing more efficient. And what boosts there were during that time. If everyone could just go to 14nm I don't think we'd seen such huge leaps. Another good example is Ryzen. It's a very radical approach to computation units which yields amazing results. Sure more cores helps and so does smaller node, but most of gains came from well designed core(s).

[porina]

25 minutes ago, The Benjamins said:

isn't it a lot of work to get a arch on a new node, AFAIK its not some copy pasta.

 

would it be worth the cost to re-engineer ice lake arch to 14nm for only a 1-2 year sales window?

 

IDK what they are going to do, all I know is they will not like it.

That's the position they're in though. They can either keep re-hashing Skylake, backport Sunny Cove, or... I'm not sure what other options there realistically are. For sure it is not a trivial exercise. If they were optimistic 10nm would be working by now it may be late to start the process. If they were pessimistic, maybe they already started doing this in the background as a backup plan.

 

19 minutes ago, S w a t s o n said:

The real oddball to me was that no one seemed to question that those parts and (R variants) were basically all of 5th gen desktop, no K skus or nothing. And nearly no one could get one. You are the first person I even sort of know that has one, let alone even knows about 5775C and brethren. Almost no one even stopped to think about it, including hardware outlets. And then intel quickly rolled out 6th gen.

There was no need for a K sku since the C parts were unlocked anyway. It was close to a paper launch and extremely late, I think only months before Skylake. I actually bought my 5675C after I had already bought a 6700k. Reviews at the time did concentrate on the iGPU performance. I was kinda hoping to set some overclocking records there, but then Skylake Iris Pro parts that came later had more CUs and kill that part of my plan. I still hope to set some records for that generation CPU, hence only getting a 5775C last year, as the HT does make a difference there.

 

12 minutes ago, ONOTech said:

It sounds like the 10nm yields are poor (<30%??, that's so low!). Anyone know why the process has been so difficult for Intel?

Their original 10nm process was very aggressive and quite simply it blew up in their face. Just didn't work. They were forced to nerf it, but even then it wasn't smooth sailing.

[Orangeator]

I pray Intel gets their sh*t together soon. We don't need AMD taking Intel's dominate position they were in before Ryzen... 

[The Benjamins]

13 minutes ago, Orangeator said:

I pray Intel gets their sh*t together soon. We don't need AMD taking Intel's dominate position they were in before Ryzen... 

I don't think its a issue just that AMD will take a good chunk of market share but around 2022 intel should be fine.

[mcbaes72]

Intel responded to rumors... 

 

Intel: "Yes, There Will be 10nm Desktop CPUs"

[pas008]

2 hours ago, GoldenLag said:

So how is intel going to make GPUs on 10nm if they cant create large enough dies for CPUs?

 

Perhaps 10nm is better at big dies and lower clocks and voltages?

you think thermal density has been causing them the problems

all seem to be clocked way lower

 

maybe they know they need to keep the clocks now to remain competitive

 

and part of me thinks they are going to use their own chiplet design very soon(omni directional emib/foveros) but try to keep clocks up and figure away to raise ipc for time being

along with trying to fix some of these security issues

[GoldenLag]

54 minutes ago, Orangeator said:

I pray Intel gets their sh*t together soon. We don't need AMD taking Intel's dominate position they were in before Ryzen... 

dont worry, AMD lasted 8-5 years on practically pennies, and returned with Ryzen. 

 

and Intel is a lot wealthier than AMD, so i dont think you need to worry about Intel going away. they have a lot of time and money to get their shit together. 

[Zando_]

56 minutes ago, Orangeator said:

I pray Intel gets their sh*t together soon. We don't need AMD taking Intel's dominate position they were in before Ryzen... 

Yeah, keeping competition is way better for the industry. Would be interesting to see what AMD would do if they were the top dogs though. 

[GoldenLag]

Just now, pas008 said:

you think thermal density has been causing them the problems

all seem to be clocked way lower

unless the density is twice that of 7nm, i dont think so. 

 

7nm TSMC can get quite toasty around 4,5 ghz on Zen 2 and at that point thermal density becomes an issue. 

 

so maybe? the clockspeed regression is probably a factor of them increasing IPC in the design and smaller nodes simply not giving good clockspeeds. 7nm isnt very impressive, and 7nm+ i dont think gave much improvement in clockspeed either. 

 

[GoldenLag]

3 minutes ago, Zando Bob said:

Would be interesting to see what AMD would do if they were the top dogs though. 

removing non-x lower clocking parts perhaps?

 

their X-series or higher clocked variants are borderline scams tbh. 

[sazrocks]

2 hours ago, Princess Luna said:

  Hide contents

¯\_(ツ)_/¯

 

I get it you are the usual AMD fan but how about don't just go shooting random numbers just to try making a point your brand is better.

