News Intel Announces Delay to 7nm Processors, Now One Year Behind Expectations

[bit_user]

Apple customers don't care what CPU is in their system, as demonstrated by how many times Apple has changed architectures over the years.

That's ancient history. A lot of people dual-boot their Macbooks or Powerbooks to Windows, which wasn't even possible until Intel got on x86.

I agree with your other points. Aside from Windows-compatibility, I agree that the vast majority of Mac users don't care if it's POWER, x86, ARM, or even VAX.

 

[bit_user]

Do you not realize how high volume Ice Lake SP will be?

Promises, promises... Show me the chips! Like every thing else promised by Intel, Ice Lake SP has slipped, and slipped, and slipped again.

And if Ice Lake SP is so great and will ramp so quickly, why did Intel bother with Cooper Lake (another 14 nm Xeon that just launched 1 month ago)? That's because Intel knows 10 nm is gonna be rough and they are minimally invested in it. Their plans hinged on getting past it quickly.

 

[RodroX]

Getting their AI accelerator into the Mac was probably one of the key reasons. People expect Apple technologies to be revolutionary. Slipping extra CPU cores into the box doesn't generate enough hype.

Yeah Apple is always right on the edge of technology.
I still can't believe those amazing mac pro wheels they launched some months ago (only U$699... and they give you 4 of them!!!! My God!!!!!!), I mean WHEELS!, Can you believe it?!

 

[Ncogneto]

Intel's problems are more complicated than their process technology. It's the fact they are still trying to port to 10 nm and 7 nm while still using a monolithic die design. Until they adopt a chiplet design, they will continue to struggle, because the memory controller does not respond well to a die shrink. Meanwhile, AMD is light years ahead of Intel in chiplet design and interconnect technology (infinity fabric). Anyone that understood the tech, could have seen this coming.

 

[Ncogneto]

Yeah Apple is always right on the edge of technology.
I still can't believe those amazing mac pro wheels they launched some months ago (only U$699... and they give you 4 of them!!!! My God!!!!!!), I mean WHEELS!, Can you believe it?!

The amount of free publicity Apple got out of those $700 wheels was priceless. too bad you don't understand that. They launched a very low volume, outrageously priced product, and got millions of dollars worth of free advertising.

 

[bit_user]

It probably gives them quite a bit more control over customizing chips specifically for their devices, which IMO is a HUGE advantage

Apple already gets customized Intel CPUs and AMD GPUs for its machines. This has been going on, for a long time, actually.

 

[bit_user]

my god what is with this deicidium?

I can't figure him out. He claims not to be an Intel employee or trying to pump Intel's stock (though I forget if he flat-out denied owning any shares). And he's not the gamer/fanboy type. So, I get the sense there's something he's not telling us. Or, maybe he's just defending Intel for the challenge of it?

I don't know, but the fact that he's been spinning for Intel better than their own CEO really speaks volumes.

 

[bit_user]

The yield standpoint comes from a monolithic design versus a chiplet design.

This would be a plausible excuse if Intel were at least shipping 6- or 8- core Ice Lake chips, but they're not. That speaks volumes about Intel's 10 nm yields.

I'm curious to see if they ship any Ice Lake SP dies with all cores enabled. I doubt we'll see that at launch, at least.

 

[InvalidError]

Intel's problems are more complicated than their process technology. It's the fact they are still trying to port to 10 nm and 7 nm while still using a monolithic die design.

There is no "porting to 10nm and 7nm", Intel has native designs for 10nm and 7nm: Sunny Cove, Willow Cove and Golden Cove. There is some BACK-porting to bring some modified version of Sunny/Willow Cove to the desktop in Rocket Lake/14nm.

On the monolithic side of things, practically all multi-chip packages ever made lose to equivalent monolithic designs. If you want the absolute best performance per core possible, monolithic is the way to go. Chiplets only make sense when aggregate system throughput and total cost are more important than per-core throughput.

