Thursday, December 17th 2020
Intel's Manufacturing Outlook for the Future Doesn't Inspire Confidence in Successful Competition, According to Susquehanna Call
Susquehanna is a global trading firm which has various interests in silicon manufacturing - and part of that interest is naturally materialized in Intel. In a recent group call from the firm, some details on Intel's manufacturing and product design woes came to light, which point towards even more execution slips than we've already seen. During the call, a number of points were broached, including dismal yields for Intel's 10 nm manufacturing process as of its introduction in late 2018 (which is why it never saw mainstream adoption from the company). News that Intel is looking for a new CEO also don't instill confidence on current CEO Bob Swan's capacity to steer the Intel behemoth.
Improved yields on 10 nm are being reported due to deployment of Intel's SuperFin technology, which improved yields to upwards of 50%, but still keeps them under the ones achieved in Intel's 14 nm process; an eye-opening tidbit in that Cannon Lake on 10 nm originally saw yields of only 25% usable chips per wafer; and that backporting Rocket Lake meant Intel had to deal with unfathomably large chips and high power consumption characteristics. And to add insult to injury, there is still not a definite timetable for 7 nm deployment, with delays being expected to be worse than the previously reported 6-12 months. This all paints a somewhat grim picture for Intel's capacity to compete with TSMC-powered AMD in many of its most important markets; the blue giant won't topple, of course, but it's expected that five years from now, we'll be looking at a very different outlook in the market between AMD and Intel. You can check the talked-about points in the call via the transcript after the break. You should still take the transcript with a grain of salt.
Sources:
Reddit, via Notebookcheck
Improved yields on 10 nm are being reported due to deployment of Intel's SuperFin technology, which improved yields to upwards of 50%, but still keeps them under the ones achieved in Intel's 14 nm process; an eye-opening tidbit in that Cannon Lake on 10 nm originally saw yields of only 25% usable chips per wafer; and that backporting Rocket Lake meant Intel had to deal with unfathomably large chips and high power consumption characteristics. And to add insult to injury, there is still not a definite timetable for 7 nm deployment, with delays being expected to be worse than the previously reported 6-12 months. This all paints a somewhat grim picture for Intel's capacity to compete with TSMC-powered AMD in many of its most important markets; the blue giant won't topple, of course, but it's expected that five years from now, we'll be looking at a very different outlook in the market between AMD and Intel. You can check the talked-about points in the call via the transcript after the break. You should still take the transcript with a grain of salt.
- 10 nm in "full production" and "shipping for revenue" since late 2018. Charlie still doesn't think it's viable. Much more real than it once was though, but not good enough.
- Superfin is unquestionably far better than base 10 nm. Better in just about every way. Yields are better, but still on the low side.
- CNL had a yield of <25% even with iGPU disabled (seems a bit high to me but huh).
- 10 nm SF yields are far better but not as good as 14 (no surprises there). Past 50%.
- Not enough 10 nm capacity to handle the entire product stack yet. Charlie doesn't like Rocket Lake, considers the backport "painfully stupid". Will be huge and a power hog.
- 14 nm shortage caused by the fact that Intel had to bloat up die sizes.
- OEMs told they were to get a "small fraction" of orders of TGL. Improvements to yield won't fully help because no. of wafer starts is still low.
- He doesn't think 10 nm will ever exceed 14 nm capacity-wise.
- Ice Lake delayed again due to a bug, but this is a good thing. Looking like it'll come mid-late Q2 to early Q3 (although Charlie originally said March + 1.5-2 month delay?).
- Performance looks pretty poor. Akin to CL-SP actually - performance per socket remains the same but power rises. Core count for shipping parts dropped from 38 -> 36.
- Ice Lake is disappointing. Sapphire Rapids is way better. Comes in Q2 22.
- SPR was originally supposed to launch in Q2 21 to ship for Aurora but... yeah. Even then it was basically going to be a broken, alpha product which would be refined over multiple tapeouts.
- Once SPR fully rolls out and is fully sorted out it'll actually be pretty solid.
- The problem with SPR is timeline. ICL-SP was supposed to compete with Rome (and it would still lose even with the on-paper and significantly better specs than reality). Milan is going to likely beat SPR... over a year earlier and before ICL-SP rolls out. By the time SPR releases, Genoa will be here or right around the corner. Genoa was supposed to compete with GNR which is late 2023 now.
- If both companies iterate perfectly on their roadmaps as planned (much easier on AMD's side right now), Intel can not catch up to AMD until late 2024 or early 2025.
- AMD's biggest problem? Unsurprisingly - capacity.
- 7nm is significantly more delayed than the 6-12 months first claimed. PVC is in-hands basically now and it was the first 7nm product. Now it's on TSMC. Sometime December this year anyway.
- Intel may not know how much of a delay we're looking at with 7.
- Charlie thinks there are several problems with 7, not just 1. However, Charlie hasn't heard anything absolutely solid on the specifics here.
- Doesn't think Intel could sell fabs. Too many things problems involved. Fabs would need to be evaluated for the buyers needs etc etc.
- Foveros is more suited to low-power chips than servers/high power.
