"A Chat With Charlie"
Update for EOY 2019 of the chat seven months ago. "What's up at Intel"? (not an AMD-centric chat)
Charlie (SemiAccurate) focuses on news that no one else has, or that they break before anyone else.
Q: What's up with Intel?
In short, it's a mess; they didn't deliver
10 nm is still technically non-existent
Qtys are non-economical
Slow, power hungry, hot, expensive
It doesn't push the bar.
"They have issues".
Q: What's going on with 10 nm?
It was a choice they made, back in spring 2017, when they told everone in that slide that both 10 nm and 10+ nm would be slower than 14 nm
The process has better power efficiency, but performance is lower
Performance directly drives ASP
So they juiced it past its optimal point to get it to tie with 14nm by boosting power; self-defeating.
The GPU is better, but it takes up a LOT of die area.
Q: Speak to four things from the last call: 2.7X scaling, COAG (contact over active gate), Cobalt, SAQP (Self Aligned Quad Patterning); any updates? Also, margins and yield.
Intel delayed 10 nm, but claimed there were no changes.
Tech Insights says otherwise; one major change - COAG
The yields are below sanity levels.
Why claim no changes, all is well: Just to keep the share price up
They are burning money to keep 10 nm a viable thing, but it is simply not economically viable.
Ice Lake yields are in the toilet; there are no upsides to this.
"It's dead".
Q: % of product going to 10 nm. Start with laptop
Very low. Laptops account for 70% of Intel units.
Intel is still hurting in capacity. PC growth was only a few percent
If 10 nm was even 5% of laptops, the capacity problem would have gone away. They just aren't there.
10 nm is a PR spin.
10+ nm Ice Lake are technically on shelves, but they aren't economically viable or in any volume.
10++ nm might be faster than 14 nm; it won't ship until EOY 2020.
Ice Lake Sunnycove did have an 18% IPC boost, which was the only reason they could bring a chip to market, but the clocks are lower
Their only chance for decent 10 nm is tiger Lake / Willow Cove, at the end of 2020.
Q: Talk about the reorg.
It has happened.
DCG (Data Center Group) is now in DPG (Digital Platform Group), so Xeon servers are lumped in with [missed that].
"Shuffling of the deck chairs"
Huge cuts to current DCG staff in Q1.
AI and MobileEye, currently autonomous, in danger of being subsumed.
It is supposedly to streamline things, not have a lot of people doing the same job.
Charlie doesn't buy it. Two possibilities:
Serious structural reform? No.
So that leaves simply making it hard to make comparisons to prior years.
Q: Memory? Habana?
Intel is desparate to exit the memory business, but no one wants it.
Optane doesn't appear to be selling above cost.
Habana: They acquired them because Intel's internal AI products were simply not competitive.
They paid way too high a price for the company.
It was a white flag on internal AI.
Q: Shortages over when?
In H2 of 2018, the shortages were due to server performance hits over security vulnerabilities.
In December of 2018, they admitted that it was due to increasing die size, which dropped the number of chips that they can produce.
This is due to AMD.
Intel's sweet spot, with 60% of sales, used to be 80-90 mm^2.
AMD core increases drove that up to 125 mm^2, so they lost one third of their production capacity.
Intel's original plan was to have a four core max at 14 nm. That is now 6 to 8 cores, even 10.
Ten core CPUs have technically been out for a quarter now, but they still don't exist.
Also yields go down as the die size goes up.
How do they solve it? Three ways:
- 10 nm online at the end of 2020 (best case, but yield problems).
- 7 nm online at the end of 2021
- AMD takes enough share where their capacity problem is solved.
Charlie mentions the Dell meltdown, and the letter mentioning "something in the fabs"
Sounds like whatever went wrong would be a material (as in SEC) issue.
"Something went bump in the night at their fabs".
Q: Desktop 10 nm?
Best case 2021, NOT Ice Lake, Tiger Lake 10++ nm Laptops EOY 2021
Tiger Lake Desktop lower performance
On server, 2021 best case, not likely competitive.
Q: Discounting for Laptop/Desktop/Server?
Laptop not much discounting yet.
AMD's Q1 Ryzen 4000 will put Intel under pretty severe pressure.
Up to now Intel had a free ride on laptops
Desktop: "It's not a fair fight".
AMD desktop "Better in every way".
Intel creates performance tiers by crippling features; AMD gives you them all.
"AMD crushes Intel on Desktop".
Intel has no hope of fixing this for 18 months.
Discounting on desktop: Charlie doesn't know.
Desktop workstations are really Xeons
When Threadripper 3 came out, Intel slashed their whole line of Xeon W prices in half.
TR3 is selling a lot, but no one can get Xeon W's.
Intel stopped making them, because they knew they wouldn't sell.
Like the 9900KF, the mythical 5 GHz beast. They sent some to reviewers, put a few online.
But they can't make them.
Desktop Workstation is a good proxy for servers.
Q: When will Intel become competitive in laptops after Renoir? What about Zen 4?
In laptops, Renoir will do to them what they did in desktop: destroy them in the marketplace.
Intel will have no response until Tiger Lake at the end of 2020
That has a good chance, but yields may well be so low that it doesn't matter.
