Intel Unveils "Arrow Lake" for Desktops, "Lunar Lake" for Mobile, Coming This Year

[btarunr]

Intel in its 2024 International CES presentation, unveiled its two new upcoming client microarchitectures, "Arrow Lake" and "Lunar Lake." Michelle Johnston Holthaus, EVP and GM of Intel's client computing group (CCG), in her keynote address, held up a next-generation Core Ultra "Lunar Lake" chip. This is the Lunar Lake-MX package, with MOP (memory on package). You have a Foveros base tile resembling "Meteor Lake," with on-package LPDDR5x memory stacks. With "Lunar Lake," Intel is reorganizing components across its various Foveros tiles—the Compute and Graphics tiles are combined into a single tile built on an Intel foundry node that's possibly the Intel 20A (we have no confirmation); and a smaller SoC tile that has all of the components of the current "Meteor Lake" SoC tile, and is possibly built on a TSMC node, such as N3.

"Lunar Lake" will pick up the mantle from "Meteor Lake" in the U-segment and H-segment (that's ultraportables, and thin-and-light), when it comes out later this year (we predict in the second half of 2024), with Core Ultra 2-series branding. Intel also referenced "Arrow Lake," which could finally bring light to the sluggish pace of development in its desktop segment. When it comes out later this year, "Arrow Lake" will debut Socket LGA1851, "Arrow Lake" will bring the AI Boost NPU to the desktop, along with Arc Xe-LPG integrated graphics. The biggest upgrade of course will be its new Compute tile, with its "Lion Cove" P-cores, and "Skymont" E-cores, that possibly offer a large IPC uplift over the current combination of "Raptor Cove" and "Gracemont" cores on the "Raptor Lake" silicon. It's also possible that Intel will try to bring "Meteor Lake" with its 6P+8E Compute tile, Xe-LPG iGPU, and NPU, to the LGA1851 socket, as part of some mid-range processor models. 2024 will see a Intel desktop processor based on a new architecture, which is the big takeaway here.



View at TechPowerUp Main Site

 

[Minus Infinity]

Let's not forget that clown Gelsinger unveiled Mediocre Lake in late 2021 saying launch was imminent.

 

[remixedcat]

wait.. intel has so many of thier own fabs yet goes to TSMC? something isn't right w the puddin!

 

[watzupken]

In my opinion, Intel’s messy CPU launch cadence will kill demand for their chips. What they are saying is we have Meteor Lake in first half of 2024, but something newer is coming up in second half. Then why would people want to buy Meteor Lake when it’s going to be something better not too far off? It may make sense that they want to stir up excitement on their products, but it will also make encourage people not in a rush to wait.

 

[cvaldes]

In my opinion, Intel’s messy CPU launch cadence will kill demand for their chips. What they are saying is we have Meteor Lake in first half of 2024, but something newer is coming up in second half. Then why would people want to buy Meteor Lake when it’s going to be something better not too far off? It may make sense that they want to stir up excitement on their products, but it will also make encourage people not in a rush to wait.

There's always something newer and better around the corner.

I know this is very difficult for many TPU forum participants to grasp but a lot of enterprise purchasing is dictated by budgetary cycles and not every company uses the calendar year as their fiscal year. Joe Consumer -- on the other hand -- may not be subject to a "use it or lose it" budget spending mandate.

And yes, a large part of Intel's revenue comes from enterprise customers. It's not just Joe Consumer walking the CPU aisle at their local Microcenter. It's also about that purchasing agent at the US General Accounting Office who needs to turn in a purchase order for thousands of desktop and notebook PCs before the federal fiscal year (September 30th) is over.

And not every state follows the federal fiscal calendar. The State of California's fiscal year begins on July 1st. And there's no requirement to follow calendar quarters or months either.

Nvidia itself has a fiscal year that ends near the end of January. Apple's ends on the last Saturday of the ninth month, September (all of Apple's quarters are exactly thirteen weeks long with the rare 14-week quarter every few years).

Anyhow this forces corporate purchasing departments to set a window in which to place an order to ensure the expense hits the proper fiscal year. And remember that CPU and GPU releases aren't joined at the hip either so often these purchases are forecasted months ahead of time. It's not Joe IT Guy driving to their local Best Buy to pick up a bunch of retail CPU, GeForce, and Radeon boxes.

 

[remixedcat]

thing is all the confusion isn't good for companies like dell, lenovo, system 76, framework, hp, asus, acer, etc.... I hope that it makes them all get more ryzen systems!

