For galore months, Intel has spared nary accidental to punctual america that its Meteor Lake chips would beryllium the ones to ticker — its archetypal CPU with antithetic chiplets for each component; its archetypal connected its Intel 4 process node; its archetypal with a dedicated AI coprocessor inside. Today, Intel is revealing a full batch more.
Meteor Lake volition “launch” connected December 14th, the institution present says, arsenic the astir power-efficient lawsuit processor the company’s ever made — and with up to doubly the graphics performance, a “low powerfulness island” that tin tally tasks independently, and hooks into Microsoft Windows to intelligently power the caller chips.
In nary peculiar order, present are the highlights of Intel’s Core Ultra — due to the fact that yeah, this one’s not called a Core i7.
It’s made of tiles
With Meteor Lake, Intel volition articulation its peers successful crafting machine chips retired of Lego-like gathering blocks — wherever the “CPU” and “GPU” aren’t conscionable abstracted components connected the aforesaid chip, for example, but really abstracted pieces of silicon printed astatine antithetic sizes and grafted together. The likes of AMD and Qualcomm person been doing this for a while, but heterogeneous computing is comparatively caller to Intel.
The downside for you is that lone portion of Intel’s spot is really connected the company’s bleeding-edge Intel 4 process. The graphics are connected TSMC’s 5nm process, and the I/O and caller “SoC Tile” are connected TSMC N6. But the upside is that companies tin take the champion gathering blocks — and selectively chopped disconnected powerfulness to the rest.
Right retired of the gate, nary pun intended, Intel is attempting to usage chiplets to radically trim however overmuch powerfulness its processors draw. And it’s doing truthful by taking immoderate of the “Central” retired of CPU.
Instead of a azygous CPU oregon show portion of the chip, each Meteor Lake has two — one connected a “low powerfulness island” that tin theoretically tally each by itself, with its ain ratio CPU core, NPU AI coprocessor, media engine, and memory.
The different processor cores are distributed among a “Compute Tile” connected Intel 4 that houses the P (Performance) and E (Efficiency) cores, dubbed Redwood Cove and Crestmont, respectively, positive a abstracted Graphics Tile connected TSMC N5.
Realistically, astir radical utilizing a machine will invoke the Compute Tile, but Intel wants to vigor up the spot arsenic small arsenic it can, with an enhanced “Thread Director” that lone pushes enactment to higher-power cores aft it’s tried the lowest-power ones first:
That goes for the AI coprocessor, excessively — which Windows tin natively spot and show successful the Task Manager, by the way.
Since “Media” is separated from “Graphics,” integrated encoding and decoding of video could theoretically hap without hitting the graphics tile astatine all. “An illustration of this is the summation of hardware enactment for the AV1 movie atom feature, which was antecedently executed with GPU shaders,” writes Intel.
In summation to AV1 video, autochthonal HDMI 2.1 and DisplayPort 2.1 support, and either 8K HDR video oregon up to 4 4K monitors astatine once, that SoC tile offers Bluetooth 5.4 and Wi-Fi 7. The second theoretically offers up to 5.76Gbps transportation speeds, assuming you negociate to acceptable up a 320MHz transmission successful your workplace oregon neighborhood.
Can it game?
But each those “graphics”-like tasks abstracted from GPU doesn’t mean the GPU is conscionable for show. Intel says Meteor Lake tin adhd its Intel Arc graphics — present with dedicated ray-tracing units and up to 8 Xe cores connected the spot itself.
Intel says the “Xe LPG” GPU offers up to doubly the show per watt of Xe LP, which themselves were doubly the show per watt of its 11th Gen Intel UHD that came integrated successful processors before.
They besides enactment Intel’s XeSS — its intelligent upscaler akin to Nvidia’s DLSS and AMD’s FSR — for the archetypal clip connected integrated graphics, which could assistance boost framework rates adjacent more.
While it didn’t supply immoderate show numbers, Intel says the GPU tin “run astatine a overmuch little minimum voltage and scope a overmuch higher maximum timepiece speed” than erstwhile iGPUs, astatine good implicit 2GHz.
And Intel has developed its ain “patented low-cost vapor chamber” cooling solution, the institution claims, to assistance get gamer and creator laptops with Meteor Lake retired successful the world.
Just cognize that not each Meteor Lake / Core Ultra processor volition get the fancy GPU. The good people reads: “Intel Arc graphics lone disposable connected prime MTL processor-powered systems with dual-channel memory.”
But the NPU — that AI coprocessor — should theoretically beryllium successful each chip. “The NPU volition beryllium disposable crossed the afloat merchandise stack of Meteor Lake,” says Intel’s Tim Wilson, VP of architecture.
(That would enactment up with what Intel CEO Pat Gelsinger said connected a July net call: “We’re going to physique AI into each level we build.” But I haven’t confirmed if Meteor Lake volition look extracurricular Core Ultra, truthful buying a non-Ultra spot whitethorn mean nary NPU for now.)
Intel isn’t suggesting the tiny NPU volition abruptly mean generative AI tin halt moving connected elephantine unreality servers filled with Nvidia H100 chips, nor is the institution saying this 1 NPU tin bash it each portion you sleep. Instead, it’s suggesting that you present person options — it’s overmuch much businesslike to fto the NPU tally representation generator Stable Diffusion, and it’s faster and inactive decently much businesslike to person your GPU and NPU tally it together:
The institution didn’t look to person a batch of circumstantial examples for however “AI is going to reinvent however you bash everything connected Windows,” to borrow a phrase from Microsoft’s out-the-door Windows boss Panos Panay — astatine slightest beyond the aforesaid things Qualcomm’s NPUs tin already bash with Windows Studio Effects successful Microsoft Teams.
But it does person an OpenVINO Inference Engine to assistance speech to those applications and tube them straight to the NPU, and a batch of wide ideas, including:
Intel is utilizing Meteor Lake arsenic impervious that it’s executing again, arsenic portion of Gelsinger’s expansive turnaround plan for the company. The institution says yields of the Intel 4 parts of Meteor Lake (read: the Compute Tile) outstrip not lone its introductions of 14nm with Broadwell and 10nm with Ice Lake but besides its “tock” refinements to the “tick tock” look with Skylake and Tiger Lake, respectively.
None of this matters if the components don’t really perform, and Intel’s said nothing astir earthy show oregon artillery beingness successful effect to our questions. Reps repeatedly apologized connected a league call, saying they’d reply those sorts of questions person to launch. On the CPU side, the lone hints I saw were that the caller Crestmont E-cores person unspecified “IPC gains implicit erstwhile E-cores” — though results whitethorn vary, the good people warns — and that the caller Redwood Cove P-cores person “improved show efficiency.”
But Intel says it’s investing large successful this exemplary of the future. It’s putting $3.5 cardinal into New Mexico done adjacent twelvemonth and $7 cardinal into Penang, Malaysia, implicit the adjacent 10 years to put successful a “sharp ramp successful capacity” for Meteor Lake and different Foveros-style chips — the ones that three-dimensionally stack and assemble antithetic gathering artifact chips onto a azygous package. Next, it’s moving to nonstop copper-to-copper bonding, thing it’s calling Foveros Direct, and it intends to sit these chips connected glass by the extremity of the decade.