The newly developed design targets mobile devices.
Intel kicked off its Innovation 2023 event by unveiling Meteor Lake, which will be the company’s new platform targeting mobile devices. The development consists of four small chips, located on a core slab made on Intel’s 22nm node. The latter ensures the delivery of microchips through Foveros technology.
From an individual chip standpoint, Intel will only produce so-called Compute Tile CPU chips, on Intel’s 4nm process. The other three chips will be manufactured at TSMC. Among them, the graphics board containing the graphics controller is at 5nm, while the SoC chip and the IO tile chip reach the 6nm node.
The physical architecture is quite complex in itself, and the Foveros technology makes production very expensive, but Intel offers surprises even when it comes to individual chips. It would be logical to think that the processor part is entirely contained in the Compute board, but the giant company from Santa Clara has twisted the previously published hybrid design, recently referred to as 3D hybrid, but in fact there is no mention of that third dimension. The point is that 16 non-standardized power cores can work in Meteor Lake. Among these cores, there can be 6 high-performance P-cores codenamed Redwood Cove, 8 low-power E-cores codenamed Crestmont, and a low-consumption E-core (also Crestmont). From now on, you have to rely on not two, but three different performance cores, and since the smaller cores do not support Hyper-Threading, the number of threads that can run will be 32. There are regular P cores and E cores in the computing board, but the cores The low-consumption electronics are actually part of the SoC board, that is, they are completely separate from the main processor part. For this reason, the topic manager responsible for scheduling also changes.
Intel’s strategy here is to check the running task with the thread manager, and if it can realistically run on the SoC’s electronic cores, it tries to keep it there. If this is not possible due to performance limits, the system copies the data to the computing board and the regular electronic cores come to the fore, if they are fast enough. If that’s not good either, the higher performing P cores will still be a last resort. According to Intel, building from the ground up is a better approach because in previous hybrid designs, most tasks started on the P cores, but sometimes to no avail, because the P cores were sufficient. The system thus loses its energy efficiency, but Meteor Lake is built in stages and only touches the P-cores when it is absolutely necessary. In this model, the goal is to achieve maximum efficiency.
Another important design element is the graphics tiles, which are based on the Xe-LPG architecture. The latter copies the technical capabilities of Xe-HPG used in dedicated GPUs, including ray tracing acceleration, only Intel has omitted XMX matrix processors from the system. This is not of much use anyway, as it only improves the operation of the XeSS process, but the latter can also run on vector engines using standard code, and the multiprocessor contains 16 of the latter. This means 128 256-bit vector drivers for the entire IGP.
Other components are on the SoC board, the multimedia block for decoding the AV1 format, the display engine that handles HDMI 2.1 and DisplayPort 2.1, the dual-channel memory and the PCI Express controller, plus a new sub-module as well. It was introduced, which Intel refers to as NPU (Neural Processing Unit). It is designed to accelerate AI tasks and is based on the Movidius design. Finally, the I/O board contains the actual implementation of the necessary interfaces. Another innovation is ISSE (Intel Silicon Security Engine), which can improve security at the platform level.
According to Intel, several versions of each microchip are manufactured, depending on the number of active components the specific processor works with. This will have a role in reducing the cost, because Foveros itself is unfortunately not a cheap technology, that is, in the case of processors that were not originally used for the specific CPU, it is optimal to leave as many components as possible outside the chips.
Meteor Lake processors may arrive in mid-December, but only a few models will likely be in the first round.