Intel's unannounced Core Ultra 7 254V Lunar Lake chip leaks out in new benchmarks — scores worse than entry-level Core Ultra 5 228V in early multi-core tests, but on par in single-core

abufrejoval said:

Castrating chips isn't really a money maker per se.

True, and the tile approach could be ensuring very high yields for Lunar Lake. But Intel loooves to do it. I imagine Alder Lake-N has high yields as a small monolithic die, but as far as I can tell >50% of the 8-core dies have been neutered into quad-cores (and a few dual-cores for the embedded market). With no 6-core option considered. It's a bit bizarre. But it's going to make the Wildcat Lake launch more interesting for me to watch.

abufrejoval said:

To me it makes zero sense, that LL production is being ramped up now, to my understanding it's always been an expensive chip to make, so if any money was to be made, it was at the initial hype end, not in the fading dusk. At least not unless TSMC is drastically reducing the cost of LL wafer starts, but even that would only be part of the total. No real reason those stacked DRAM chips should get drastically cheaper, and the packging pipeline is probably one of the trickiest beasts to manage.

They admitted the Lunar Lake memory on package approach was too expensive, but is it dissimilar to what Apple is doing?

As for what a 254V is, my guess based on the lineup is it will just be a 256V (16 GB) with even lower clocks, but retaining the full 8 Xe cores and 12 MiB L3. As the cache is not tied to CPU core counts like they usually do, maybe they can split the difference and only enable 10 MiB.

usertests said:

True, and the tile approach could be ensuring very high yields for Lunar Lake. But Intel loooves to do it. I imagine Alder Lake-N has high yields as a small monolithic die, but as far as I can tell >50% of the 8-core dies have been neutered into quad-cores (and a few dual-cores for the embedded market). With no 6-core option considered. It's a bit bizarre. But it's going to make the Wildcat Lake launch more interesting for me to watch.

That's been one of the reasons, I've never gotten an AD Atom: the N300 was overpriced and everything else most likely neutered 8-core dies, too bad Fritzchens Fritz never took a close look!

I am really missing a floor plan for those Atom dies, because I always wanted to know of much of their real-estate is actually in the cores. Atoms are so I/O heavy that with the later process nodes they could have become entirely I/O pad bound, which would have proven the intentional castration theory even more.

And that was interesting because I once found a guy responding to an Anandtech article who said that Intel had been forbidden from excessive castrating chips for competitive reasons by some West Coast court: never saw that mentioned or discusses again.

But then Atoms themselves were mostly considered "counter revenue" and a defensive measure against ARM eating up x86 territory from the rear. And while they seem to have been very cheap to produce for Intel, OEMs hated them to the point of revolt, because they couldn't make money with them, but were forced to turn them into product via Intel package deals.

But that is a type of battle Intel can no longer do, because that was only possible with cheap self-fabbed chips and their power over OEMs while AMD was weak. And, actually one might ask if beginning that war was a big part of its downfall after.

But with all that changed, culling good chips for Lunar Lake seems crazy and if it's not barrel scraping, evidently my "analysis" is much worse than I thought.

usertests said:

They admitted the Lunar Lake memory on package approach was too expensive, but is it dissimilar to what Apple is doing?

Of course, it's somewhat similar. And one might argue that AMD's Strix Halo take is the smarter way of doing it.

But both Intel and AMD in that case made the mistake of believing that copying the Mx "Lego" approach would work outside the Fruity Cult's premium market. Apple can ask HBM prices for DDR technology approaching GDDR speeds at the very top, Intel and AMD can't, because they don't operate in a walled garden.

And on top they can't reach the 512-bit Ultra rank performance the Fruity Cult can sell at those HBM prices.

AMD at least doesn't share Intel's cost structure for the DRAM, so it's perhaps the smarter approach, but with what I believe to be bespoke chips for Strix Halo CCDs and IOD, it's a very expensive design if they can't sell it at scale. And at the current asking price and quite a lot of its life-cycle already run down, I don't see them ramping to the volumes they'd need.

