Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/intel-may-cancel-meteor-lake-desktop-cpus-in-favor-of-raptor-lake-refresh.17265/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

Intel May Cancel Meteor Lake Desktop CPUs in Favor of Raptor Lake Refresh

This isn't totally without precedent. When tigerlake first came to market in September 2020, it was just the 4 core U skus, with the 8 core H series came a good bit later. A bit after the 4 core die came out, Rocketlake (8 core icelake backported to 14nm) came out on desktop. In this case the decision was probably made back when it was uncertain if 10nm would be ready for HVM with larger die 8 core products. I wouldn't be surprised if it was also partially because 10nm was not yet clocking fast enough for desktop, and if back in like Jan 2020 D1 was probably the only fully ramped HVM site (constraining capacity enough to preclude a full 10nm lineup for 2020).

As for Meteorlake I would assume that intel will have it’s hands full with laptop, and if i4 can’t yet hit the absurd clocks of i7 then who cares if ipc is like 15-20% better if core counts are also smaller. Maybe the team that would have been responsible for a desktop die couldn’t get core counts to scale. Maybe intel doesn’t want 32 core parts on desktop anymore since they are launching real workstation parts for the first time since like coffeelake. If that’s the case I can’t blame them the 7900/50x and 13700/900K are crazy overkill for a consumer platform.
 
Last edited:
This isn't totally without precedent. When tigerlake first came to market in September 2020, it was just the 4 core U skus, with the 8 core H series came a good bit later. A bit after the 4 core die came out, Rocketlake (8 core icelake backported to 14nm) came out on desktop. In this case the decision was probably made back when it was uncertain if 10nm would be ready for HVM with larger die 8 core products. I wouldn't be surprised if it was also partially because 10nm was not yet clocking fast enough for desktop, and if back in like Jan 2020 D1 was probably the only fully ramped HVM site (constraining capacity enough to preclude a full 10nm lineup for 2020).

As for Meteorlake I would assume that intel will have it’s hands full with laptop, and if i4 can’t yet hit the absurd clocks of i7 then who cares if ipc is like 15-20% better if core counts are also smaller. Maybe the team that would have been responsible for a desktop die couldn’t get core counts to scale. Maybe intel doesn’t want 32 core parts on desktop anymore since they are launching real workstation parts for the first time since like coffeelake. If that’s the case I can’t blame them the 7900/50x and 13700/900K are crazy overkill for a consumer platform.
Just to go a little further back -- this also happened with Haswell Refresh (4690K/4790K) when 14nm wasn't quite ready.

I agree with all of this -- this is probably due to not having the volume and performance needed for desktop compared to their current products. Also, it looks like 13th gen is doing well enough against Zen 4 they don't need to do anything super special here.
 
What AMD and TSMC are doing is not a secret. How can Intel find its upcoming product is not competitive so late?
I think qualifiers are in order. For laptop 6+8 is fine and you can benefit from the like 40% lower power at iso performance. I see no reason why meteor lake wouldn’t out perform N5 zen4 mobile given the likely inferior ipc, worse node, fewer cores, less accelerators. This article is referencing desktop. In desktop cpus are being clocked to the moon and multithreaded workloads are more common. It is not surprising that i4 is unable to clock as well as ultra mature 10nm. Considering the only thing we’ve seen so far is 6+8 I assume the team that was responsible for a higher core count desktop die has failed to execute on time. If you can’t push i4 super hard yet either then it’ll probably just be lower power raptorlake at iso performance (obviously it could still trounce the midrange and alderlake). If this is the case, then I see no reason why they would launch anything above like an i5 for desktop with the 6+8 tile, and just focus almost all of their volume into laptops where they will have a big win until zen5 laptops become a thing.

As for how this is possible ask the design team that was responsible for a dozen SPR tape outs and almost two years of delays. I am almost more concerned for design side than I am for IFS. Same issues with ARC as well. If the desktop team is having tons of bugs then it might make sense to can it. If it is another product that is a year and a half late; then by then Arrowlake is due and what’s the point. Better to axe it and make sure Arrowlake doesn’t have issues (because it can not have issues).
 
I agree with hist78, why was a refresh inserted in desktop so late? Because it was planned all along, a shell game.

Intel has an EUV problem, only 1 site has EUV (D1X). They need HVM1 (Ireland) to really ramp up Intel 4 production, and Ireland is just starting EUV SL1 as another forum post recently highlighted.

So the marketing timelines are not realistic. EUV is a massive bottleneck and it affects everything, despite the public statements.
 
I agree with hist78, why was a refresh inserted in desktop so late? Because it was planned all along, a shell game.

Intel has an EUV problem, only 1 site has EUV (D1X). They need HVM1 (Ireland) to really ramp up Intel 4 production, and Ireland is just starting EUV SL1 as another forum post recently highlighted.