Where the heck is that graph coming from!? 400w from a 3900x!? I’d think that I know if my 3900x was kicking out 400W lol.

[Zando_]

2 minutes ago, GoldenLag said:

removing non-x lower clocking parts perhaps?

 

their X-series or higher clocked variants are borderline scams tbh. 

Aren't they just higher TDP variants with higher stock clocks? Don't see how that's a scam, it'd be more for people who don't want to touch the CPU at all. Given that people buy Intel K chips and never OC them at all (just buy them for the higher stock clocks), it's not too surprising. 

On the 2000 series they had XFR and PBO IIRC, I forget what/if there was a difference on the 1000 series. 

[Brooksie359]

2 hours ago, Taf the Ghost said:

The node yields. It's just the original yields were about .5%. Not 5% but .5% and they couldn't yield a iGPU at all. They've dialed back a lot of the technical advances and while they can show off silicon at the density they've claimed, reality is that it isn't that much more dense than 14nm and clocks a lot worse. And that was just to get the node out the door. Yields are still going to be terrible and a single server die likely is yielding maybe 20% per wafer. Maybe. Full die SKUs from Icelake Server parts are going to be extremely expensive and really only fit in a niche roll. 

 

What's more fun is Intel's 7nm is their first EUV node. And they're going to be doing a lot of layers with EUV. Which means the whole "7nm in 2022!" mantra may not happen. Even with two full years to get there. It will be some large, generational leaps, but only because their design teams haven't had a node to launch anything new on in ages.

They have been working on 7nm for awhile now alongside 10nm so much so that there were a couple times rumors went around that Intel would nump straight to 7nm because 10nm wasn't working out. I honestly don't doubt that they will hit 7nm in 2022. 

[GoldenLag]

4 minutes ago, Zando Bob said:

Aren't they just higher TDP variants with higher stock clocks?

enable PBO, and there really isnt much difference. and even without PBO there isnt a whole lot to see either. 

5 minutes ago, Zando Bob said:

Don't see how that's a scam

i guess im wrong to call it a scam, but from a value perspective, its bad and its marketing as somethign "extra" when its really not. 

6 minutes ago, Zando Bob said:

On the 2000 series they had XFR and PBO IIRC, I forget what/if there was a difference on the 1000 series. 

on the 1k series there wasnt much reason to buy X series, you could OC on the stock cooler to a similar level, even without OC the difference was nothing. worst offender was 1700 to 1800x. 

7 minutes ago, Zando Bob said:

On the 2000 series they had XFR and PBO IIRC,

the 2700x had some compelling reasons, such as the boost clock that you couldnt get by OC. other than that, the 6 core had no compelling reason to get the X variant. 

[porina]

13 minutes ago, sazrocks said:

Where the heck is that graph coming from!? 400w from a 3900x!? I’d think that I know if my 3900x was kicking out 400W lol.

That looks like it was total system power from the wall. Ball park 250W delta from idle to load, but that would include power supply (in)efficiency. Say 80% PSU, that's down to 200W for CPU + mobo VRM inefficiency + anything else using power under load like ram. CPU socket somewhere below that. Given 105W TDP Ryzens are 142W PPT at stock doesn't seem too insane.

[Taf the Ghost]

7 minutes ago, Brooksie359 said:

They have been working on 7nm for awhile now alongside 10nm so much so that there were a couple times rumors went around that Intel would nump straight to 7nm because 10nm wasn't working out. I honestly don't doubt that they will hit 7nm in 2022. 

They really should have had 7nm for 2020. As it stands, that's already going to be 2 years late, though they didn't start the cycle until a bit later, so it's technically not behind schedule from development reasons. That we know of. Yet.

[Zando_]

3 minutes ago, GoldenLag said:

enable PBO, and there really isnt much difference. and even without PBO there isnt a whole lot to see either. 

i guess im wrong to call it a scam, but from a value perspective, its bad and its marketing as somethign "extra" when its really not. 

on the 1k series there wasnt much reason to buy X series, you could OC on the stock cooler to a similar level, even without OC the difference was nothing. worst offender was 1700 to 1800x. 

the 2700x had some compelling reasons, such as the boost clock that you couldnt get by OC. other than that, the 6 core had no compelling reason to get the X variant. 

Aha. I knew the 2700X was solid (I had one), wasn't sure on the others. I've had an R5 1600 that did 4Ghz easy (and then refused to do 4.1 with over 1.4v, lmao), the aforementioned 2700X (was dissapointed because it wouldn't do above the stock boost so I moved back to X58 because it was more fun), a 2200G, and got to tinker with an R5 2600 for a friend the other day, got it to a rough 4GHz stable with 3200Mhz CL16 RAM (could've pushed it more with more time but I just wanted to give him an improvement in CS:GO with better single core). I'd say the non-X ones are more fun due to overclocking better than the boost can do it, but the 2700X offered better actual performance. Haven't used any Zen 2 chips so IDK how they do. 