 

[Solaris17]

lot of enterprise purchasing is dictated by budgetary cycles

and limits. Iv blown $xxx,xxx just so I get the same budget the next year. if you dont then your cap is lowered, which REALLY sucks in year 3 or 5, when contracts come back or pricing increases.

 

[A&P211]

There's always something newer and better around the corner.

I know this is very difficult for many TPU forum participants to grasp but a lot of enterprise purchasing is dictated by budgetary cycles and not every company uses the calendar year as their fiscal year. Joe Consumer -- on the other hand -- may not be subject to a "use it or lose it" budget spending mandate.

And yes, a large part of Intel's revenue comes from enterprise customers. It's not just Joe Consumer walking the CPU aisle at their local Microcenter. It's also about that purchasing agent at the US General Accounting Office who needs to turn in a purchase order for thousands of desktop and notebook PCs before the federal fiscal year (September 30th) is over.

And not every state follows the federal fiscal calendar. The State of California's fiscal year begins on July 1st. And there's no requirement to follow calendar quarters or months either.

Nvidia itself has a fiscal year that ends near the end of January. Apple's ends on the last Saturday of the ninth month, September (all of Apple's quarters are exactly thirteen weeks long with the rare 14-week quarter every few years).

Anyhow this forces corporate purchasing departments to set a window in which to place an order to ensure the expense hits the proper fiscal year. And remember that CPU and GPU releases aren't joined at the hip either so often these purchases are forecasted months ahead of time. It's not Joe IT Guy driving to their local Best Buy to pick up a bunch of retail CPU, GeForce, and Radeon boxes.

I prefer to shop at my local Fry's.

 

[cvaldes]

I prefer to shop at my local Fry's.

There are still some Frys?

Many of the stores in the SF Bay Area shuttered before the pandemic and the rest of the Bay Area chain were closed in 2021. But they were the walking dead long before that. Even in the final year of the Sunnyvale Frys, half of the store was curtained off and the aisles were twice as wide as before. And many of the shelves were barren. A lot of suppliers would not put their products on the shelves unless they were paid in advance, no credit terms.

If there are still some Frys most likely they are renting the brand name.

I prefer to shop at Central Computer. Those guys don't suck.

But again GAO purchasing agents don't shop at mom-and-pop PC stores either.

Intel releases new silicon following their own timetable. Customers -- whether they be enterprise or consumer -- still have a choice.

And there's nothing that says that any given organization needs to buy all of their gear from one sole vendor. The GAO purchasing agent could buy a selection of various systems with an assortment of components from different suppliers. In fact, this is likely the case.

So if Arrow Lake and Lunar Lake don't fit into a corporate purchasing cycle, there will be something else for them when that cycle is up. Much of Intel's attention is (rightfully) on datacenter anyhow.

 

[Squared]

From what another outlet reported, it sounded like Intel touted AI capabilities for Arrow Lake and a new super efficient architecture for Lunar Lake. This seems odd since they're supposedly both getting Lion Cove. I'm beginning to wonder if Arrow Lake is just Meteor Lake-S on a higher clock speed node.

 

[R0H1T]

And yes, a large part of Intel's revenue comes from enterprise customers.

And that number is going down, especially since Covid. I know it's hard to understand but not everyone needs Intel Inside® also I bet (some)people are eventually getting fired for buying Intel.

 

[remixedcat]

And that number is going down, especially since Covid. I know it's hard to understand but not everyone needs Intel Inside® also I bet (some)people are eventually getting fired for buying Intel.

Lots of data centers are getting epyc too lol

 

[Eternit]

In my opinion, Intel’s messy CPU launch cadence will kill demand for their chips. What they are saying is we have Meteor Lake in first half of 2024, but something newer is coming up in second half. Then why would people want to buy Meteor Lake when it’s going to be something better not too far off? It may make sense that they want to stir up excitement on their products, but it will also make encourage people not in a rush to wait.

It was similar with Haswell refresh, Broadwell and Skylake.
I think Intel knows their 7nm (Intel 4) is not working well and they don't want to spend money on fixing it when 20A works better and is almost ready. They just want to have some return from inwestment in 7nm to avoid shareholders' frustration.

 

[Daven]

I predict Arrow Lake will use Meteor Lake p and e cores but change something insignificant to justify a new name. Looks like Lunar Lake is just a reshuffling of tiles like deck chairs on the titanic. This doesn’t look good for Intel.

