There is a good chance for Intel to recover, but that remains to be proven.
From their long pipeline of future CMOS manufacturing processes with which Intel hopes to close the performance gap between them and TSMC, for now there exists a single commercial product: Meteor Lake, which consists mostly of chips made by TSMC, with one single Intel 4 die, the CPU tile.
The Meteor Lake CPU seems to have finally reached the energy efficiency of the TSMC 5-nm process of almost 4 years ago, but it also has obvious difficulties in reaching high clock frequencies, exactly like Ice Lake in the past, so once more Intel has been forced to accompany Meteor Lake with Raptor Lake Refresh made in the old technology, to cover the high-performance segment.
Nevertheless, Meteor Lake demonstrates reaching the first step with Intel 4.
If they will succeed to launch on time and with good performance, later this year, their server products based on Intel 3, that will be a much stronger demonstration of their real progress than this Meteor Lake preview, which has also retained their old microarchitecture for the big cores, so it shows nothing new there.
Only by the end of 2024 it will become known whether Intel has really become competitive again, after seeing the Arrow Lake microarchitecture and the Intel 20A manufacturing process.
N5 is interesting because it's the first process fully designed around EUV and because it was pretty much exclusive to Apple for almost two years. It launched in Apple products in late 2020, then crickets until about late 2022 (Zen 4, RTX 4000, Radeon 7000). Launches of the other vendors were still on N7 or older processes in 2020 - RTX 3000 for example used some 10nm Samsung process in late 2020. All of those were DUV (including Intel 7 / 10ESF). That's the step change we are looking at.
Exactly. N5 is sort of an outlier, it's a process where a bunch of technology bets and manufacturing investment all came together to produce a big leap in competitive positioning. It's the same kind of thing we saw with Intel 22nm[1], where Ivy Bridge was just wiping the floor with the rest of the industry.
Improvements since have been modest, to the extent that N3 is only barely any better (c.f. the Apple M3 is... still a really great CPU, but not actually that much of an upgrade over the M2).
There's a hole for Intel to aim at now. We'll see.
[1] Also 32nm and 45nm, really. It's easy to forget now, but Intel strung together a just shocking number of dominant processes in the 00's.
The reason N5 came together for TSMC is because they run more experiments per unit time than Intel does. They're doing this 24 hours a day across multiple shifts, which makes it possible for them to improve a given process faster. It remains to be seen if Intel can actually pull ahead or not without a major culture change, or if "this time" they can succeed at becoming a trusted foundry partner that can drive enough volume to support the ongoing investment needed in leading edge fabs.
> The Meteor Lake CPU [...] has obvious difficulties in reaching high clock frequencies,
Not sure where that's coming from? The released parts are mobile chips, and the fastest is a 45W TDP unit that boosts at 5.1GHz. AMD's fastest part in that power range (8945HS) reaches 5.2GHz. Apple seems to do just fine at 4GHz with the M3.
I'm guessing you're looking at some numbers for socketed chips with liquid cooling?
The 5.1 GHz Intel Core Ultra 9 processor 185H is the replacement for the 5.4 GHz Intel Core i9-13900H Processor of previous year. Both are 45-W CPUs with big integrated GPUs and almost identical features in the SoC.
No liquid cooling needed for either of them, just standard 14" or 15" laptops without special cooling, or NUC-like small cases, because they do not need discrete GPUs.
Both CPUs have the same microarchitecture of the big cores.
If Intel had been able to match the clock frequencies of their previous generation, they would have done that, because it is embarrassing that Meteor Lake wins only the multi-threaded benchmarks, due to the improved energy efficiency, but loses in the single-threaded benchmarks, due to lower turbo clock frequency, when compared to the last year's products.
Moreover, Intel could easily have launched a Raptor Lake Refresh variant of i9-13900H, with a clock frequency increased to 5.6 GHz. They have not done this only to avoid an internal competition for Meteor Lake, so they have launched only HX models of Raptor Lake Refresh, which do not compete directly with Meteor Lake (because they need a discrete GPU).
During the last decade, the products made at TSMC with successive generations of their processes had a continuous increase of their clock frequencies.
On the other hand Intel had a drop in clock frequency at all switches in the manufacturing processes, at 14-nm with the first Broadwell models, then at 10-nm with Cannon Lake and Ice Lake (and even Tiger Lake could not reach clock frequencies high enough for desktops), and now with Meteor Lake in the new Intel 4 process.