 

[Ncogneto]

There is no "porting to 10nm and 7nm", Intel has native designs for 10nm and 7nm: Sunny Cove, Willow Cove and Golden Cove. There is some BACK-porting to bring some modified version of Sunny/Willow Cove to the desktop in Rocket Lake/14nm.

On the monolithic side of things, practically all multi-chip packages ever made lose to equivalent monolithic designs. If you want the absolute best performance per core possible, monolithic is the way to go. Chiplets only make sense when aggregate system throughput and total cost are more important than per-core throughput.

What ever. My comment stands. The fact is the memory controller does not respond well to a die shrink. Newsflash, it's not about the "best performance per core possible", it's about the best performance per dollar possible. Period. End of discussion.

"Chiplets only make sense when aggregate system throughput and total cost are more important than per-core throughput"

Yeah, so things like higher yields, better binning ability, and the fact chiplets can actually work on 7 nm don't matter. Gimme a break............ Monolithic can be regarded as "the way to go" only if they can keep pace with the die shrinks of the chiplet design.

it's not about comparing a monolithic 14 nm processor to a 14 nm chiplet processor, now is it?

 

[bit_user]

The only things we have that would be an apples to apples comparison is on server hardware.

I said the ISA is more efficient to implement. That's the main reason Intel failed in the phone market.

Anyway, if you need proof of the potential of ARMv8-A, the best example is Apple:

https://www.anandtech.com/show/13392/the-iphone-xs-xs-max-review-unveiling-the-silicon-secrets/4​

On those comparisons the ARM chips have the same TDP but lower clocks and single threaded performance of about 1/2-2/3 the Epyc or Xeon equivalents. They are able to pull close on massively threaded applications where the SMT4 is able to help.

That's some old info, you got there.

https://www.anandtech.com/show/15578/cloud-clash-amazon-graviton2-arm-against-intel-and-amd/7​

Unfortunately, that comparison is not against the current 7002-series EPYCs. Amazon's Graviton 2 has no SMT, so they are comparing a 64-core CPU against a 32-core/64-thread EPYC (and a current-gen Cascade Lake Xeon).

Until recently, ARM hasn't been making performance-oriented cores. So, even the existing server chips don't show the ISA in its best light. However, things aren't looking so good for x86:

https://www.anandtech.com/show/15578/cloud-clash-amazon-graviton2-arm-against-intel-and-amd/9​

 

[bit_user]

Until they adopt a chiplet design, they will continue to struggle, because the memory controller does not respond well to a die shrink.

Evidence to the contrary:

https://www.tomshardware.com/news/amd-ryzen-4000-renoirs-memory-controller-is-scary-good​

Renior is 7 nm and, like their other APUs, uses a monolithic die.

 

[InvalidError]

The fact is the memory controller does not respond well to a die shrink.

Memory controllers, HSIO lanes and everything else are absolutely fine on 7nm. The only downside is that they don't scale down much silicon-wise. However, when you go chiplets, you still end up spending a whole lot of die area on chiplet interconnects anyway so you aren't necessarily saving all that much die area there.

Newsflash, it's not about the "best performance per core possible", it's about the best performance per dollar possible. Period. End of discussion.

Millions of gamers on Intel's K-series CPUs still stomping AMD's best using a five years old architecture beg to disagree.

 

[bit_user]

On the monolithic side of things, practically all multi-chip packages ever made lose to equivalent monolithic designs.

Serious question: have you (or anyone else) seen an IPC comparison between Renior and Ryzen 3800X? That would be a good way to know how much efficiency AMD is losing by their multi-die approach* (although the 3000-series also has the disadvantage that its L3 and memory controller are an earlier process node).

You can argue abstract points all day, but real data is king. Even if you're right (and I'm not saying otherwise), it's important to know the scale of the impact we're talking about.

* IIRC, communication between the 3800X's two CCXs still goes through the I/O die.

 

[Ncogneto]

Evidence to the contrary:

https://www.tomshardware.com/news/amd-ryzen-4000-renoirs-memory-controller-is-scary-good​

Renior is 7 nm and, like their other APUs, uses a monolithic die.

