ne plus ultra

Intel intros first Meteor Lake chips with faster GPUs and worse single-core speed

They're Intel's first CPUs with a neural processing unit for on-device AI tasks.

Andrew Cunningham – Dec 14, 2023 3:00 PM | 93

Intel's Core Ultra CPUs combine multiple silicon dies into a single chip. Credit: Intel

Intel announced technical details about its next-generation Core CPUs a few months ago. Codenamed Meteor Lake, the chips are good and bad news for Intel's chip manufacturing ambitions—they're simultaneously the first chips to use the new Intel 4 process and the first of Intel's mass-market consumer processors to use silicon manufactured by someone other than Intel (in this case, TSMC). They're also a showcase for Intel's Foveros packaging technology, which welds several pieces of silicon ("tiles," in Intel's words) rather than integrating everything into a single monolithic die.

Today, Intel is announcing the first wave of actual Meteor Lake processors, which Intel says will be available in some PCs starting today—expect to see quite a few of these PC designs announced this week and even more at CES next month.

The lineup includes 11 chips across two different product families: H-series processors for thin-and-light workstation and gaming laptops and U-series chips that will end up in Ultrabooks. Eight of these are launching today, and three more are expected in Q1 of 2024. The two families share many similarities, including Intel's first built-in neural processing unit (NPU) for accelerating machine learning and AI workloads, but in short, the H-series chips use more power and include more P-cores and GPU cores.

The 11 CPUs that Intel is announcing today. Credit: Intel

Ars Video

The slide above has all the key specs, but at a high level: the H-series chips include either six or four Redwood Cove P-cores, eight Crestmont E-cores, and two Crestmont LP E-cores. (Meteor Lake includes a pair of even lower-power E-cores in its SoC tile, allowing the main CPU tile to be shut off entirely when CPU usage is low.) They also include Intel Arc-branded, TSMC-manufactured GPU tiles with either eight or seven Xe cores, base CPU power limits of 28 W, and maximum Turbo Boost power limits of either 64 or 115 W (likely depending on the specific laptop you buy).

The U-series chips are cut back considerably to fit inside a lower power budget. They all include just two P-cores plus eight E-cores and two LP E-cores, and integrated GPUs with only four Xe cores; the Arc branding has been dropped for these despite their use of the same GPU architecture, something Intel was already doing with the Iris GPU branding in previous generations. Their base power limit is 15 W, and their turbo power limit is 57 W, though the two additional models coming in early 2024 have lower base and Turbo limits of 9 W and 30 W, respectively.

None of the chips announced today use the top-tier Core Ultra 9 branding, but the Core Ultra 9 185H will launch early next year with a marginally increased 5.1 GHz peak clock speed, higher base and turbo power usage, and most of the same basic specs as the Core Ultra 7 165H. Intel has announced non-Ultra Core 7, Core 5, and Core 3 branding for slower and cheaper CPUs, but none of these are being announced today.

The chips also include the Intel AI Boost NPU, support for either LPDDR5, LPDDR5x, or DDR5 (no more DDR4 support for Intel's chips), integrated Wi-Fi 7 and Bluetooth 5.4, Thunderbolt 4, eight PCIe 5.0 lanes for dedicated GPUs and 12 PCIe 4.0 lanes for SSDs, and HDMI 2.1 and DisplayPort 2.1 support.

Eventually, Intel will announce both higher- and lower-end Meteor Lake processors. Credit: Intel

Much of Intel's announcement centered on the NPU, but right now, their usefulness is still pretty limited, partly because so few PCs have actually included this hardware before this year (and they've mostly been running the Arm version of Windows). For now, Windows uses NPUs primarily for things like real-time background removal and other video-processing tasks.

Expanded availability in mainstream x86-based PCs, combined with Microsoft's all-encompassing push to include more AI-related features into all of its products, will probably lead to more features that can take advantage of on-device AI processing in 2024. This could be a useful privacy feature since when you use Microsoft's generative AI features right now, it generally involves sending information to servers that may use your input for further training of AI models.

