“Generating High-Speed Images with Stable Diffusion using Meteor Lake’s Built-in NPU: A Demo and Details Unveiled at Intel Press Briefing”

COMPUTEX TAIPEI 2023 will be held at the “Taipei Nangang 1 & 2 Exhibition Hall” in Taipei, Taiwan for four days from May 30th to June 2nd (Taiwan time). At COMPUTEX TAIPEI before the corona wreck, Intel sent speakers to the main keynote speech every year, but this year (2023) Intel will not officially participate in COMPUTEX TAIPEI, and will attend nearby venues. Participation is in the form of private meetings and exhibitions.

In the meantime, Intel has held a press briefing for members of the press to reveal more details about “Meteor Lake,” which the company plans to announce and launch later this year, especially this time around the built-in NPU. (Neural Processing Unit, a processor specialized for AI inference processing) was demonstrated.

Meteor Lake to be launched later this year, hybrid architecture and improved power efficiency

Meteor Lake, which Intel plans to launch later this year (pictured at last year’s Intel Vision)

Meteor Lake, which Intel plans to announce and launch in the second half of this year, will be a next-generation client PC product that will succeed the current 13th generation Core (development codename: Raptor Lake). Intel already outlined Meteor Lake back in August of last year.

As detailed in the article above, the biggest feature of Meteor Lake is that it is made by applying 3D chiplet technology, which Intel calls with the development codename of Foveros. In Meteor Lake, Intel’s Intel 4 (Intel’s process node formerly called 7nm The four tiles, a compute tile manufactured by a notebook that introduces EUV for the first time, a GPU tile manufactured by TSMC’s 5N, and an SoC tile manufactured by TSMC’s 6N, are mounted in a 3D direction. It’s becoming

Intel’s Roadmap for Client PCs

Overview of Meteor Lake

Intel has yet to reveal details of the compute tiles or GPU tiles, but the CPUs were implemented with new power-saving features and a hybrid architecture focused on improving performance (performance hybrid architecture). It only revealed that it has improved power efficiency compared to the 12th generation Core (Alder Lake) and 13th generation Core (Raptor Lake).

Details regarding the GPU tile have not been disclosed, but this time it was explained that it will be based on Arc, which was introduced as a single GPU last year (2022). The built-in GPU of the 12th generation / 13th generation Core continues to be equipped with Intel Xe Graphics (Xe-LP) introduced in the 11th generation Core (Tiger Lake), and there is no GPU update. was Since it will be Arc-based in Meteor Lake, new features introduced in the Arc generation such as XeSS (Xe Super Sampling), AI function (XMX), AV1 decoder / encoder, etc. Built-in GPU will also be supported.

Please refer to the previous article for details on the architecture of Arc.

Meteor Lake will appear as a completely new architecture as a new product for client PCs, and in 2024, Arrow Lake will be introduced, where the compute tile manufacturing process node will be miniaturized to Intel 20A.

In addition, Lunar Lake and its successor are planned as a version with further improved power efficiency, the compute tile is refined to Intel 18A, and this is Arm SoC for Arm version Windows (WoA) with further review on power management It will be a product that realizes power efficiency comparable to that of As a result, it will be possible to realize notebook PCs that are smaller and can be driven by batteries for a long time with x86 CPUs.

Intel’s plan is an ambitious manufacturing technology innovation to rapidly develop and introduce 5 nodes in 4 years, which is scheduled for the “IDM 2.0” strategy promoted by the company’s CEO Pat Gelsinger. It depends on the strategy, and for Intel to introduce Arrow Lake and Lunar Lake on schedule, it depends on whether Intel’s manufacturing technology strategy can be executed as planned.

Meteor Lake makes it possible to generate images at high speed with Stable Diffusion without a single GPU

Hardware that realizes AI in Meteor Lake, it is possible to use not only NPU but also CPU / GPU

This time, Intel has revealed details about the NPU (Neural Processing Unit) that the company will introduce in Meteor Lake as a new detail of such Meteor Lake.

In 2016, Intel acquired Movidius, which developed the VPU, which is now known as the NPU, and has provided USB modules and other products. The feature of Movidius VPU is that it has a built-in memory inside the chip, and it is a mechanism that does not increase power consumption by not accessing external memory, enabling AI inference with extremely low power consumption of 1 W or less. It was a big feature to do.

However, after that, Intel ended external sales of VPU derived from Movidius, and now it was in the form of providing a single VPU called Keem Bay to some PC makers. By using Keem Bay, it is possible to implement Windows AI solutions promoted by Microsoft, such as Windows Studio Effects, which will be provided to some OEM manufacturers in the 13th generation Core generation. was in shape.

However, as Intel itself announced in September last year, Intel has revealed that it has integrated the Movidius-derived VPU into Meteor Lake as an NPU, and since Meteor Lake, the NPU will be standard equipment. Become.

Mr. John Layfield, Intel Vice President and General Manager, Client AI Division

This time, Intel revealed details about the implementation of its Movidius-derived NPU. Intel Vice President and Client AI Business Division Manager John Layfield said, “The NPU installed in Meteor Lake is a product derived from Movidius’ VPU. If the first product is the 1st generation, it is currently provided by Intel. Keem Bay, which is being developed, will be a product based on the 2nd generation architecture.In contrast, Meteor Lake will be equipped with the 3rd generation architecture developed internally by Intel, ”said Meteor Lake He explained that the built-in NPU is an NPU that has evolved further from Keem Bay.

Using an NPU specialized for AI processing increases processing power by 10 times and power efficiency by 5 times

According to Layfield, the NPU installed in Meteor Lake has further improved performance compared to Keem Bay. Specifically, “The processing power is 10 times higher than when processing with a CPU or GPU, but the power efficiency is 5 times higher.” Layfield explained.

This time, Intel demonstrated that it supports multiple functions in Windows Studio Effects, and also demonstrated image generation using Stable Diffusion’s generation AI using Meteor Lake. Although the detailed conditions are unknown, it was confirmed that the image was generated in about 30 seconds without a standalone GPU.

Windows Studio Effects effects (automatic framing, fixation of gaze, background blurring) on ​​a PC equipped with Meteor Lake impose a high load on the CPU and GPU alone, but even if the load on the CPU and GPU is not so high, it can be achieved with the NPU. It has been made public how it can be processed.

When running Windows Studio Effects on a laptop equipped with Meteor Lake, note that the load on the CPU and GPU has not changed much, but the load on the NPU has increased. It’s labeled Movidius VPU, but it was explained that this was an engineering sample (which means it could be branded differently).