Intel Process Technology Update: 10nm Server Products in 1H 2020, 7nm in 2021 | Hacker News Reader

Intel Process Technology Update: 10nm Server Products in 1H 2020, 7nm in 2021

anandtech.com

56 pointskristianp15 comments

Is there reason to believe them now? Honest question, I don’t follow the field much too close.

Also, is it too little too late? Where will TSMC be in 2021?

Intel has actually been selling 10nm chips since late last year (in very low volume, a single SKU i believe) so it seems more likely their estimates are on track now.

As for Intel vs TSMC - The node 'sizes' aren't directly comparable, Intel's 10nm is roughly equivalent to TSMC's 7nm and Intel's 7nm is expected to be similar to TSMC's 5nm. TSMC is targeting 2020 for 5nm so Intel is a bit behind still but not as far as the node names would make it appear.

Yea every other time this conversation starts, there's a table floating around showing how the process sizes actually compare on various measurements. It doesn't line up exactly, but yes it's close to what you say, Intel 14nm ~= TSMC/Samsung 10nm, Intel 10nm ~= TSMC/Samsung 7nm.

ish.

The 10nm chip Intel sold last year was bullshit -- low volume and single SKU because it wasn't anything to write home about.

People also keep trying to defend Intel by saying their 10nm is roughly equivalent to the TSMC 7nm, but that's a pretty BS defense seeing as Intel hasn't shipped a real product with it.

Meanwhile TSMC is already working out the kinks on their 5nm process for next year. No matter how you slice it, Intel has fallen behind.

Intels 10nm is nowhere as good as TSMC 7nm.

In theory is was going to be a bit better than TSMC 7mm, with only 7+nm equaling or exceeding it.

In practice, it's slower than Intel 14nm, denser (better be!!), and consumes less power. And has pesky yield issues, Semiaccurate is convinced it'll never be a profitable node.

That cpu was like that Bob Dylan cd collection Sony made 100 copies of.

At this point, Intel has basically come to accept that it needs to redefine 10nm to be whatever meaningful improvement over 14nm it can manage to ship by the end of this year. I'm not sure what exactly this will entail, but I'm sure that there are enough usable ideas from their failed 10nm that they still have a significant advance over 14nm. Even if the fab process ends up not being much of an advance over 14nm, they have a backlog of microarchitecture changes waiting to ship on their first post-14nm process (eg. LPDDR4 support, GPU improvements leading toward their discrete GPUs).

Their 7nm process is quite different and the biggest potential roadblock there is EUV, which is the same potential roadblock everybody else is working against.

I think their general idea is to make 10nm CPUs on very small dies, so that they can get a reasonable acceptance rate. Then push as many ancillary functions as possible onto 14nm silicon.

Bam! A tiny 10nm part surrounded by a bunch of 14nm pieces has a high enough yield that they can get production cranking and start selling "10nm" CPUs.

I don't think Intel's chiplet strategy is getting off the ground soon enough to help them deliver 10nm client CPUs this year. AMD's way closer to shipping something like that, because they're using fairly traditional packaging techniques instead of something like Intel's EMIB. For this year, Intel needs to ship a regular single-die client CPU with integrated graphics, which means they need a 10nm process that solves whatever prevented Cannon Lake's integrated GPU from functioning. (This probably means at a minimum redesigning that GPU to use lower density transistor libraries.)

Semiaccurate's theory, based on their impression that only 1-2 fabs are dedicated to 10nm, is that it has some of the things they're planning on using in 7nm, and they might as well make some chips to sell while they debug them.

Intel's 10nm is a write off. Everything about it is a mess. They're basically trying to get some SKUs out in real volume to save some face, and then try to get their 7nm out in time.

TSMC plans to have their 5nm+ in volume in 2021 (5nm ships next year).

As for how Intel's 7nm stacks up against TSMC's 5nm+? We have no idea. We can't even go off of Intel's 10nm process cause its a complete clown show. Needless to say, by 2021 Intel's prior process advantage will be at least greatly eroded.

I think Icelake will likely be 10nm+, as originally scheduled, because it would be too late to make any significant changes. So they basically skipped the 10nm and Cannonlake generation. They said they are stock piling these parts now, I am going to guess yield isn't anywhere near where it should be.

It is first time ( but I may be wrong ) Intel's has specifically mentioned Node Size with its competitor's name in an investor notes. Intel said they will have 7nm in 2021, and it is roughly the same as TSMC's 5nm. That is only half of the story because if everything works according to schedule TSMC will have 5nm in 2020, and 5nm+ in 2021. And by 2021, TSMC will have already shipped at least 100M to 150M 5nm unit.

Intel still has the lead in high performance node, but that will be tested when AMD launches their EPYC 2 with TSMC's High Performance 7nm later this year.

It is worth mentioning Intel shipped their first 14nm Broadwell in late 2014, which was already 6 - 8 months behind their original schedule. And they will be shipping 10nm in late 2019. For 5 years we had very little performance per clock improvement or node improvement.

We used to value Intel's CPU because it had highest performance per clock ( Now we know that was achieved with security trade off ) and industry leading node. Now both of these are mostly gone.

It would be interesting to see if Apple really moves to their own ARM CPU on Mac. ( Which I really don't care, fix the Macbook keyboard first )

I am still not sure what they meant "client systems on shelf for 2019". All the other sources seem to imply it strictly for mobile (Ultrabook and NUC). So the server 10nm will be first (H1 2020) and then desktop later than that? Does this match Intel's past behavior for their desktop / server product release schedule?

Mobile (Ultrabook and NUC) == client.

Reply on news.ycombinator.com