So the marketing timelines are not realistic. EUV is a massive bottleneck and it affects everything, despite the public statements.
Some quick napkin math. From google maps search D1 is bigger than Ireland with D1x being like half the site. Given it’s nature, it doesn’t have to bother making 22 and 45nm usb controllers and the like. As of right now OR, AZ, and Israel are the only 10nm sites. To date intel has 3 generations of client products, FPGAs, pontevechio, 2 generations of large die xeons, networking chips, and presumably a miriyad of other smaller products in HVM or ramping. Let us be conservative and instead of saying D1 is 33% of this capacity they are like 25%. Given the current PC slowdown it also would make sense if 10nm starts were partially reallocated to the other HVM sites so the HVM can focus on HVM and D1 can focus all of its efforts on meteorlake and future nodes.

We also know the 2+8 die is about 40mm^2, and from just looking at the die shots floating around the internet we can guestimate that the 6+8 die is about 72mm^2 (less than one third the size of alderlake and even smaller than an A series soc). Per Dan Apple starts HVM in Q2 for truly massive September launches of the iPhone. If intel really is having production on the production stepping start up in Q1, D1 only needs to make one tiny died product, on a node with fewer mask layers than 10nm, while already being one of intel’s larger sites, and the EUV utilization is more similar to N5 than to N3. I just don’t buy the line that without Ireland they can’t launch Meteorlake in any real volume.

If they are launching in Q3 or 4 then I see no reason why they wouldn’t have the wafers to land contracts for a good portion of the new thin and light designs, as well as throwing some K skus at desktop right before Black Friday. This isn’t to say that capacity won’t be constrained (it will definitely be), but we also aren’t talking GF 7nm HVM with 10-15k wspm level bad here either. My guess is that initially it won’t be any worse than normal ramps, but with subsequent capacity being slow to come online. In short I think that for Meteorlake they should be fine, but without more HVM sites Granite Rapids HVM will not be possible (doubly so if they are splitting wafers between Meteorlake which I’m guessing is at the limits of what they can spit out).
 
Last edited:
It's a strange development if Intel does cancel the desktop version of the Meteor Lake processors. The reason stated in the article is even more strange.

Seems to be real.
The other rumor (which already had a few month ago) is that laptop Meteor Lake 2023 all made by TSMC.
 
So many sketchy rumors and “leaks” right now (like the compute dies of MTL being made by TSMC… I just cannot imagine them dual-designing MTL compute dies for both Intel and TSMC processes), but likely there’s some moving around.

I put my bet on the EUV squeeze for ramp up of Intel 4/3/20A being very real, and combined with a rumor that Intel 4 cannot clock as high as the optimized Intel 7 (of course it can’t), I think Intel is prioritizing Granite Rapids and Sierra Forest on Intel 3 over any client platform.

That then would set them up to again use another client product as the test bed for their next-gen RibbonFET process on 20A, which would be Arrow Lake. In early ‘22 they claimed 20A would be manufacturing ready in 1H ’24, and so a refreshed RPL-S in 2H ‘23 and MTL-P/U in Q4 ‘23 would perfectly align with an ARL-S launch in 2H ‘24 (1 year out from RPL-S refresh, 6 months from MTL-P/U).

From all the way back in early 2020 everyone said it would take until 2024 for Intel to get their house in order, and that appears to continue to be true… but as for these rumors, Intel has made no change to the February ‘22 publicly detailed process roadmap and I think they would have to if these rumors were true. Perhaps we’ll hear more at next earnings, but I doubt it. Perhaps likely is not Arrow Lake that’s moving to TSMC N3, but Lunar Lake that’s moving to a more advanced version of N3, like their N3X
 
So many sketchy rumors and “leaks” right now (like the compute dies of MTL being made by TSMC… I just cannot imagine them dual-designing MTL compute dies for both Intel and TSMC processes), but likely there’s some moving around.

I put my bet on the EUV squeeze for ramp up of Intel 4/3/20A being very real, and combined with a rumor that Intel 4 cannot clock as high as the optimized Intel 7 (of course it can’t), I think Intel is prioritizing Granite Rapids and Sierra Forest on Intel 3 over any client platform.

That then would set them up to again use another client product as the test bed for their next-gen RibbonFET process on 20A, which would be Arrow Lake. In early ‘22 they claimed 20A would be manufacturing ready in 1H ’24, and so a refreshed RPL-S in 2H ‘23 and MTL-P/U in Q4 ‘23 would perfectly align with an ARL-S launch in 2H ‘24 (1 year out from RPL-S refresh, 6 months from MTL-P/U).

From all the way back in early 2020 everyone said it would take until 2024 for Intel to get their house in order, and that appears to continue to be true… but as for these rumors, Intel has made no change to the February ‘22 publicly detailed process roadmap and I think they would have to if these rumors were true. Perhaps we’ll hear more at next earnings, but I doubt it. Perhaps likely is not Arrow Lake that’s moving to TSMC N3, but Lunar Lake that’s moving to a more advanced version of N3, like their N3X
Intel and TSMC’s next earnings will be the most important semi events of the 2023 calendar year no doubt. The picture will clear up then.
 
Last edited:
Back
Top