In 2021, Innosilicon launched the Fantasy One Graphics Card, based on IMG BXT GPU IP. A single Fantasy One card can achieve 4K-level performance for both desktop and cloud gaming with 5 TFLOPS of rendering power, and this performance can be doubled by running two cards together on a server. The card has been perfectly tailored for the expanding Android cloud gaming market, delivering exceptional cost and power efficiency for entertaining mobile experiences via the cloud.

Now, the Innosilicon team has shared their latest accomplishments with the Fantasy One card: forty-four instances supported simultaneously on a single board. Each game runs at about 30fps with a low power draw and with just one card.

This is a significant milestone on their development journey that highlights the capabilities of their product and the benefits to data centre businesses of exploring newer graphics card solutions for their services. In the below interview, Huck He, Principal SoC Architect at Innosilicon, discusses their vision, how they achieved these results, and where they are going next.

What are the latest trends that Innosilicon sees in cloud gaming?

Game content is diversifying and improving in quality. Game graphics, interaction methods, sound experience, and delay improvements—to name just a few—are being constantly upgraded in order to meet users' increasingly high gaming experience requirements.

The second trend is multi-terminal synergy, or enabling a ubiquitous gaming experience no matter what device is used: smart TVs, tablets, PCs, and smartphones are all equally valid options for cloud gaming.

Cloud gaming is gradually integrating other technology fields to promote technological innovation. Many cities now have 5G coverage, enabling gamers to stream more detailed content at faster speeds and with lower latency. The emergence of convincing AI-generated content is offering game developers a new approach to differentiating gameplay. And the roll-out of meta-universe applications will require the same kind of processing and communication infrastructure that the cloud gaming industry is currently proving.

Game developers are exploring ways to create cloud-native content instead of just offering users the ability to stream traditional games from the cloud in a bid to boost revenue per user. This content could take advantage of the greater processing capabilities of the cloud to offer larger-scale games with more complex rendering for high-resolution displays at high frame rates or permeate the game with advanced features like AI-generated content to change gameplay.

What considerations do data centre businesses need to have when selecting the right graphics card for their services?

When picking a graphics card, data centre operators consider the ecosystem of the graphics card vendors and their popularity in a particular segment, for example, cloud gaming. They pay attention to the rendering arithmetic of graphics cards, the achievable accuracy, the training and extrapolation arithmetic of AI, and the target applications and cloud-base capabilities of the cards. On top of that, there is a growing move towards the green data centre, which encourages safe, low-energy, environmentally conscious approaches to meeting the needs of businesses; this in turn can improve cost efficiency by lowering power costs.

How has working with an IP solution helped you create the best possible graphics card for your customers?

We made the decision to base our graphics card on Imagination’s GPU IP in order to accelerate our time to market and ensure that our fundamental graphics architecture is based on proven and high-quality technology. This decision gave our engineering teams the time to add new features to the graphics cards that help our products stand out in the cloud gaming marketplace.

In our GPU research and development process, the GPU design department and IP design department have been working closely together to create a set of high-performance, high-reliability, low-power consumption, and strong compatibility IP solutions for GPU products.

For example, the LPDDR IP used in Fenghua (Fantasy) Two is not only compatible with LPDDR4/4x/5/5x, but we also designed and optimised the internal request scheduling mechanism of DDR for the data access characteristics of GPU cores. This gives GPU cores more bandwidth for memory access and lower latency when processing critical tasks. The DDR controller’s internal request scheduling mechanism is likewise designed to provide the GPU core with greater memory access bandwidth and lower latency for critical tasks.

Because CoreActive has a robust library of mature IPs for different process nodes of various foundries, we are able to choose the process node more flexibly and are not constrained by IP selection when designing GPU chips.

In which areas did you choose to focus your engineering effort to create silicon that would be ideal for Android cloud gaming?

Focus points in adapting silicon for Android cloud gaming include support for virtualisation, support for new versions of OpenGL® ES or Vulkan® API, optimisation of linearity for multi-client scenarios, and reliability.

In hardware development, we concentrated more architectural, algorithmic, and design resources on researching and implementing hardware modifications for virtualisation and API upgrades. Secondly, we invested in the software driver development department to create drivers for higher versions of Android and optimise their stability and performance. Lastly, we concentrated some of our testing efforts on new features to ensure reliability and performance in mainstream gaming scenarios and to ensure that the hardware can communicate and operate well with Android devices.

What were the challenges in scaling the number of gamers that Fantasy One could host? How did you overcome them?

Expanding the number of gamers that can be accommodated on a graphics card is an important goal for our commercial customers, but one that comes with some difficulties. One of the major challenges is the linearity of the software scheduling. As the number of clients increases, the number of requests sent to the GPU and the data accompanying them increases exponentially. Our driver engineers put a lot of effort into optimising the data handling scheduling: maximising the use of the on-chip data handling engine while optimising the data throughput efficiency of the graphics controller, narrowing down the gaps and wastage in the task scheduling, which also reduced the system CPU load and wait time.

The isolation and protection of client data is another problem that needs to be solved. In order to ensure that the data of multiple clients does not interfere with each other, we designed a highly efficient and flexible configuration of the memory mapping mechanism to manage the memory space of different clients so as to achieve physical isolation of the client data without reducing access efficiency. Our solution is an evolution of Imagination’s HyperLane virtualisation technology that creates up to eight individual control lanes per core, each securely isolated in memory, with flexible performance management techniques.

What features of IMG BXT have added the most value to Fantasy One?

Imagination’s tile-based deferred rendering architecture brings significant power efficiency benefits to the FengHua (Fantasy) Series GPUs, enabling us to support more ways of cloud gaming than competitors at the same power consumption level.

What has been the response from customers and the market on Fantasy One?

Fantasy One has received rapid market acclaim upon its launch. After testing and evaluation, our customer found that Fantasy One has outstanding advantages in terms of energy efficiency ratio, number of customers supported by a single card, and price/performance ratio when compared against competing solutions in typical cloud gaming scenarios.

What is next for Innosilicon?

Fantasy One and Fantasy Two are just the beginning for Innosilicon, and the results have been very good since their launch. Our next step will be to continue to cultivate two product lines: one for the low-power market and one for the server market.

In the low-power market, in addition to continuing to improve performance, we will actively explore application scenarios in different industries and launch products that focus more on energy efficiency, low cost, and cost-effectiveness.

In the server market, we will continue to introduce products with more advanced technology and stronger performance. For example, we will adopt chiplet technology, which can break through the limitations of process bottlenecks and better improve performance.