 

[dyonoctis]

In my opinion, Intel’s messy CPU launch cadence will kill demand for their chips. What they are saying is we have Meteor Lake in first half of 2024, but something newer is coming up in second half. Then why would people want to buy Meteor Lake when it’s going to be something better not too far off? It may make sense that they want to stir up excitement on their products, but it will also make encourage people not in a rush to wait.

If you really need a new computer now, you will get one now. If you can wait until September 2024, then it means that you want a new computer, but didn't "need" one. *shrug*
MTL was also delayed, and honestly looks like a proof of concept right now. MTL launching as planned after ADL would have made the launch less awkward

 

[Squared]

I never imagined that a complaint about two processor lines would be that they are released just one year apart. Usually this is what customers want, so much so that CPU makers rebrand existing products when they don't have a new one (like Ryzen 8000 series).

 

[AMDK11]

From what another outlet reported, it sounded like Intel touted AI capabilities for Arrow Lake and a new super efficient architecture for Lunar Lake. This seems odd since they're supposedly both getting Lion Cove. I'm beginning to wonder if Arrow Lake is just Meteor Lake-S on a higher clock speed node.

From what I know, LunarLake is supposed to have, among others: new IGP
and NPU, and the x86 core board is to be based on a newer lithographic
process. According to Intel, this is what I would describe as a new
architecture. Intel, however, does not reveal anything about the P
cores. Both ArrowLake and LunarLake are based on x86 LionCove cores,
with the difference that LunarLake will reportedly have LionCove+, which
will receive a larger L1.

The RedwoodCove (MeteorLake) core is basically a slightly improved
Golden/RaptorCove with a larger L1-I of 64KB in 16-way mode instead of
32KB in 8-way mode. I suspect that in LunarLake LionCove+ will get 64KB
L1-D of 16-Way this time instead of 48KB of 12-Way.

Moreover, from what I have heard and from previous leaks, LionCove will receive, among others, an 8-way x86 decoder (GoldenCove 6-way) and ROB 700+ (GoldenCove ROB 512).Apparently there will be 3 MB L2, divided into 512 KB with low latency and the remaining 2.5 MB with a higher delay.As for L3, it is speculated that in LionCove it will be about 4-5 MB instead of 3 MB.In terms of the number of transistors, the core is to be much larger than Golden/Raptor/RedwoodCove (+40-50 %).

 

[Squared]

Both ArrowLake and LunarLake are based on x86 LionCove cores,
with the difference that LunarLake will reportedly have LionCove+, which
will receive a larger L1.

The last time Intel kept the microarchitecture but changed the L1 cache, the name changed from a radiant yellow thing to a tree thing (Sunny Cove -> Willow Cove). So I'd expect Arrow Lake to get Lion Cove (named for a lion's golden mane, I assume) and Lunar Lake to have some sort of "tree" Cove if these rumors are true. (Sunny Cove on 14nm was Cypress Cove, Golden Cove on Intel 4 is Redwood Cove. The only exception is Raptor Cove and that's probably just because Meteor Lake follows it and someone at Intel is really obsessed with dinosaurs.)

 

[AMDK11]

The last time Intel kept the microarchitecture but changed the L1 cache, the name changed from a radiant yellow thing to a tree thing (Sunny Cove -> Willow Cove). So I'd expect Arrow Lake to get Lion Cove (named for a lion's golden mane, I assume) and Lunar Lake to have some sort of "tree" Cove if these rumors are true. (Sunny Cove on 14nm was Cypress Cove, Golden Cove on Intel 4 is Redwood Cove. The only exception is Raptor Cove and that's probably just because Meteor Lake follows it and someone at Intel is really obsessed with dinosaurs.)

SunnyCove and WillowCove, when it comes to x86 logic, are essentially the same core, but not exactly. CypressCove is simply Sunny (10nm) moved to 14nm. Sunny/CypressCove is 512KB inclusive L2, as is Skylake 256KB. The difference in WillowCove is that L2 is changed to 1.25 MB non-inclusive. Like WillowCove (TigerLake), the L2 cache does non-inclusive only (AlderLake/RaptorLake/MeteorLake).

RaptorCove differs from GoldenCove only in larger L2 because it received 2MB instead of 1.25MB, so the x86 itself is the same.RedwoodCove is some x86 modifications, although they are so small that they do not result in a noticeable increase in IPC.

As for LionCove+, Intel will certainly provide a different name, but until we know the specifics, assuming it will only be an L1 modification, we will conventionally call this modification LionCove+.