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

Jun 04, 2024 10:00 AM - 5 months ago 61630

Last year, Intel boasted that its Meteor Lake processors, dubbed Core Ultra, represented nan company’s biggest architectural displacement successful 40 years. But Intel didn’t settee down aft that: coming it’s revealing really Lunar Lake, its adjacent laptop spot coming this fall, will overhaul nan look yet again.

Facing the existential threat of Arm and nan opportunity of AI PCs, Intel has apparently ditched its famous tick-tock cadence for a full caller system-on-chip design, 1 that not only triples nan size and much than quadruples nan capacity of its AI accelerator, but promises up to 14 percent faster CPU capacity astatine nan aforesaid clockspeed, 50 percent much graphics performance, and up to 60 percent amended artillery life than past year’s model.

“It’s x86 powerfulness for illustration you’ve ne'er seen it before,” claims Intel method marketer Rob Hallock, who says Intel tweaked each portion of nan spot to make it happen. He says it’ll “definitely” beat Qualcomm, too.

Here’s an early glimpse astatine a existent Lunar Lake chip, wherever you tin intelligibly spot its 2 chunks of onboard representation beneath nan main silicon.

Image: Intel

The biggest change? If you bargain a Lunar Lake laptop, it won’t person abstracted representation sticks aliases chips! Lunar Lake now bakes 16 aliases 32GB of LPDDR5X representation into nan package itself, pinch nary expertise to link much RAM. It’s a alteration that reduces nan powerfulness depletion of moving information done nan strategy by astir 40 percent, according to Intel. For those who request much memory, Hallock says a abstracted Arrow Lake architecture is coming to laptops later this year.

After hours of poring done descent decks and presentations, positive a speedy chat pinch Hallock, here’s everything other I conscionable learned.

8 cores, nary hyperthreading

Last year’s Meteor Lake contained a chaotic caller “3D capacity hybrid architecture” pinch loads of Performance (P), Efficiency (E), and moreover a brace of brand-new Low Power Efficiency (LP-E) cores connected a abstracted tile dubbed nan “low powerfulness island.”

That land was built for illustration a smartphone, a first for Intel, pinch its ain Wi-Fi, Bluetooth, show controllers, memory, and those low-power CPU cores. The idea: you could theoretically prevention artillery life by ne'er heating up different tiles and bigger cores unless you’re doing bigger tasks.

Image slider: Meteor Lake’s LP cores vs. Lunar Lake’s containment zones.

But it didn’t work. Apps for illustration Microsoft Teams coiled up warming nan full chip. So Intel is axing nan LP-E cores aft only 1 twelvemonth — successful favour of a caller 4 by 4 system. You get up to 4 caller “Lion Cove” P-cores and 4 caller “Skymont” E-Cores successful a Lunar Lake chip. Those E-cores now tally arsenic accelerated arsenic an LP-E halfway astatine one-third nan power, aliases standard to 2x aliases 4x nan capacity (single-threaded vs. multi-threaded).

And pinch a caller thread director, Windows tin now create “containment zones” that really support “most existent workloads” connected nan Skymont E-cores, Microsoft and Intel claim.

The LP-E cores “did not constrain each nan workloads we had hoped aliases wanted nan land to handle, truthful we were not getting nan afloat entitlement of artillery life and ratio we had hoped for,” says Hallock.

Image: Intel

“This is cardinal to Lunar Lake artillery life: we tin tally much workloads successful a little powerfulness situation connected a little powerfulness halfway pinch less things turned on, and still springiness you a awesome personification experience,” says Intel chap Stephen Robinson.

Microsoft Teams uses 35 percent little powerfulness connected Lunar Lake acknowledgment to those changes, Intel claims — though Intel says it can’t yet construe that to hours for me.

Similarly, Intel has yet axed Hyper-Threading because nan SMT technology eats much powerfulness and existent property than it’s worth. “Adding much cores is somewhat much dice area” than nan doubled portions of circuitry needed to make HT work, Hallock admits, but he says nan E-cores are truthful compact and tin now that HT simply nary longer makes sense.

More capacity everywhere

Speaking of tin E-cores, Intel’s Skymont has different surprise: this year’s E-core is much powerful and businesslike than past year’s P-core at emblematic laptop clockspeeds — pinch up to 20 percent much single-threaded performance. It conscionable can’t standard up to astir arsenic galore gigahertz:

“Raptor Cove for highest performance, Skymont for much capacity ISO powerfulness aliases ISO capacity astatine little power,” says Intel.

Image: Intel

Here we are zoomed into nan bluish container connected nan companion graph. Label your axes, Intel!

