Intel Meteor Lake Analysis - Core Ultra 7 155H only convinces with GPU performance

We want to start with an overview of the new mobile processor line-up since there are some changes. There are basically two types, the Core Ultra U models with a TDP between 15-57 Watts as well as the Core Ultra H chips with a TDP between 28-115 Watts. This means the U-series stays in the same range as before, while the new H-series now replaces the former Raptor Lake-P models like the Core i7-1360P as well as the Raptor Lake-H45 chips like the Core i7-13700H or the Core i9-13900H.

All new processors offer a total of 10 efficiency cores (more on that shortly), only the number of performance cores differs depending on the model. Th high-end chips Core Ultra 155H, Core Ultra 7 165H as well as the Core Ultra 9 185H have 6 fast performance cores, which is similar to the previous H45 processors like the Core i7-13700H. These three new CPUs offer a total of 16 cores/22 threads. The two new models Core Ultra 5 135J and Core Ultra 5 125H only have 4 performance cores, so the are pretty much the direct successors to Raptor Lake-P chips like the Core i7-1360P. All new H-series models also get the new Arc GPU with 8 Xe cores (Core Ultra 7 & Core Ultra 9) or 7 Xe cores (Core Ultra 5).

The regular U-series chips as well as the H-series models can be equipped with up to 64 GB LPDDR5/x-7467-RAM or up to 96 GB DDR5-5600 RAM. There will be two more U-series chips in 2024 with a lower TDP range between 9-30 Watts and they are also equipped with slower LPDDR5x-6400 RAM. All new Core Ultra mobile processor are also equipped with a dedicated NPU (neural Processor) with two Gen3 Neural Compute Engines.

The major design change of the Meteor Lake processors is the switch from a monolithic design to a chiplet/tile design, where the chip consists of multiple parts. There are certain advantages to this approach, including better scaling capabilities and not all of the tiles have to be manufactured on the same node. Meteor Lake consists of four different tiles: SoC tile as the main component with all the basic functionality, compute tile with P- and E-cores, GPU tile as well as an I/O tile. The compute tile is manufactured in the new Intel4 process, so it is Intel’s first 7 nm chip. All the other tiles are manufactured by TSMC (SoC & I/O tile: TSMC N6, GPU tile: TSMC N5). The chiplet design also results in a slightly higher power consumption though.

In order to get a meaningful comparison between the different processor and graphics adapters, we look at the performance results in synthetic benchmarks, but also the power consumption, which results in the efficiency. All the power measurements were performed with an external screen, so we can eliminate different panel types that would distort the results. Keep in mind that we measure the power consumption of the whole system and do not solely rely on the displayed values for the CPU and GPU consumption. 

We already mentioned that the underlying architecture did not change, but it was further refined according to Intel. We were quite surprised that the single-core performance is actually a bit lower compared to the previous Raptor Lake chips. The new Core Ultra 7 155H is closer to the AMD’s current Zen4 processors instead, but the new Meteor Lake chip is clearly beaten by Apple’s M3 generation and also Qualcomm’s Snapdragon X Elite. Our results are pretty much on par with the reference scores from Intel, so the single-core performance is roughly the same. 

Our multi-core tests show performance gains and the best comparison processor is the Core i7-13700H, which also offers 6 performance cores for a total of 20 threads. The Core i7-1360P on the other hand only has 4 performance cores (16 threads) and we have shown multiple times that the Core i7-1360P in slower than the Core i7-13700H at 40 Watts. The Core Ultra 7 155H has two more efficiency cores (total of 22 threads), so it is not surprising that the multi-core performance is slightly better. However, the performance depends on the power limits and the two units with the Core Ultra 7 155H are slightly faster than the Core i7-13700H in the Schenker Vision 14 (45/40 Watts), but slower than the Core i7-13700H in the Microsoft Surface Laptop Studio 2 (82/50 Watts). The Core Ultra 7 155H can operate in a TDP range between 28-115 Watts, so the actual multi-core performance will vary heavily between different laptops.

Intel claims the GPU performance of the new Arc GPU with 8 Xe cores (128 EUs) has doubled compared to the old Iris Xe Graphics G7 (96 EUs), and we can confirm that claim, at least roughly. This means the new Arc GPU is also faster than AMDs Radeon 780M of the Zen4 CPUs. The two dedicated Intel GPUs Arc A350M and Arc A370M are also beaten, but Apple’s M3 GPUs are more powerful. The Core Ultra 5 chips of the H-series will be equipped with a slower 7-core Arc GPU, which will be closer to the Radeon 780M or even slightly slower.

So far, we do not have any numbers for the new Core Ultra U-models with the new Intel Graphics, but considering it only has four Xe cores with lower maximum clocks, we assume the performance will be down by around 50 %. This means this new GPU called Intel Graphics could be pretty much on the same performance level as the old Iris Xe Graphics G7 (96 EUs).

The big advantage of the synthetic benchmarks is not confirmed in our gaming tests so far, but upcoming driver updates can change the situation noticeably. The best-case scenario so far is that the performance is on par with the Radeon 780M, but it might not be much faster than the Iris Xe Graphics G7 in older titles. We will perform additional gaming benchmarks over the next couple of days and also take a look at Intel’s XeSS upscaling.

We use the game The Witcher 3 for our efficiency test, because the CPU does not have a big impact in this scenario. We can see a big advantage of at least 34 % compared to the old Iris Xe Graphics G7, but the new Arc GPU is not quite on par with the Radeon 780M yet. Apple’s M3 GPUs are still in a completely different league.

We do not see any major improvements in the synthetic benchmarks PCMark 10 as well as the cross-platform benchmark CorssMark. 

Intel often shows the performance at 28 Watts in the marketing material, but the problem is that there will probably not be many Meteor Lake H chips running at such low TDP figures. Even the old Raptor Lake-P models usually ran at much higher power limits and this TDP range is often used by U-series chips, even though they cannot keep up in terms of performance (especially multi-core & GPU).

We still evaluated the multi-core performance of the new Core Ultra 7 155H at fixed power limits of 28 Watts, 35 Watts, and 45 Watts in Cinebench R23. The performance is better compared to Raptor Lake-P (Core i7-1360P) and Raptor Lake-H45 (Core i7-13700H), but the big problem for Intel is AMD’S Ryzen 7 7840U, which offers more performance at all three power levels.

Intel’s marketing material claims significant power savings while idling and during lighter workloads including Netflix streaming, which is supposed to be handled by the two LowPower efficiency cores of the SoC tile alone, so the compute tile can be turned off. We will try to check these claims over the next couple of days and update this article accordingly. However, we might have to wait a bit until final drivers and software versions are available.

A Neural Processing Unit can found inside the Intel Core Ultra 7 155H. This part of the processor has the job of processing AI applications particularly efficiently. That means models such as Stable Diffusion or language models can already run on the device. With the UL Procyon AI Interface benchmark we put the GPU head-to-head with the NPU and simultaneously checked the power consumption. In both scenarios, it quickly became clear that the NPU's performance was only slightly below that of the Arc iGPU when it came to processing neural networks but the energy requirements were nearly halved. This definitely makes the AI accelerator more efficient than using the graphics chip for the same applications.