What they really need to to is just clarify what the X actually means, and only put it on applicable models. On Intel the K just means the chip is overclockable, and on HEDT an X means that's the top dog highest core count lad, otherwise it's basically just a K. But the current Zen 2 chips have the 3600, 3600X that's basically the 3600 with a higher TDP so it runs hotter? And then the Ryzen 7s are all X chips, as are the 9s. 

[pas008]

20 minutes ago, GoldenLag said:

unless the density is twice that of 7nm, i dont think so. 

 

7nm TSMC can get quite toasty around 4,5 ghz on Zen 2 and at that point thermal density becomes an issue. 

 

so maybe? the clockspeed regression is probably a factor of them increasing IPC in the design and smaller nodes simply not giving good clockspeeds. 7nm isnt very impressive, and 7nm+ i dont think gave much improvement in clockspeed either. 

 

true

 

their 14nm is so redefined that maybe they need to have 10nm++ to get the clocks at same area and its too late for that game

 

[GoldenLag]

1 minute ago, Zando Bob said:

But the current Zen 2 chips have the 3600, 3600X that's basically the 3600 with a higher TDP so it runs hotter?

its supposed to have a higher boost frequency, but as we have allready seen. the boost is short and needs a lot of voltage. and the thermal density just runs the temp higher. 

 

you get some increase in multicore boost, but PBO more or less removes that and games come more and more multithreaded. the real advantage they have is the higher base clock, imo. 

[cj09beira]

3 hours ago, porina said:

Yes, but we have to consider architecture and process as two parts. Intel 10nm sucks, but what would that same architecture do on 14nm?

also we have to keep in mind that one cant say oh intel just needs this node and this architecture and achieve high ipc 5ghz+ because it could very well be that the higher ipc architecture has more trouble clocking high

2 hours ago, ONOTech said:

It sounds like the 10nm yields are poor (<30%??, that's so low!). Anyone know why the process has been so difficult for Intel?

 

they were aiming to keep being on top, the densest node, 36nm in one of the dimensions (MMP) vs 40 nm for tsmc's 7nm, all the while using a 193nm (wavelength) light source with quad patterning (you etch multiple times to improve resolution of the etch), they were also going to substitute copper on some layers to another metal to improve resistance, in the end they bit more than they could chew, if they had pull it off though it would have been pretty dam impressive

 

49 minutes ago, pas008 said:

you think thermal density has been causing them the problems

all seem to be clocked way lower

 

maybe they know they need to keep the clocks now to remain competitive

 

and part of me thinks they are going to use their own chiplet design very soon(omni directional emib/foveros) but try to keep clocks up and figure away to raise ipc for time being

along with trying to fix some of these security issues

resistance has been increasing quite a bit which a good part of the current problem and lack of reduced power consumption at high clocks.

intel is going chiplet soon as well, they have too, i just hope amd has good enough R&D money by then

49 minutes ago, GoldenLag said:

dont worry, AMD lasted 8-5 years on practically pennies, and returned with Ryzen. 

 

and Intel is a lot wealthier than AMD, so i dont think you need to worry about Intel going away. they have a lot of time and money to get their shit together. 

ya they already announced (they aren't even going to hide it this time) they have 3000M dollars set a side for "discounts" 

Spoiler

aka bribes, the whole slide was how they had set a side 10 times the profits amd made just for the bribes  

16 minutes ago, GoldenLag said:

its supposed to have a higher boost frequency, but as we have allready seen. the boost is short and needs a lot of voltage. and the thermal density just runs the temp higher. 

 

you get some increase in multicore boost, but PBO more or less removes that and games come more and more multithreaded. the real advantage they have is the higher base clock, imo. 

TL;DR its better silicon than non X (on average).

[bcredeur97]

Intel has very tight tolerances on their processes I think that's why 10nm was such a struggle for them.

However, if they can't do 10nm... then why could they do 7??? that doesn't make sense to me.

[jiyeon]

5 hours ago, The Benjamins said:

AMDs 7nm+ zen 3 CPUs will hit mid 2020 and 5nm is on track for 2021.

 

By the time Intel gets there 7nm they will be up against 5nm+

I like the image that Intel is like the Internet Explorer of the CPU space in this context regarding shrinks.

[yian88]

6 hours ago, The Benjamins said:

How on earth are they going to compete?

 

AMDs 7nm+ zen 3 CPUs will hit mid 2020 and 5nm is on track for 2021.

 

By the time Intel gets there 7nm they will be up against 5nm+

Or I suspect they might face 1st Gen 3nm GAA if Samsung has something to say about it.