Ultimately both Lunar Lake and Strix Halo may turn into very bloody noses for both vendors, although AMD at least has the production cost advantage, but less of a premium market and too much last gen dGPU competition.

usertests said:

As for what a 254V is, my guess based on the lineup is it will just be a 256V (16 GB) with even lower clocks, but retaining the full 8 Xe cores and 12 MiB L3. As the cache is not tied to CPU core counts like they usually do, maybe they can split the difference and only enable 10 MiB.

The LP-DDR5X stacks Intel uses for Lunar Lake may be bespoke co-designs, don't recall who made them, but they may also be inventory burning holes into somebody's balance sheets today. No idea if Intel had to pay for them up-front or if a RAM vendor is now sharing losses, but I can see some people not getting the bonus they were hoping for.

Perhaps the worst is that I'm not sure fire-sale Lunar Lake NUCalikes may be a good buy, because software support for the likes of Windows 12 or Linux may go out of the window with staff cuts and too much of LL being very specific, a similar fate to the current Snapdragon Elite notebooks, that have run out of software support basically from the start.

abufrejoval said:

That's been one of the reasons, I've never gotten an AD Atom: the N300 was overpriced and everything else most likely neutered 8-core dies, too bad Fritzchens Fritz never took a close look!

Despite Intel's disappointing choices, I think the quad-core systems being dumped on the market for around $100-150 were still a good option for some people. It's like a modernized low-power Skylake chip with better integrated graphics and video capability. They definitely acted as an effective counter to Raspberry Pi, inspiring videos like Jeff Geerling's When Did Raspberry Pi become the villain?.

But with the apparent successor Wildcat Lake around the corner and being enabled in Linux over the past few months, I wouldn't recommend it anymore. Wildcat Lake could easily double single-threaded performance, which is massive for the low-end. If it uses Xe3, doubled graphics performance is also plausible, putting it in performance territory that was previously unreachable by Alder Lake-N, and more similar to Alder Lake-U.

We have this Alder Lake-N mockup, and maybe that's it?! https://locuza.substack.com/p/info-snack-alder-lake-m-raptor-lake

abufrejoval said:

But that is a type of battle Intel can no longer do, because that was only possible with cheap self-fabbed chips and their power over OEMs while AMD was weak. And, actually one might ask if beginning that war was a big part of its downfall after.

Reuters and others are asking questions about 18A yields. A recent MLID leak showed a bunch of products fabbed at TSMC with use of 18A only for some low-end dies/tiles. So it's plausible that Wildcat Lake could end up being the majority of what Intel produces on 18A before they attempt to make 14A the real TSMC competitor.

But even if they eat their own dog food, 18A must be more expensive than the previously used 10nm ESF nodes, and that will affect pricing. It might be worth it if it delivers huge perf/efficiency gains.

abufrejoval said:

Perhaps the worst is that I'm not sure fire-sale Lunar Lake NUCalikes may be a good buy, because software support for the likes of Windows 12 or Linux may go out of the window with staff cuts and too much of LL being very specific, a similar fate to the current Snapdragon Elite notebooks, that have run out of software support basically from the start.

I've seen the Phoronix headlines about Intel Linux people being fired, but I don't see any scenario where the bespoke x86 chip faces a similar fate as Snapdragon Elite. Is there really much more work that needs to be done on Lunar Lake specifically?

Intel Lunar Lake sees up to 20% performance boost in Linux vs. Windows with iGPU caveatsASUS WMI Fix Submitted For Linux 6.12-rc5 To Handle Lunar Lake Performance IssueIntel Xe2 Lunar Lake Graphics Compute / OpenCL Performance Looking Great
From what I've seen on Slickdeals, the lowest end Lunar Lake laptops (226V) have hit the $500-600 range. I'd love to see it go even lower, but you are getting a gaming experience similar to Strix Point's Radeon 890M graphics which can easily command such pricing.