With the 14-nm and 10-nm (now rebranded as Intel 7), Intel has succeeded to greatly increase the maximum clock frequencies after many years of tuning and tweaking. Now, with Meteor Lake, this will not happen, because they will pass immediately to different better manufacturing processes.
According to rumors, the desktop variant of Arrow Lake, i.e. Arrow Lake S, will be manufactured at TSMC in order to ensure high-enough clock frequencies, and not with the Intel 20A, which will be used only for the laptop products.
Intel 18A is supposed to be the process that Intel will be able to use for several years, like their previous processes. It remains to be seen how much time will pass until Intel will become able to reach again 6.0 GHz in the Intel 18A process.
That's getting a little convoluted. I still don't see how this substantiates that Intel 4 "has obvious difficulties in reaching high clock frequencies".
Intel is shipping competitive clock frequencies on Intel 4 vs. everyone in the industry except the most recent generation of their own RPL parts, which have the advantage of being being up-bins of an evolved and mature process.
That sounds pretty normal to me? New processes launch with conservative binning and as yields improve you can start selling the outliers in volume. And... it seems like you agree, by pointing out that this happened with Intel 7 and 14nm too.
Basically: this sounds like you're trying to spin routine manufacturing practices as a technical problem. Intel bins differently than AMD (and especially Apple, who barely distinguish parts at all), and they always have.
I have also pointed that while for Intel this repeats their previous two process launches, which is not a good sign, TSMC has never had such problems recently.
While one reason why TSMC did not have such problems is that they have made more incremental changes from one process variant to another, avoiding any big risks, the other reason is that Intel has repeatedly acted as if they had been unable to estimate from simulations the performance characteristics of their future processes and they have always been caught by surprise by inferior experimental results compared to predictions, so they always had to switch the product lines from plan A to plan B during the last decade, unlike the previous decade when all appeared to always go as planned.
A normal product replacement strategy is for the new product to match most of the characteristics of the old product that is replaced, but improve on a few of them.
Much too frequently in recent years many Intel new products have improved some characteristics only with the price of making worse other characteristics. For example raising the clock frequency with the price of also increased power consumption, increasing the number of cores but removing AVX-512, or, like in Meteor Lake, raising the all-cores-active clock-frequency with the price of lowering the few-cores-active clock frequency.
While during the last decade Intel has frequently progressed in the best case by making two steps forward and one step backward, all competitors have marched steadily forwards.
> I have also pointed that while for Intel this repeats their previous two process launches, which is not a good sign, TSMC has never had such problems recently.
I'll be blunt: you're interpreting a "problem" where none exists. I went back and checked: when Ivy Bridge parts launched the 22nm process (UNDENIABLY the best process in the world at that moment, and by quite a bit) the highest-clocked part from Intel was actually a 4.0 GHz Sandy Bridge SKU, and would be for a full 18 months until the 4960X matched it.
This is just the way Intel ships CPUs. They bin like crazy and ship dozens and dozens of variants. The parts at the highest end need to wait for yields to improve to the point where there's enough volume to sell. That's not a "problem", it's just a manufacturing decision.
You can't compare optimized clock frequencies on 2 year mature process with first run on a new process... AMD and Nvidia both improve stable frequencies with process improvements at TSMC even on the same nodes over time (RTX 4060 TI, tsmc N4, 2.31ghz base, 2.54ghz boost vs RTX4060 - TSMC N4, 1.83ghz base, 2.46ghz boost).
Most chipmakers saw gains moving from n5 to n5p at tsmc, which wasn't even a process jump simply maturity and optimization on the existing node.
From their long pipeline of future CMOS manufacturing processes with which Intel hopes to close the performance gap between them and TSMC, for now there exists a single commercial product: Meteor Lake, which consists mostly of chips made by TSMC, with one single Intel 4 die, the CPU tile.
The Meteor Lake CPU seems to have finally reached the energy efficiency of the TSMC 5-nm process of almost 4 years ago, but it also has obvious difficulties in reaching high clock frequencies, exactly like Ice Lake in the past, so once more Intel has been forced to accompany Meteor Lake with Raptor Lake Refresh made in the old technology, to cover the high-performance segment.
Nevertheless, Meteor Lake demonstrates reaching the first step with Intel 4.
If they will succeed to launch on time and with good performance, later this year, their server products based on Intel 3, that will be a much stronger demonstration of their real progress than this Meteor Lake preview, which has also retained their old microarchitecture for the big cores, so it shows nothing new there.
Only by the end of 2024 it will become known whether Intel has really become competitive again, after seeing the Arrow Lake microarchitecture and the Intel 20A manufacturing process.