October is clearly significant for imagine technology, an IP company based in the UK. First, it was announced that Simon Beresford Wylie (former chief executive officer of arqiva) would be the CEO of imagination. On Tuesday, imagine released the latest generation of img B Series High-Performance GPU IP. This multi-core architecture GPU IP has four series of cores with 33 configurations.
This will help Imagination further strengthen its position in the mobile GPU IP market while competing in the desktop and cloud GPU market. To what extent will the ray-tracing GPU to be launched next year increase its competitiveness?
For any processor architecture, the pursuit of higher performance will not stop. Kristof beets, senior director of product management at imaging technology, said: "after using the high-performance img a series power VR GPUs released in December last year, customers asked if we could provide higher performance multi-core GPUs. Therefore, on the basis of a large amount of investment and the technical advantages of a series, we added multi-core technology and launched img B series, which expanded the GPU product line of imagine with an amazing 33 new configurations. "
At present, the IMG B series can provide authorization, and each product series has been authorized by the manufacturer.
The first multi-core architecture GPU IP, more suitable for small chip trend
It is reported that the B series can provide 6 tflops (trillions of floating point operations per second) computing power. Compared with the previous generation of img a series products, the power consumption is reduced by 30%, the area is reduced by 25%, and the filling rate is 2.5 times higher than that of the competing IP core.The B-series offers the highest performance density (performance per mm2), while offering a variety of new configurations to achieve lower power consumption and up to 35% bandwidth reduction for a given performance goal, according to imagine.
Specifically, IMG B-series GPU IP provides four series of professional cores: Bxe, BXm, BXT and bxs for specific markets
IMG BXE: for HD display applications, which is specially designed for user interface (UI) rendering and entry-level game design. Each clock cycle can process from 1-16 pixels, and supports resolutions from 720p to 8K. Compared with the previous generation of kernels, the area is reduced by 25%, and the filling density is 2.5 times that of competitors.
Img BXT: for high performance applicationsThe flagship B-series GPU has built-in four cores, which can provide 6 tflops performance, and can process 192 gigapixel (1 billion pixels) per second. The AI computing power reaches 24 tops, and it can provide the highest performance density in the industry.
Img bxs: for automotive applicationsIt meets the ISO 26262 safety standard and provides a complete product series from entry level to advanced level. It can increase the speed of automatic driving graphics processing to more than 60% of the competitive products. It can provide solutions for the next generation of human-machine interface (HMI), UI display, infotainment system, digital cockpit and surround view. The configuration with high computing power can also support automatic driving and auxiliary driving.
The most noteworthy feature of B-series GPU is multi-core architecture.Kristof beets said in his communication, "from past experience, when the high performance of a single core is brought into play, it will move to a multi-core architecture. One of the most important reasons is that it is difficult to improve the performance of the architecture with multiple single cores, and there are a lot of control logic to manage. The multi-core architecture we have developed integrates innovative decentralized management methods, providing efficient scalability features that are in line with the trend of the chiplet architecture industry. "
David Harold, chief marketing officer of imaging, added: "we have observed the trend of small chips in the GPU market, and this decentralized design can not only reduce development costs, but also reduce development time. But the signal transmission of small chips is a problem. "
"In the past, GPU had many secondary cores, which were very dependent on some resources of the main core. Our multi-core architecture design avoids central modules. Small chips have independent resources and can run independently. At the same time, we also avoid some complex logic. Therefore, our design is very in line with the trend of small chips. " Kristof beets further said.
Based on such a multi-core architecture design, imagination optimizes each product family of BXT and BXm kernels, and realizes multi-core extension of GPU kernel by using the expansion characteristics of multiple main cores. Each core can run independently or be combined into a large GPU to perform tasks jointly, which can provide maximum performance for a single application or support different cores to run independent applications.
Bxe series provides the expansion mode of main core and secondary core, which is an area optimization solution. It provides high performance through a single GPU core, and uses hyperlane technology of imagination for multi task processing.
The bxs automobile GPU kernel also takes advantage of the extensibility of multiple main cores to support performance expansion, as well as security checks across multiple cores to ensure correct operation.
It is worth noting that the release of the B series are given 4-core architecture design. Kristof beets explained, "the four core architecture is determined based on our experience and user feedback. The 4-core architecture is the balanced architecture of performance and energy efficiency that can best meet the market demand. In the future, it will continue to improve single core performance and expand multi-core architecture. We have experience in architecture with up to 14 cores, so check that we are not a challenge."
"We need to see that in the server market, the expansion of multi-core is not the most important, but the expansion of cabinet and Ethernet based expansion." Kristof beets also said.
