Welcome to the rabbit hole!
Thanks for the insight into this kind of nebulous business. Edge seems to be fine given how many random intel kiosks and industrial PCs exist in the wild, but 5G basestations are something that I have no clue what the space looks like (other than QCOM/BCOM probably having a presence).
I hope they can hold market share after next gen EPYC releases. Do you think that's plausible given intel 3 server chips will release around 6 months later?
What do you think about IFS manufacturing 5G chips for Ericsson and one other company I forgot the name of?
I think the opportunity is a positive future indicator for IFS, albeit probably not a very high volume one.
The question becomes, does Intel's competitive situation get worse, better, or roughly the same, I think.
I wouldn't really call the Ericsson deal an IFS win.
What do you purpose intel makes with intel 3 in the client next year? If you are using an internal node, there is no reason not to shoot for 20A for the compute die. Intel 3 is supposed to be HVM ready in 2H23 while 20A is supposed to be in 1H24. As long as 20A products can hit the beginning of the year OEM refresh cycle, just use the node that is better. As for DC, SRF is supposedly coming in 1H24, so 20A is out of the question (if it even has the full feature list to even support the SRF dies). I think it is also a safe guess that GNR will have a larger die than ARL, so it makes less sense to use it as a ramp vehicle (doubly so because they said they wanted a fast follow after SRF). As for N3E, if intel 3 does indeed have better performance per watt, then there is no reason for intel to use that in their server products either. As for the density gap, if we look at the HP cells that intel is so fond of, intel 4/3 3 fin device is barely behind N3E's 3 fin device.
wikichip has even closer numbers for the HP cells, but I have no clue where they got their N3E numbers from, so I don't want to definitively source them (plausible as there numbers are).
Intel 4 Deep Dive - Semiwiki
As I previously wrote about here, Intel is presenting their Intel 4 process at the VLSI Technology conference. Last Wednesday Bernhard Sell (Ben) from Intel gave the press a briefing on the process and provided us with early access to the paper (embargoed until Sunday 6/12). “Intel 4 CMOS...semiwiki.com
Given we don't know about power consumption, I'm going to defer to Daniel's judgment, and history, and assume the power efficiency isn't going to match TSMC's. He seems to have insight into this segment I certainly don't. But, you're right, it's not factual, and if it is very power efficient, well, it's very good for servers. But, that's a really big if, and seems unlikely, given what has been said here.
You misunderstand Dan and my explanation of power. The power bands we were talking about were for mobile/wearables. For DC/laptop intel's process are capable of matching TSMC (just look at intel's 10/7 products vs AMD's N7P products). Low power in foundry terminology means smaller than an Ipad. Desktop is comical high power. Laptop is flexing between mid power and high power bands. DC is running all out at medium-low voltages. You also misunderstand intel's claims. The claim has always been "unquestioned performance per watt leadership". It has never been claimed that intel will have the highest frequency CPUs. For proof of this look at the intel 4 whitepaper. The 20% improvement was near the bottom of the power band.
One SPR is a server product not a desktop product. Two server products take MUCH longer to validate than client parts. Three SPR is on intel 7 not intel 4. Four as I stated, why would intel make new products on intel 3 for client products when 20A should be ready in time.
Yes it should be easy, they are the same node. It seems you have a fundamentally flawed understanding of what intel 4 and 3 are. To quote intel in 2022 "intel 3 will add the following to intel 4: denser design libraries, increased transistor driver current, reduced via resistance, increased use of EUV, and an 18% improvement to performance per watt". Given what we know now about intel 4 it is best to think of intel 3 as intel 4+ with the rest of the libraries/any missing devices now enabled in the PDK.
Emerald rapids is a different chip, and no SPR is not a refresh. It is a new core, on a new platform, in a new disaggregated SOP, on intel 7.
By this logic why not use 20A?
I'm assuming you mean some ADL-N 8 e-core cheapo laptop part. Intel hasn't announced such a cost optimized part, so it is a bit premature to talk about one of those parts not already using intel 3 being proof intel 3 being a disaster.
Intel using server parts for it is a good indication, since this is the BU that needs every edge they can get to catch up to AMD/ARM.
18A is the only IFS node with no commited customer orders at this time. Intel 16 and 3 both have signed customers.
TSMC being 90% of the 3"nm" foundry market doesn't mean intel 3 is bad. What it means is that intel's roughly N5 density foundry competitor node (intel's own words on how they expected their "7nm process technology" to compete with TSMC's then most advanced process) is coming out 3 years later. It also means that unsurprisingly intel's ecosystem is not as extensive as TSMC's.
By this logic was 32nm SOC a bad 32/28nm node because client moved to 22nm for ivybridge rather than using the low power optimizations that intel added to 32nm SOC?
Your understanding is flawed because you say that intel 3 is uncompelling but intel 4 is good. They are the same node, yet you act as if they are different things. On your point of the overall success of the technology, it is WAY too early for outsiders to call. If we see few if any trailing edge products (be they internal or external) using intel 3 post MTL/GNR/SRF then we can confidently say intel 3 was a "not a great node for them". But unless we see intel ireland winding down it's intel4/3 line in a couple of years, then I refer you to my 32nm SOC only being used by phones/chipsets rather than having a sandy bridge refresh example.
I agree that the reasons why 18A doesn't have any external customers right now when the others do is understandable. Regardless, it is patently false to say there are no external customers for intel 3 and imply that 18A does. As for focusing on intel 3 for IFS, it has always been talked about less than 18A (for blindingly obvious reasons). 20/18A has always been the thing intel can't stop talking about from the day intel accelerated was announced (back then they were funnily enough talking about foundry commitments on 20A presumably because 18A wasn't part intel accelerated).
My point on this was that ADL-N was not something intel announced until it was out. If intel wants some ADL-N next type thing on intel 4/3, then we wouldn't know about it until it launches. Additionally as I've stated multiple times, it makes no sense to make a full top to bottom lineup in addition to the current ARL skus.
Claims that intel 3 is not a compelling technology because it isn't as dense as N3E is disengous when it was originally intended to compete with N5. By this logic 28nm was uncompeling because 22nm had a WAY bigger PPA lead than N3E is presumed to have over intel 3. Based on techinsight's numbers intel 3 is 2/3 of N3E. Technologically speaking there is nothing inherently wrong with intel 3. Assuming intel delvers, I would even go so far as it being excellent for what it is. When was the last time TSMC delivered 18% PPW from an annual performance kicker? At the end of the day though, this doesn't really matter to end customers because foundry is not about technology (it never has been), it is about service. Even if we wind back the clock to 2021, intel 3 would still probably be a loser vs N5 in foundry because TSMC has the best service and ecosystem. My larger point is that for what it is meant for, intel 3 is good. But between the still nascent IFS ecosystem, still in progress IDM2.0 shift, and late TTM intel 3 has a lower value prop than the on paper numbers would suggest.
As for talking more about intel 3 than 18A, then I guess we are living in different universes then. For intel accelerated they had QCOM talking up 20A. were talking about. In early 2022 they had the following slides:
Was 14nm worse than 16FF then because it had 0 intel custom foundry customers? Additionally, as a recall TSMC got big even while they were behind the rest of the industry. UMC runs quite the tight operation in spite of being behind TSMC for almost all of it's existence.
If your thoughts were that intel 3 was going to start some sort of coup like 14LPP did, and that the node intel said will be most comparable to foundry 5"nm" technology was magically going to get denser once the name changed from 7nm to intel 3; then we have nothing to talk about.