Third-party apps will also be able to take advantage of the NPU. Intel says it "is partnering with more than 100 software vendors to bring several hundred AI-boosted applications to the PC market," and while the company offered few examples, it's easy to see apps like Adobe's enhancing or adding to their existing generative AI features on PCs with some local hardware acceleration.

Lower single-core performance

The technology behind the Core Ultra chips, and the new packaging technology that Intel is using to create them, may ultimately be more interesting than their actual performance. To be clear, performance looks fine, but for the CPU especially, we're talking about very mild improvements (and some actual regressions in single-core speeds) compared to Intel's 13th-gen chips, and GPU performance that's in the same ballpark as AMD's Ryzen 7040 series processors (and, presumably, the extremely similar Ryzen 8040).

For months now, leaked benchmarks of Meteor Lake chips have pointed to mixed performance improvements over 13th-generation Raptor Lake laptop chips. In its slides, Intel is careful to highlight areas where performance looks good and downplay places where it looks bad, but they basically confirm what the leaks have suggested: big gains for the integrated GPU, modest multi-core performance improvements, and a small regression in single-core performance. Note that all of Intel's provided performance numbers highlight H-series chips, and we have no idea how U-series chips will perform at this point.

Intel claims that the multi-core performance of the Core Ultra 7 165H is roughly 8 percent faster than the Core i7-1370P and 11 percent faster than an AMD Ryzen 7 7840U. Intel claims that its single-threaded performance is 12 percent faster than the 7840U but roughly 7 percent slower than the i7-1370P. This decrease can be explained partly by a roughly 4 percent decrease in clock speed, from 5.2 to 5.0 GHz. But the rest of the decrease could suggest that the new CPU architecture doesn't perform quite as well or that the new P-cores can't run as fast as the old ones for as long.

Graphics performance, on the other hand, looks great if you compare it to the roughly 3-year-old Intel Iris Xe integrated GPU, but only a shade better than AMD's RDNA 3-based Radeon 780M integrated GPU. The new GPU can be as much as two times faster than the old one, though this heavily depends on the game you're playing. Compared to the Radeon 780M, Intel says that Meteor Lake's best Arc GPU is roughly 5 percent faster across an average of 18 games. The version of the Arc GPU with 7 Xe cores instead of 8 is about the same speed as the Radeon 780M.

Intel's slides also highlight its power efficiency, with examples chosen to make Intel look particularly good. Intel says that the Core Ultra 7 165H roughly matches the Core i7-1370P's performance as long as the chips are using 18 W of power or less, but when you're using more power, Meteor Lake always gets you better performance in the same power envelope. Intel also compared Meteor Lake's performance-per-Watt favorably to the Ryzen 7840U and Apple's new M3 chip (though especially in the case of the M3, it's worth noting that Intel's peak power consumption can go a whole lot higher and that using twice as much power doesn't necessarily get you twice the performance). We'll need to run some of our own efficiency tests to compare the processors head-to-head.

Intel also says that Meteor Lake will use less power than the 7840U in all kinds of common scenarios, most notably during video streaming and when idle on the Windows desktop. Meteor Lake's video decoding hardware has also been moved into the SoC tile rather than staying in the GPU tile, allowing Intel to switch off the GPU when all you need is video playback.

Intel will likely be able to improve speeds as it iterates on Meteor Lake's design and refines both its own Intel 4 manufacturing process and the Foveros packaging process, and the GPU improvements are useful for people whose thin-and-light laptops are occasionally pressed into service as portable gaming machines. It's just too bad that the big things Meteor Lake does—graphics performance, the new NPU, the new manufacturing tech—aren't accompanied by an equally dramatic improvement to CPU performance and efficiency.

Listing image: Intel

Andrew Cunningham Senior Technology Reporter

Andrew is a Senior Technology Reporter at Ars Technica, with a focus on consumer tech including computer hardware and in-depth reviews of operating systems like Windows and macOS. Andrew lives in Philadelphia and co-hosts a weekly book podcast called Overdue.

93 Comments