This is Lunar Lake — Intel’s utterly overhauled AI laptop chip that ditches memory sticks

4 Giugno 2024 News, Rassegna Stampa

Last year, Intel boasted that its Meteor Lake processors, dubbed Core Ultra, represented the company’s biggest architectural shift in 40 years. But Intel didn’t settle down after that: today it’s revealing how Lunar Lake, its next laptop chip coming this fall, will overhaul the formula yet again.

Facing the existential threat of Arm and the opportunity of AI PCs, Intel has apparently ditched its famous tick-tock cadence for a whole new system-on-chip design, one that not only triples the size and more than quadruples the performance of its AI accelerator, but promises up to 14 percent faster CPU performance at the same clockspeed, 50 percent more graphics performance, and up to 60 percent better battery life than last year’s model.

“It’s x86 power like you’ve never seen it before,” claims Intel technical marketer Rob Hallock, who says Intel tweaked every part of the chip to make it happen. He says it’ll “definitely” beat Qualcomm, too.

a:hover]:text-gray-63 [&>a:hover]:shadow-underline-black dark:[&>a:hover]:text-gray-bd dark:[&>a:hover]:shadow-underline-gray [&>a]:shadow-underline-gray-63 dark:[&>a]:text-gray-bd dark:[&>a]:shadow-underline-gray”>Image: Intel

The biggest change? If you buy a Lunar Lake laptop, it won’t have separate memory sticks or chips! Lunar Lake now bakes 16 or 32GB of LPDDR5X memory into the package itself, with no ability to connect more RAM. It’s a change that reduces the power consumption of moving data through the system by approximately 40 percent, according to Intel. For those who need more memory, Hallock says a separate Arrow Lake architecture is coming to laptops later this year.

After hours of poring through slide decks and presentations, plus a quick chat with Hallock, here’s everything else I just learned.

Last year’s Meteor Lake contained a wild new “3D performance hybrid architecture” with loads of Performance (P), Efficiency (E), and even a pair of brand-new Low Power Efficiency (LP-E) cores on a separate tile dubbed the “low power island.”

That island was built like a smartphone, a first for Intel, with its own Wi-Fi, Bluetooth, display controllers, memory, and those low-power CPU cores. The idea: you could theoretically save battery life by never heating up other tiles and bigger cores unless you’re doing bigger tasks.

But it didn’t work. Apps like Microsoft Teams wound up warming the entire chip. So Intel is axing the LP-E cores after only one year — in favor of a new 4 by 4 system. You get up to four new “Lion Cove” P-cores and four new “Skymont” E-Cores in a Lunar Lake chip. Those E-cores now run as fast as an LP-E core at one-third the power, or scale to 2x or 4x the performance (single-threaded vs. multi-threaded).

And with a new thread director, Windows can now create “containment zones” that actually keep “most real workloads” on the Skymont E-cores, Microsoft and Intel claim.

“This is key to Lunar Lake battery life: we can run more workloads in a lower power environment on a lower power core with fewer things turned on, and still give you a great user experience,” says Intel fellow Stephen Robinson.

Microsoft Teams uses 35 percent less power on Lunar Lake thanks to those changes, Intel claims — though Intel says it can’t yet translate that to hours for me.

Similarly, Intel has finally axed Hyper-Threading because the SMT technology eats more power and real estate than it’s worth. “Adding more cores is slightly more die area” than the doubled portions of circuitry needed to make HT work, Hallock admits, but he says the E-cores are so compact and capable now that HT simply no longer makes sense.

Speaking of capable E-cores, Intel’s Skymont has another surprise: this year’s E-core is more powerful and efficient than last year’s P-core at typical laptop clockspeeds — with up to 20 percent more single-threaded performance. It just can’t scale up to nearly as many gigahertz:

The four “Lion Cove” P-cores, meanwhile, offer a 14 percent performance increase clock for clock, though Intel wouldn’t provide clockspeed numbers so we can truly compare. But overall, says Hallock, performance is “generationally very significantly up” compared to last year.

In the GPU realm, Intel is even more sure of itself: the company says its Xe2 GPU offers 1.5x the graphics performance of Meteor Lake (in 3DMark Time Spy) which itself was 2x the performance of the previous generation. It’s still got the same number of Xe cores and other functional units, but with a variety of performance and efficiency improvements.