The easiest 32 GB model to get to has been the 258V, at around $750. That's a shame since you are looking at paying +50% or more for virtually no performance gain, only the doubled RAM being significant.

usertests said:

Despite Intel's disappointing choices, I think the quad-core systems being dumped on the market for around $100-150 were still a good option for some people. It's like a modernized low-power Skylake chip with better integrated graphics and video capability. They definitely acted as an effective counter to Raspberry Pi, inspiring videos like Jeff Geerling's When Did Raspberry Pi become the villain?.

I'm not sure the RP was ever considered much of a threat, but all those Android set-top boxes might have become a much bigger market without those Atoms.

usertests said:

But with the apparent successor Wildcat Lake around the corner and being enabled in Linux over the past few months, I wouldn't recommend it anymore. Wildcat Lake could easily double single-threaded performance, which is massive for the low-end. If it uses Xe3, doubled graphics performance is also plausible, putting it in performance territory that was previously unreachable by Alder Lake-N, and more similar to Alder Lake-U.

I'm so happy with Zens in that space, I'm simply no longer looking at Intel there. The 5825U Barcelo sells below €200 with 8 Zen 3 cores and a GPU that does well enough for any productivity work at 4k, while the newer 8845HS Hawk Point isn't that far from Strix Point in terms of CPU and GPU performance, but sells below €400 for well equipped Mini-ITX boards. 64GB of DDR4-3200 is still €150, very much the same price as DDR5-5600 so all that DDR4 scarcity craze seems just talk.

And at just €450 I got a Dragon Range 7945HX 16-core BD 790i Mini-ITX board from Minisforum with basically full desktop power at lower TDP that trades blows with Intel's biggest desktop CPUs, while being far more efficient.

These aren't AMD's latest, but they remain far ahead of anything Intel can offer in that market. Within a €200-450 price range you get much better CPU performance, efficiency and price and if you want a bit of iGPU bang, Hawk Point or Phoenix don't need to hide behind any Intel iGPU. None of these will tempt gamers, while the Dragon range has 28 lanes of PCIe v5 to offer, plenty for any dGPU.

And you don't have to choose between E-cores and P-cores. If it's about price you dial back a generation on the AMD side, for a similar price but much better performance. If it's about Wattage AMD allows you to make that choice mostly via the TDP setting on the board: Barcelo will go from 10 to 35 Watts, Hawk Point/Phoenix 15-55 Watts and while my Dragon Range seems to turbo to 100 Watts, it can also be throttled to 45 Watts, I believe.

No need to sacrifice core counts or PCIe lanes when an octa 5825U can be had for less than €200, none of the P/E scheduling nonsense you really don't need with a hypervisor, very good Linux support and only Microsoft server editions suffer from AMD cheaping out on signing drivers with the proper keys.

usertests said:

We have this Alder Lake-N mockup, and maybe that's it?! https://locuza.substack.com/p/info-snack-alder-lake-m-raptor-lake

Thanks for the mockup, seems to confirm my theory that the cores are such a small part of that Atom SoC, that it would be very hard to explain how defects piled up in all those octa-cores... pathetic market manipulation and I'm glad I never bought one.

usertests said:

Reuters and others are asking questions about 18A yields. A recent MLID leak showed a bunch of products fabbed at TSMC with use of 18A only for some low-end dies/tiles. So it's plausible that Wildcat Lake could end up being the majority of what Intel produces on 18A before they attempt to make 14A the real TSMC competitor.

But even if they eat their own dog food, 18A must be more expensive than the previously used 10nm ESF nodes, and that will affect pricing. It might be worth it if it delivers huge perf/efficiency gains.

I've not come across Wildcat Lake data, perhaps because nobody really believes they'd have something interesting to offer there. Intel can't make two different core designs deliver any significant benefit, perhaps not even with the biggest hyperscaler designs where custom ARMs are hurting the hardest. They wind up with too many SKUs, too many different dies at low volumes and high cost.

usertests said:

I've seen the Phoronix headlines about Intel Linux people being fired, but I don't see any scenario where the bespoke x86 chip faces a similar fate as Snapdragon Elite. Is there really much more work that needs to be done on Lunar Lake specifically?