The laptop market is most vulnerable to slush funds and underhanded games.
Look at the Microsoft Surface 3; Microsoft got a windfall of dumptruck loads of cash from Intel, all because of the threat of one AMD SKU.
The CPU MSRPs vary by $300-$400, but Intel's one sells for only $100 more, plus they throw in Windows Pro, which eats up most of that $100.
Q: Zen 4 motherboards, and Genoa?
That's two generations away.
First comes Zen 3/Milan.
Chips have been back for months now, and they perform significantly better than expected.
Same motherboards. AMD is in a really good position for 2020.
Q: Servers? 10 nm pushed out?
OEMs were told directly that servers were pushed out, but Intel denies that.
Splitting hairs, or intentionally misleading? You decide.
Ice Lake server is 2021.
Q: What the heck is going on here? TDP leaks?
Intel's server program is a mess.
Charlie doesn't know why. Management?
Intel's engineers are very good.
Institutionally, everything is melting down.
Management: Everything's great, they claimed that 10 nm was shipping for revenue in 2017.
Yields are improving, better than expected, etc. 10 nm is always just around the corner.
Four years later, it still doesn't work.
But things keep sliding.
Are they wordsmithing, or are they lying through their teeth?
"Go Away. Stop Lying".
It's a broken record. Really hard to figure out where things went off the rails.
Q: SemiAccurate report on Intel server push-outs?
Ice Lake server is now a 2021 product. Intel directly informed OEMs of that, despite subsequent denials.
Sapphire Rapids was to have been in H1 2021. It is now 2022 - at least a full year slip.
The program is in deep deep trouble [there may have been a third "deep"]
"Borderline mind-boggling"
Granite Rapids was to have been its successor; that is now H2 2023.
Everything until late 2023 has been pushed out.
Cooper Lake was to have been Q1 2020; it is now H2.
How do we know? Development systems are unavailable. That isn't possible with a Q1 launch.
"The roadmap is a mess"
Here's why I stated that "I am worried about Intel":
From my August 2018 article, Why Intel has no chance in servers, and they know it:
Intel knew that there was no chance of catching up with AMD until after Sapphire Rapids
Which meant Granite Rapids. But that is now coming in H2 2023.
Intel has no chance of catching up until then.
There is no sign of AMD slipping.
A one year slip is probably OK
A two year slip is problematic
Any longer, and I question their viability, they may no longer be an ongoing entity (SEC speak).
Cooper lake (14 nm) was also pushed out.
TDPs are very important
At Computex, Intel raised Cooper Lake's TDB from 250 to 300 watts.
That is not an innocious change. 250 watts is the limit for air cooling.
Six months before its purported launch, Intel basicly told its OEMs "You have to water cool".
The OEMs were not happy, and used very colorful language.
Forcing water cooling screwed over everyone in the infrastructure.
Why did Intel do this? Not to lose so badly on the benchmarks, and to be able to move from 40-48 cores to 56 cores.
"Look, close to 64"!
They single-handedly gave AMD a big advantage: "We don't require water cooling".
Last Q from CR: Will Intel's 7 nm be better? EUV?
Granite Rapids is still 2021 per Intel. It's really 2023.
No visibility on 7 nm.
10 nm: "Everything is wonderful" They are claiming the same for 7 nm.
I'm pessimistic that Intel came down with a bout of honesty.
TSMC 5 nm is essentially Intel's 7 nm.
TSMC is ramping their 5 nm now.
Intel is best case a year behind.
TSMC is solid. Intel, no reason to believe.
Email question phase:
Q: What did Intel tell their OEMs regarding Ice Lake server?
Samples Oct 2020, volume production starts November 2020, sales in 2021.
Q: Not impacting financials, why?
Servers come in small, medium, and large sizes.
If you bid on a small, they offer you a large for the same price: more cores, more performance.
Same silicon, just the fuses not blown, keep the customer.
Cannibalizing their tier strategy.
Near the breaking point.
But sill competitive at the low end of servers.
Q: DCG cuts to personnel, why? And any outsourcing?
Already outsourcing to Samsung [I don't think these are CPUs].
That still didn't help with the capacity problem.
The DCG cuts are sales, not tech jobs.
The cuts are those redundant with other orgs.
The cuts will extend to other orgs in Q1.
Q: Discrete GPU push and capacity constraints
If they are made in house, that doesn't help with capacity issues.
7 nm is way out.
Big GPU unveil at CES.
It's their DG1. That is basically just an eval board. Not gaming.
Q: What about AMD's new board member?
No idea; "he took my seat".
Q: Is there an AMD-Samsung mobile phone GPU development?
Yes, they said that.
Q: Can Jim Keller change things?
Maybe. The first CPU he could have affected would be Granite Rapids.
The first architecture that he could affect would be after Granite Rapids, 2025.
Sapphire Rapids was fully done when he came on board, so the platform for Granite Rapids was handcuffed.
Closing CR Q: What would you do to turn things around at Intel?
Up to Granite Rapids, scrap it all and start over.
Their roadmap is simply not competitive.
A roadmap takes 4 to 5 years.
Start over, there is no light at the end of the tunnel