And while assuredly not all Lunar Lake chips will be equal, Intel says it didn’t have to split its Xe laptop GPU into two different flavors for lower and higher wattage: Xe2 can now scale across the spectrum of light and medium-weight laptops all by itself. Intel also says the GPU offers 67 TOPS of AI performance, in addition to the NPU.

Meteor Lake didn’t truly kick off the AI laptop generation the way that Intel hoped. In fact, you could argue it left early adopters in the cold — with just 11.5 TOPS worth of AI acceleration, their NPU falls far shy of Microsoft’s 40 TOPS requirement for Copilot Plus PCs.

But Lunar Lake triples the amount of NPU hardware on the die, doubles the memory bandwidth, and boosts the clockspeed from 1.4GHz to 1.95GHz —offering up to 48 TOPS and an estimated 2x to 4x performance overall.

The tripled hardware does draw a little more power, but Intel says it’s substantially faster: just 5.8 seconds for 20 iterations of Stable Diffusion for Lunar Lake vs. 20.9 seconds for Meteor Lake, while pulling 11.2 watts instead of 9 watts previously.

Intel says its software partners are currently building 350 AI features for PCs due through 2025.

Lunar Lake now natively supports H.266 VVC video for an additional 10 percent filesize reduction over AV1 “at the same quality”.

You get Wi-Fi 7 and Bluetooth 5.4 baked into the chip — though it still requires a companion PCIe module for the physical radios and antenna connectors.

Intel claims it takes 55 percent less time to wake up wireless when waking up the machine.

eDP 1.5 with “panel replay” and other techniques can save up to (emphasis on “up to”) 351mW of power by not repeatedly drawing the same images on screen.

There’s a new dedicated Partner Security Engine on the chip that’s “effectively Microsoft Pluton,” Intel says, with its own processor, fuse, and crypto portion.

Lunar Lake’s onboard memory means motherboards can and will shrink — “there are some interesting design win decisions coming,” says Hallock.

Lunar Lake’s Compute Tile is indeed built on TSMC’s N3B processor, with the platform controller on TSMC’s N6, though Intel says it does all the design, assembly, and packaging.

Intel says it can dynamically adjust the speed of the RAM to reduce Wi-Fi interference.

Intel says it can dynamically adjust clockspeed in 16.67MHz intervals (down from 100MHz) to optimize performance.

Every Lunar Lake system gets two Thunderbolt 4 ports — “You are guaranteed a minimum of two Thunderbolt 4 ports on every Lunar Lake system that you touch,” says Hallock.

Intel, like Qualcomm, will sell a Mac Mini-like AI PC development kit later this year — one that Intel says will be upgradable to Panther Lake chips when they’re available.

https://www.theverge.com/2024/6/3/24169115/intel-lunar-lake-architecture-platform-feature-reveal

This is Lunar Lake — Intel’s utterly overhauled AI laptop chip that ditches memory sticks

Last year’s Meteor Lake contained a wild new “3D performance hybrid architecture” with loads of Performance (P), Efficiency (E), and even a pair of brand-new Low Power Efficiency (LP-E) cores on a separate tile dubbed the “low power island.”

Speaking of capable E-cores, Intel’s Skymont has another surprise: this year’s E-core is more powerful and efficient than last year’s P-core at typical laptop clockspeeds — with up to 20 percent more single-threaded performance. It just can’t scale up to nearly as many gigahertz:

Meteor Lake didn’t truly kick off the AI laptop generation the way that Intel hoped. In fact, you could argue it left early adopters in the cold — with just 11.5 TOPS worth of AI acceleration, their NPU falls far shy of Microsoft’s 40 TOPS requirement for Copilot Plus PCs.

Evidenziatore

Ricerca avanzata

Speaking of capable E-cores, Intel’s Skymont has another surprise: this year’s E-core is more powerful and efficient than last year’s P-core at typical laptop clockspeeds — with up to 20 percent more single-threaded performance. It just can’t scale up to nearly as many gigahertz:

Meteor Lake didn’t truly kick off the AI laptop generation the way that Intel hoped. In fact, you could argue it left early adopters in the cold — with just 11.5 TOPS worth of AI acceleration, their NPU falls far shy of Microsoft’s 40 TOPS requirement for Copilot Plus PCs.

Evidenziatore

Tag

Ricerca avanzata

Related Post