33 configurations, increase mobile market share, attack desktop and cloud market at the same time
Imaging obviously hopes to expand into GPU market with higher performance while maintaining the competitiveness of mobile market through 4 series and 33 configurations of GPU.
Data show that in the GPU IP market, imagination has the highest market share of 35.5%, arm ranks second with 34.5%, and Qualcomm takes the remaining 30% market share. In addition, in the automotive GPU IP market, imagine's market share is as high as 43%.
Mobile GPU IP market share
Therefore, for mobile devices, Bxe, BXm and BXT are available from entry-level to high-end B-series. Among them,The performance and positioning of Bxe and BXm are similar. "There's a very close relationship between the two series," says Kristof beets,Both series of products are very sensitive to bandwidth and cost,Bxe emphasizes more graphics rendering capabilities, and BXm has both strong graphics rendering capabilities and greater floating-point computing capabilities. "
In order to better meet the needs of the cloud market, imagine's GPU can be flexibly combined or split, and it can also provide hardware virtual function in multi instance environment, bringing better flexibility and security.
As for why the bxs series was specifically launched for the automotive market, David Harold said, "it's based on our understanding of the automotive industry and to better meet the unique needs of the automotive market. Many GPU IP on the market have not been optimized separately for the automobile market, which not only can't meet the needs of this market technically, but also can't let customers pass the relevant certification of automobile industry more quickly. "
"Our bxs is an optimization for the needs of the automotive market. It not only meets the needs of the automotive market in terms of performance characteristics, but also conforms to the ISO 26262 safety standard, and can help customers pass product certification faster." David Harold said: "bxs uses risc-v firmware, which is the first in the industry. By adopting risc-v, we also hope to find more opportunities in the automotive market."
The high performance brought by the multi-core architecture, as well as the comprehensive scalable combination, make imagine's B-series can be applied to mobile devices, consumer devices, Internet of things, microcontrollers, digital TV (DTV) and automobile markets. But there is also a key problem, the increase in the number of cores brings more data and the increase of image resolution, so the bandwidth problem can not be ignored.
For this reason, in addition to multi-core architecture, B-series also introduces an important compression technology imgic.Imgic compression technology can provide four levels of compression, complete compression and compression rate of 75% (near perfect quality), 50% (visual lossless), 25% (the most bandwidth saving) lossy compression.
It is reported that for high-end mobile phones with 4K display and 120Hz refresh rate, the read-write speed is close to 8GB / s. Using imgic's lossless mode, the average 4GB / s can be reduced to the average 4GB / s without affecting the quality; if lossy compression is used, the bandwidth can be reduced to 6GB / s, 4GB / s and 2GB / s respectively when the compression rate is 75%, 50% and 25%.
Ray tracing GPU to be launched in the first half of next year
When the A-Series GPU was launched last year, imagine revealed that it would launch ray tracing products in the next few years. At the launch of the B series, Kristof beets further revealed the progress of imaging in ray tracing. "Based on our 10-year research on ray tracing, we believe that it will take level 4 for ray tracing to achieve the best user experience, higher computing power, lower bandwidth and desktop level efficiency," he said
The so-called level 4 is the level division of imagination for ray tracing to help the industry understand and compare the differences of various ray tracing more clearly. There are six levels in total. Level 4 is the BVH processing and consistency sorting function in hardware.
According to Kristof beets, imagination will officially launch level 4 ray tracing technology in the first half of next year.
Imaging's ray tracing technology will be geared to the mobile and desktop markets. At present, NVIDIA has launched a desktop GPU that supports ray tracing. Therefore, it is worth looking forward to how much changes imaging's ray tracing GPU can cause in the desktop market. In addition, the energy consumption of the mobile phone is not well solved due to the energy consumption of the mobile phone.
As far as imagination is concerned, if it wants to enhance its competitiveness in the IP market, it naturally needs to maintain and enhance its market share in the mobile market and automobile GPU IP market, and expand to more markets at the same time. This is why imagination said when it released a series GPUs last year, it said that the next B series and C series GPUs would maintain 30% performance improvement.
Of course, in order to better meet the market demand, imaging also needs collaborative technology such as AI synergy, so that its GPU can not only process images, but also process programmable AI, and also need to cooperate with AI accelerator to achieve better AI performance. It is revealed that imagination will release its fourth generation neural network accelerator (NNA) product next month.
In addition, in the current international situation, the Chinese market is also a good opportunity for imagination. David Harold said imaging has not developed any IP in the United States and has established research and development facilities in Shanghai and Beijing.
GPU, AI and wireless connection are all working on imagination. Can it break the circle successfully?