Intel core ultra with AI

Not quite a revolution, but this is a solid step forward for Intel in every key area – including AI and graphics

Intel has finally released its much-trailed Core Ultra chips – codenamed Meteor Lake – in its bid to power “AI PCs”. That is, laptops and desktops with silicon that incorporates not merely a CPU and GPU, but a neural processing unit (NPU) as well. While question marks still hang over the usefulness of AI PCs, there’s a good chance a laptop you buy in 2024 will include one. The Core Ultra is now Intel’s mainstream notebook chip, ranging from the Ultra 5 125U to the Ultra 9 185H.

The processors are the first consumer chips built on the company’s Intel 4 process node, using its Foveros 3D hybrid architecture. This allows Intel to stack chiplets (it calls them “tiles”) atop one another, and it adds up to Intel’s biggest architectural shift for over a decade.

Meteor Lake has four active tiles: a compute (CPU) tile, a graphics (GPU) tile, an SoC tile and an I/O tile. The external foundry TSMC will manufacture the I/O, SoC and GPU tiles to Intel’s design, while Intel manufactures the CPU tiles on its Intel 4 process. All four of these active tiles ride on top of a single Intel-produced Foveros 3D base tile that ties together the functional units with sufficient bandwidth and low enough latency for the chip to function as close to one monolithic die as possible.

Meteor Lake has three compute units that can process AI workloads, namely the CPU, NPU and GPU. AI workloads will be directed to each unit based as required. As before, Intel has a mixture of P-cores and E-cores, with P-cores handling latency-sensitive single-threaded and multithreaded work, while the E-cores step in to handle both background and heavily threaded tasks. However, these two types of cores are now augmented by two new low-power-island E-cores located on the SoC tile. These cores are geared for the lowest-power tasks. Intel calls this new three-tier core hierarchy the 3D performance hybrid architecture.

You can’t have a P

The entire (current) lineup of Intel Core Ultra chips are lacking in the usual ‘P’ series of the i3/i5/i7/i9 CPU’s the ‘P’ standing for Performance, which immediately raises some questions. What, for example, has happened to the P-series of chips from previous Core families? The answer: they’ve gone, having been merged with the H-series to cover middle-tier and high-end laptops. The U-series remains for low-power laptops and ultra portables. As the table details, they either consume 9W or 15W as base power, compared to the 28W of the H-series (aside from the 45W Ultra 9 185H).

The entire collection has the same new neural chips, but the graphics offering varies considerably. At the top of the pile sits the Core Ultra’s Arc-branded GPUs, which offer a big step up over Intel’s previous-generation Iris Xe integrated GPUs, but they’re limited to H-series processors. Additionally, Intel claims that the new processors will be extremely efficient, adding substantial battery life thanks to both the node shrink and the use of the NPU. For example, Intel says the 28W Ultra 7 165H uses 25% less power than the Core i7-1370P while streaming Netflix video. The reason? The latter chip uses P-cores and E-cores for that task, while the new Core Ultra only uses the low-power (U) efficiency cores. Confused? Don’t panic, there’s an article in the bog about explaining the P’s and U’s (pun intended).

Intel also makes big claims about its power efficiency versus AMD-based laptops – “up to 79% lower power [at idle in Windows] than AMD at the same 28W envelope for ultrathin notebooks”– but we will hold judgement on that as we’ve not seem or been able to run the tests ourselves.

It also claims “leadership CPU compute for Ultrathin PCs”, specifically sharing results for a Core Ultra 7 165H that suggested it was “up to 11% faster” in multithreaded tasks than an AMD Ryzen 7840U at similar power settings. Intel also claimed big victories, ranging from +19% to +41%, using the Ultra 7 155H over the 7840U and 1360P in UL Procyon, PugetBench for Premiere Pro and PugetBench for Adobe Lightroom. The small print mentioned that an Apple M3 Chip was also tested, but the results curiously aren’t listed on all of Intel’s slides.

Graphic improvements

For graphics, Intel is touting its new Arc GPU with the overarching promise of double the performance per watt – a claim that our early tests confirm, give or take. The Arc branding will only be used by H-series Intel Core Ultra processors with at least 16GB of dual-channel RAM. Otherwise, you get Intel Graphics.

While that’s confusing, and there isn’t a way to write is other than ‘as is’ without 3 pages of annotations, just wait – it gets worse. The new graphics architecture is called Xe-LPG, the low-power gaming alternative to Xe-HPG — which is the GPU found in the desktop Intel Arc Alchemist GPUs. Xe-LPG is not to be confused with Xe LP, which was the graphics solution used in Intel’s 11th, 12th and 13th gen Core processors.

The U-series chips – with mere Intel Graphics – include four Xe-LPG cores with lower peak frequencies than their H-series counterparts. Logic suggests they will be half as fast as the H-series chips that feature eight Xe-LPG cores, but we don’t yet have testing to back this up.

Intel is touting DX12 Ultimate support, Xe SuperSampling and an engine (supporting DP4a) to help with AI acceleration. There’s also support for DisplayPort 2.1 UHBR20 (20Gbits/sec per lane, 80Gbits/sec total).

The AI question

Finally we come to the NPU tile. The Core Ultra 7 165H can deliver up to 34 tera operations per second (TOPS) on the CPU, GPU and NPU combined, but Intel hasn’t broken down the TOPS across each one individually. For comparison, the AMD Ryzen 8040 series can reach 39 TOPS, up from 33 TOPS for the 7040 family. To compare AI performance, Intel matched its Ultra 7 165H against a Core i7-1370P and Ryzen 7 7840U with Ryzen AI, with Intel’s new chip coming top across a series of (Intel-chosen) benchmarks. For example, it claims that it was 1.7x faster than the Ryzen chip in Adobe Premiere Pro’s AI workflows (colour grade, scene edit, export) and 3.2x faster for Stable Diffusion A1111. But AMD could probably pick its own benchmarks with similar figures..

What’s more interesting is the apps Intel has chosen – Wondershare Filmora, Adobe Lightroom, DaVinci Resolve – as this tells us where it feels AI makes a difference. This also means less power consumption in Zoom, with Intel claiming a 38% reduction thanks to NPU offloading. But the real question is how useful the AI features will be, and to whom. Intel says it has partnered with more than 100 software vendors to come up with over 300 “features”, and that its support for OpenVINO should allow for great development support. We shall see.

Coming to a laptop near you

This month we review the first Intel Core Ultra laptop is out in the Acer Swift Go 14, but that’s just the start. It seems straightforward for Intel’s partners to update their existing designs to accommodate the new package, with several examples in our CES feature (see p26) and hundreds more on their way. While it’s too early to say whether the Core Ultra is a sure-fire success in terms of performance, we can safely say that it has the backing of all the big manufacturers.