In all cases you need a certain critical mass. If there is only one developer left for every component, that's dead men walking. And once they start, they won't automatically stop at the minimum Intel needs to keep things running, developers are social and tend to flock here or elsewhere. My company is only software, but some cuts look much better on paper than when all of a sudden a critical service just dies completely, because institutional knowledge got lost.

In Intel's case I'm not even sure that all the design documents developers need were written before the hardware engineers got fired or left. And sometimes you just need to ask someone, who may no longer be there.

You can't reverse engineer the stuff that makes a SoC as complex as LL even come up. I've gone and read much of the initialization code for my little Rockchip based Orange Pi 5+ in search for the place we I need to activate the 60Hz option for the 4k display and it's thousands of registers that need to be set just so, for any SoC these days to just power on. You can't just hire a kid from college and expect working results before he seeks greener pastures.

And no, you can't train an AI on this, because there is only one correct solution, that is specific for every design: you can't just half-ass and guess the rest

usertests said:

Intel Lunar Lake sees up to 20% performance boost in Linux vs. Windows with iGPU caveatsASUS WMI Fix Submitted For Linux 6.12-rc5 To Handle Lunar Lake Performance IssueIntel Xe2 Lunar Lake Graphics Compute / OpenCL Performance Looking Great
From what I've seen on Slickdeals, the lowest end Lunar Lake laptops (226V) have hit the $500-600 range. I'd love to see it go even lower, but you are getting a gaming experience similar to Strix Point's Radeon 890M graphics which can easily command such pricing.

Again, looking at the range from Barcelo with its pre-RDNA iGPU and only 15 Watts to play with at the low and vs. Hawk Point at 55 Watts on an RDNA 3 iGPU, the percentage improvements may be 'giant', easily 100% faster.

But that doesn't matter when that just means you can go from 720p to 1024p on "middle" settings for gaming but you work from a big 4k screen for productivty or pleasure. Or mobile gaming on battery power simply isn't how you pass your day.

If you're looking for GPGPU compute or gaming, Strix Halo may be the point where it gets to be remotely interesting, but still not really good, at far too high a price and not for long on battery.

Lunar Lakes' USP was the ability to work a very extended working day or perhaps even an intermittend week-end on a laptop without panicking over wall power, while the system doesn't suck completely when you need to do a bit of calculation,now and then.

That turned out to be too niche a use case for the majority of road warriors to worry about.

And those who just want to show off their Fruity Cult Rolexbooks, there was already something out there.

For laptop gaming I got a Lenovo LOQ ARP9 last year for €750. It came with a 'headless' Cezanne Zen3 APU and an NVidia RTX 4060, a really nice 1920p 144Hz display, excellent keyboard, and a very robust chassis that manages to cool 150 Watt of power without becoming too loud for company. I upgraded it to 64GB and added 4TB of NVMe and also made it my Bazzite test machine via dual-boot.

No matter what they say about 8GB GPUs, It's a remarkably able gaming machine at that resolution and a really good desktop to boot. I'm pretty sure it's better at gaming than a top notch Strix Halo, and at less than 1/3rd the price.

I've used it on the lunch table for mobile work, but it was always the chairs becoming too unconfortable for long seating before the battery gave in.

There is better and cheaper alternatives to Lunar Lake for every use case, only perhaps no better alternative for all of them.
And the same is true for Strix Halo: they all aim at not being a total desaster in far too many use cases.
And that makes them expensive, but probably not valuable for most.

usertests said:

The easiest 32 GB model to get to has been the 258V, at around $750. That's a shame since you are looking at paying +50% or more for virtually no performance gain, only the doubled RAM being significant.

How the Fruity Cult managed to use RAMxiety and storage claustrophobia to push up profits has tempted notebook vendors beyond their ability to resist.

I tend to need extra RAM for VMs, but I try only to buy at commodity prices and not pay a premium, when an upgradable option isn't available.

But I can typicall just use a different machine if I need more RAM, more cores, more GPU, more storage etc.

Those who's living situation really only allows for a single device, are much easier to milk.