- Question about why ARM is more competitive in recent attempts vs previous attempts - Answer: more developed eco-system, lack of real glass-jaws compared to previous designs. Amazon is a bit more of a special case because they can design chips to exactly what they need with all the strengths and benefits that suit them best.
- Intel are going to out-source CPUs, but what and where is still being debated internally. Decision will be made in Q1 between SS and TSMC for different products. Nothing is set in stone yet.
- Intel are looking for a new CEO. There are multiple candidates, one has been shot down, 3 potential left. They are hoping to make the decision in Q1, probably after the Jan call. Does seem like Bob's going either way.
- Question: Will Intel do anything else with the backported core? A: I hope not.
- Intel won't outsource everything to TSMC because TSMC doesn't have the capacity. Intel grabbing capacity shouldn't take capacity away from AMD - TSMC should prioritise customers with long-term plans to stay with them.
- There is a 10 nm desktop product coming (Alder Lake-S) but Charlie doubts there will be much capacity for it due to it being needed elsewhere.
33 Comments on Intel's Manufacturing Outlook for the Future Doesn't Inspire Confidence in Successful Competition, According to Susquehanna Call
Yep... they still are invested in a non profitable node... staggering
The shitshow continues.
Like a drifting boat the catchup date for Intel just keeps getting further and further away......
Maybe someone will toss 'em a rope.... Or not.
Some claims look strange though, the parts about bloated dies and 14nm shortages, for example.
As for the bloated dies, it makes sense that if you develop a design for a dense process, and then have to backport it to a less dense process, all the transistors you had crammed in there now have to be spread out in less density, which leads to a bigger overall footprint. Pulling this out of my proverbial, but if you have a 200 mm2 die in 10 nm, bringing it to 14 nm would make it a 320mm2 die. Hence, bloated, lower yields, higher latency, more leakage... and all of that.
But news like this just makes me cringe everytime I even think about suggesting that any of my clients opt for either 10 or 11 gen stuff....
I say it's time for a mgmt shakeup/down, or at least for someone to go in and light a plutonium-fueled fire under the collective asses of the current mgmt team, hehehe :)
The only thing that can save them now I think is some government subsidies/military involvement to get their Fabs back up to a competitive state --
im sure that intel, which have more than 10 time more capacity all around over amd, raise top next year.
might also gpu side...
well,amd's products lean strongly TSMC products and there happends alot just now.
but intel not need it,intel build it cpus itself,no need help
lets wait intel rocket lake for gamers and then after few month, intel adler lake for all.
rocket lake is 14nm but best gaming cpu cleary
adler lake is 10nm with hybrid technology and wha i see, its raise cpus power new age.
march and june 2021 shows alot that thing...
i say Q3/2023,both amd and intel have 3-5nm cpus both,intel for sure lead gaming performance,is it amd still theory apzz score,like cinbench,we will see it.
anyway,2021 shows alot.and i mean both side.
and this is bad news for AMD fans. Expect price increase due to milking and TSMC price hikes. Well deserved for such die-hard fanatics! Lisa is eyeing $1000 for a mainstream CPU with Zen4! Let’s see if they’re as loyal as they claim...
losing their one trick pony they’ve been reselling for over a decade to real innovation is actually value for the possible extra cost that innovation gets you.
No wonder there is no word on 10nm for the desktop, but this would make even the rumored Tiger Lake-H unlikely to actually show up.
It seems they should under stand the physics and the engineering required to make these denser products.
It would also add clarity to articles like this if we had a definition of what a 10nm node is.
We know it's not gate size. The "size" of an electron in silicon is about 5nm so it is unclear what TSMC or Intel mean on 7nm or smaller nodes.
I'm obviously not expecting to see this happening but I was wondering how bad the delay is going to be.
The industry desperately needs Samsung to bring some competition to TSMC's 5nm and with a really decent capacity too, I know Samsung has enough EUV machines ready to do this but they just need to perfect the node first.
Intel could then use TSMC where it wants, Samsung where it wants and still us it's own 10nm SF for some products and 14nm for it's bulk production of low performance (cheaper) products.
If Intel had swallowed it's pride in 2015 after seeing it was having problems with the 10nm node they could have scaled back their ambitions in an attempt to deliver something, then said if it still looks like the issues won't be solved by 2017 they should have started discussion with TSMC. Intel easily has enough money to have two design teams running side by side, one working on their own ill fated 10nm node and another using TSMC. That would have meant in 2019 if the in house FABs hadn't sorted themselves out Intel was ready to go at TSMC. They could have released high end products on TSMC nodes and cheaper volume products on their own FABs.
One other avenue I think should have been explored was Intel licencing TSMCs IP for the 7nm node maybe a deal all the way through to 3nm when everything is changing to GAAFET. TSMC doesn't have anywhere near the capacity to serve their long term customers and also help Intel out in a big way and Intel has the capacity just not the IP. So put these two together and TSMC could make an absolute fortune by licencing their node knowhow to Intel and Intel wouldn't be where it has been for the last 3 years and where it will be for the next 5!
intel had allocated another $10 billion this year on buybacks, basically at the cost of their NAND business.