LightBlog

mardi 3 décembre 2019

Imagination’s new A series GPU architecture is the company’s biggest launch in 15 years

The mobile GPU market these days is dominated by Qualcomm’s custom Adreno GPUs and ARM’s Mali GPUs. However, in the past, the situation was different. There were a few more competitors in the mobile GPU market, which helped to increase competitiveness in the industry. Imagination Technologies, in particular, was one of the prime players of the industry. The UK-based company still remains one of the leaders in the mobile GPU space thanks to vendors licensing its GPU technologies, but its own PowerVR-branded GPUs are found in very few products these days. In the past, its high-end mobile GPUs were used by Apple in the company’s A series SoCs, up until the Apple A10 in 2016. Its low-end and mid-range GPUs, on the other hand, are used by companies such as MediaTek, Rockchip, etc. Now, Imagination Technologies is back with a major GPU architecture launch in the form of the new A series architecture. The company is calling it its most important GPU announcement in 15 years, and it’s promoting the new GPUs as having the world’s fastest GPU IP.

To learn how this impacts the mobile GPU market, let’s delve into the company’s background and see how it fits in the mobile GPU market.

Imagination Technologies – Background

Imagination Technologies has historically been one of the leaders in the mobile GPU market, as it has been operating since the 1990s. The company owns a number of GPU IP patents, which means SoC vendors have to pay it licensing fees. Its tile-based deferred rendering (TBDR) technology is a good example. The PowerVR brand of GPUs competed with Qualcomm’s Adreno GPUs and ARM’s Mali GPUs, but after Texas Instruments closed down in 2012, the company’s only licensee at the high-end was Apple.

Apple and Imagination were linked until 2017, as Imagination even built customized versions of its PowerVR GPUs for iPhones and iPads. Apple confirmed that it was in discussion with Imagination to acquire the company in March 2016, but there was no deal. In April 2017, Imagination announced that Apple had told the company it was transitioning to custom GPUs within 15-18 months, moving away from Imagination’s PowerVR GPUs. A couple of months later, Imagination made a shocking announcement: it was putting itself up for sale, and its market capitalization had been significantly reduced in a short time as its biggest customer moved away from using its technology. Imagination was bought in September 2017 by Canyon Ridge, a Chinese equity firm. To say the least, the developments were surprising.

In the same month, Apple launched the A11 SoC alongside the iPhone 8/iPhone X launch. The A11 was said to contain Apple’s first custom GPU. However, Apple continued to pay licensing fees to Imagination as the A11’s custom GPU still used Imagination’s GPU technology (including TBDR). According to AnandTech, Apple’s custom GPU design is commonly regarded to be a fork of Imagination’s IP in 2015, which Apple continues to develop independently thanks to an architectural license. The relationship between Apple and Imagination these days is not known, and officially, Apple’s custom GPU architecture remains a black box.

Having been acquired by Canyon Ridge, Imagination found itself in a position where it had no customer for its high-end GPUs. The company’s PowerVR Rogue architecture had found its way to the iPhone and the iPad at the high end, but after the A11, there were no other buyers. In the low-end and mid-range segments, Imagination continued to license GPU technology to vendors such as MediaTek, Rockchip, and Unisoc. Its PowerVR 8XT and 9XTP architecture didn’t make it to any shipping products, though. In fact, MediaTek migrated from using PowerVR GPUs to using ARM’s Mali GPUs. For example, the Helio P90 used Imagination’s PowerVR GM9446 GPU (that was based on the older Rogue architecture), but MediaTek then moved to ARM’s Mali-G76MC4 GPU in the Helio G90T.

In the meantime, ARM’s Mali GPUs itself were proving uncompetitive in the mobile GPU market, as they had inferior PPA (performance per area) compared to Qualcomm’s custom Adreno GPUs that custom GPU IP, and more recently, Apple’s custom GPUs. ARM’s Mali-G71 and Mali-G72 suffered from inferior performance and power efficiency compared to the Adreno GPUs of that time. While the Mali-G76 was a solid improvement, it still wasn’t enough to take the mobile GPU performance and efficiency crown in the Android market from Qualcomm.

Qualcomm was in a good place in the Android GPU market as its GPUs enjoyed best-in-class performance, but in the wider SoC market, it found itself being beaten by Apple’s custom GPUs. The custom GPUs in the Apple A11, A12, and A13 SoCs are significantly faster than the Adreno 540, Adreno 630, and Adreno 640 in the Snapdragon 835, Snapdragon 845, and Snapdragon 855 respectively. Suddenly, Apple had a big lead in mobile GPU performance, with Qualcomm relegated to second place. ARM’s Mali GPUs were some way behind in terms of PPA as well as power efficiency.