And can only support PCie 3.0, has a top clock speed of 3600 base and DDR4 3200. Amd itself advised that users will want to wait until the next gen of chips, that will use what? A chiplet design.

 

[Ncogneto]

Memory controllers, HSIO lanes and everything else are absolutely fine on 7nm. The only downside is that they don't scale down much silicon-wise. However, when you go chiplets, you still end up spending a whole lot of die area on chiplet interconnects anyway so you aren't necessarily saving all that much die area there.


Millions of gamers on Intel's K-series CPUs still stomping AMD's best using a five years old architecture beg to disagree.

Stomping? LOL, more like eeking out a narrow victory at useless 1080P gaming. Like I am going to spend 1k on a video card to game at 1080P. Give it a rest. As if anyone can tell the difference between 240 FPS and 210 FPS, lmao.

 

[Conahl]

And he's not the gamer/fanboy type

you sure about that ?

Ncogneto from what i have seen, intel only has any performance advantage, is because of its clockspeed, and takes a pretty big power usage hit to get that advantage. would love to see a comparison between intel and amd at the same clocks. both on performance, and power usage.

 

[RodroX]

The amount of free publicity Apple got out of those $700 wheels was priceless. too bad you don't understand that. They launched a very low volume, outrageously priced product, and got millions of dollars worth of free advertising.

Thanks you for saying I don't understand that. I don't know who the hell do you think you are to know what I can think or not.

 

[InvalidError]

Stomping? LOL, more like eeking out a narrow victory at useless 1080P gaming. Like I am going to spend 1k on a video card to game at 1080P.

You may not but there are competitive gamers out there seeking 240+Hz monitors.

 

[bit_user]

would love to see a comparison between intel and amd at the same clocks. both on performance, and power usage.

It's not exactly what you wanted, but the bottom of this page features a synthetic IPC comparison:

https://www.anandtech.com/show/14605/the-and-ryzen-3700x-3900x-review-raising-the-bar/6​

Now, if you want to put them on a level playing field, I'd say the focus should really be on power-parity. It doesn't really matter if one clocks a bit higher, as long as it doesn't burn more power.

 

[bit_user]

Thanks you for saying I don't understand that. I don't know who the hell do you think you are to know what I can think or not.

Agreed. The attack was unnecessary to make the point.

FWIW, I don't believe the $700 price tag was a publicity stunt, either. It's just typical Apple profit-maximizing behavior. Anyone seriously in the market for a Mac Pro already knows it, and Apple already got lots of free publicity for it, when they announced it.

The way to think about the $700 wheels is more like a $10k upholstery option, in a high-end luxury car. The manufacturer knows you can probably afford it, since you're already buying the car. They just don't want to miss any chance to make a little more money off you.

 

[NightHawkRMX]

What? Intel pushing back a node? I have never heard of this before.

I guess Intel will have to add more plus signs

 

[Conahl]

It's not exactly what you wanted, but the bottom of this page features a synthetic IPC comparison:

https://www.anandtech.com/show/14605/the-and-ryzen-3700x-3900x-review-raising-the-bar/6​

Now, if you want to put them on a level playing field, I'd say the focus should really be on power-parity. It doesn't really matter if one clocks a bit higher, as long as it doesn't burn more power.

normalizing for clock speed vs actually making both sides run at the same clocks, then running each test, would be different, wouldn't it ?

 

[jasonf2]

Wow. It is the 10 nano meter process node all over again. This is a huge win for AMD.

 

[jasonf2]

Intel for all of its power and money forgot a simple rule of economics. To be a monopoly there has to be an impossible barrier of entry. For a tech company this means that there has to be innovation and a performance lead that makes the competitors unable to gain market traction. Between process nodes and the insistence that four cores are good enough for the last decade they had dug their own hole. If they don't come out with something akin to the Nehalem in the Athlon 64 age pretty soon the hole they have dug will be their grave.