Image: Intel

The 4 “Lion Cove” P-cores, meanwhile, connection a 14 percent capacity summation timepiece for clock, though Intel wouldn’t supply clockspeed numbers truthful we tin genuinely compare. But overall, says Hallock, capacity is “generationally very importantly up” compared to past year.

Intel’s caller P-cores besides person much capacity astatine little power.

Image: Intel

In nan GPU realm, Intel is moreover much judge of itself: nan institution says its Xe2 GPU offers 1.5x nan graphics capacity of Meteor Lake (in 3DMark Time Spy) which itself was 2x nan capacity of nan erstwhile generation. It’s still sewage nan aforesaid number of Xe cores and different functional units, but pinch a assortment of capacity and ratio improvements.

And while assuredly not each Lunar Lake chips will beryllium equal, Intel says it didn’t person to divided its Xe laptop GPU into 2 different flavors for little and higher wattage: Xe2 tin now standard crossed nan spectrum of ray and medium-weight laptops each by itself. Intel besides says nan GPU offers 67 TOPS of AI performance, successful summation to nan NPU.

Xe2 replaces some Meteor Lake H and Meteor Lake U GPUs.

Image: Intel

A tripled NPU

Meteor Lake didn’t genuinely footwear disconnected nan AI laptop procreation nan measurement that Intel hoped. In fact, you could reason it near early adopters successful nan acold — pinch conscionable 11.5 TOPS worthy of AI acceleration, their NPU falls acold awkward of Microsoft’s 40 TOPS request for Copilot Plus PCs.

But Lunar Lake triples nan magnitude of NPU hardware connected nan die, doubles nan representation bandwidth, and boosts nan clockspeed from 1.4GHz to 1.95GHz —offering up to 48 TOPS and an estimated 2x to 4x capacity overall.

Tripled AI hardware.

Image: Intel

The tripled hardware does tie a small much power, but Intel says it’s substantially faster: conscionable 5.8 seconds for 20 iterations of Stable Diffusion for Lunar Lake vs. 20.9 seconds for Meteor Lake, while pulling 11.2 watts alternatively of 9 watts previously.

Intel says its package partners are presently building 350 AI features for PCs owed done 2025.

Everything other that caught my oculus aliases ear

  • Lunar Lake now natively supports H.266 VVC video for an further 10 percent filesize simplification complete AV1 “at nan aforesaid quality”.
  • You get Wi-Fi 7 and Bluetooth 5.4 baked into nan spot — though it still requires a companion PCIe module for nan beingness radios and antenna connectors.
  • Intel claims it takes 55 percent little clip to aftermath up wireless erstwhile waking up nan machine.
  • eDP 1.5 pinch “panel replay” and different techniques tin prevention up to (emphasis connected “up to”) 351mW of powerfulness by not many times drafting nan aforesaid images connected screen.
  • There’s a caller dedicated Partner Security Engine connected nan spot that’s “effectively Microsoft Pluton,” Intel says, pinch its ain processor, fuse, and crypto portion.
  • Lunar Lake’s onboard representation intends motherboards tin and will shrink — “there are immoderate absorbing creation triumph decisions coming,” says Hallock.
  • Lunar Lake’s Compute Tile is so built connected TSMC’s N3B processor, pinch nan level controller connected TSMC’s N6, though Intel says it does each nan design, assembly, and packaging.
  • Intel says it tin dynamically set nan velocity of nan RAM to trim Wi-Fi interference.
  • Intel says it tin dynamically set clockspeed successful 16.67MHz intervals (down from 100MHz) to optimize performance.
  • Every Lunar Lake strategy gets 2 Thunderbolt 4 ports — “You are guaranteed a minimum of 2 Thunderbolt 4 ports connected each Lunar Lake strategy that you touch,” says Hallock.
  • Intel, for illustration Qualcomm, will waste a Mac Mini-like AI PC improvement kit later this twelvemonth — 1 that Intel says will beryllium upgradable to Panther Lake chips erstwhile they’re available.

The Intel AI PC Development Kit, pinch 32GB of RAM, 2 USB-C, 2 USB-A, and HDMI.

Image: Intel

Intel says a large activity of Lunar Lake laptops will get later this year, pinch 80 different designs crossed 20 hardware partners astatine launch, including each nan biggest PC vendors — though not Microsoft, which chose to spell pinch Qualcomm’s chips for its Surface Laptop and Surface Pro instead. Intel’s customer spot leader Michelle Johnston Holthaus says each 80 designs should beryllium disposable up of nan holidays this year.

More