In May this year, ARM set about correcting the performance and efficiency deficit by announcing the Mali-G77 with the new Valhall GPU architecture with 40% performance improvements. The GPU scored its first two design wins in the form of the Samsung Exynos 990 and MediaTek Dimensity 1000 SoCs.

As of mid-2019, the mobile GPU market reflected the reality that custom GPUs developed by Apple and Qualcomm were better than licensed GPUs from the likes of ARM and Imagination. In terms of competition, this is a bad thing for the industry, as it acts as a barrier to competition if the licensed IP is inferior to custom IP that won’t be licensed to any third-party SoC vendor. Samsung knows this, and the company has already committed to licensing AMD’s GPU IP in future Exynos SoCs (set to launch in a two years’ time).

In the face of this background, the stage was set for Imagination’s most important announcement in 15 years.

A series GPU architecture – AXT, AXM, and AXE GPUs

Imagination has announced its new A-series GPU architecture, with the claim of it having the fastest GPU IP ever created. It’s a big claim. The company’s goal here is to show that the architecture is better than the custom GPUs made by Apple and Qualcomm. In essence, Imagination wants to make licensed GPUs viable again, and provide a better alternative than ARM’s Mali GPUs.

The A series architecture is Imagination’s 10th generation GPU architecture. It consists of three categories of GPUs: the XE, XM, and XT product lines, which signify the low-end, mid-range and high-end segments. The PowerVR branding is not used anymore to describe the architecture, as the company now uses it to describe the TBDR tech. The fastest GPU IP claim applies to the XT product line, as it’s intended as a premium series of mobile GPUs.

The A series GPU architecture is claimed by Imagination to be 2.5x faster than Imagination’s previous generation Series 9 GPU in an ISO area and process node comparison. Imagination isn’t comparing A series with the 8XT or the 9XT lines as there are no publicly available products that use those GPU series. AnandTech notes that the data is likely based on the Helio P90’s GM9446 GPU, which is a Rogue-based 9XM series GPU.

The A series has 4x more ALUs over a 9XM Rogue GPU. It’s said to have 8x faster AI performance. Finally, Imagination says that the A series architecture uses 60% lower power for similar performance.

The new A series architecture is applicable for four configurations in the XT series and one XM configuration. The XE series isn’t actually based on the new architecture, as it’s a continuation of Rogue. The AXT-64-2048 is a flagship GPU with 2.0 TFLOPS, 64 gigapixel/s fill rate, and 8 TOPS of AI performance. The AXT-32-1024, on the other hand, is said to be a premium mobile GPU with 1 TFLOPS, 32 Gigapixel/s fill rate, and 4 TOPS of AI performance. The other two configurations are the AXT-48-1536 and the AXT-16-512, while the AXM series has the AXM-8-256, which is rated for 256 GFLOPS, 8 gigapixel/s fill rate, and 1 TOP of AI performance.

In terms of mobile GPUs, Imagination’s plan is to have the AXT-32-1024 as the sweet spot for a premium mobile GPU. The larger two configurations can be used in tablets, for example.

The new GPU IP contains a lot of changes: 4x more ALUs, 128-wide architecture, and a lot more. AnandTech has covered the new A series technology in detail, so readers should check out the publication’s coverage.

A series architecture – Outlook

The outlook of Imagination’s A series architecture looks great. On the face of it, the AXT-32-1024 GPU will be much faster than even next-generation designs such as Qualcomm’s Adreno GPU in the Snapdragon 865 and even Apple’s custom GPU, to say nothing of the Mali-G77. Imagination also has it covered when it comes to PPA, with the company noting the existing problem where the Mali GPUs have the same performance while taking up much more area, leading to lower PPA. The power efficiency should be great as well.

In terms of design wins, it’s a mixed bag. Qualcomm won’t be licensing this architecture, and the vast majority of the world’s Android phones use Qualcomm SoCs. Samsung has already signed a partnership with AMD to use its Radeon IP in future Exynos SoCs. The two major vendors left are Huawei’s HiSilicon Group and MediaTek, which both currently use ARM’s Mali GPUs. As MediaTek has previously used PowerVR GPUs, it’s a possibility that the company will move back to Imagination from ARM in future SoCs. The company has already announced a flagship SoC for 2020 devices in the form of the Dimensity 1000.

Imagination’s announcement of the new A series architecture was much needed in the mobile GPU market. If it all plays out, the mobile GPU market is now much more competitive in terms of licensed GPUs that can be used by multiple vendors.


Via: AnandTech

The post Imagination’s new A series GPU architecture is the company’s biggest launch in 15 years appeared first on xda-developers.



from xda-developers https://ift.tt/35WvHk3
via IFTTT

Aucun commentaire:

Enregistrer un commentaire