The Detailed Explanation of GPU Turbo

Under the hood, Huawei uses TensorFlow neural network models that are pre-trained by the company on a title-by-title basis. By examining the title in detail, over many thousands of hours (real or simulated), the neural network can build its own internal model of how the game runs and its power/performance requirements. The end result can be put into one dense sentence:

Optimized Per-Device Per-Game DVFS Control using Neural Networks

In the training phase, the network analyzes and adjusts the SoC’s DVFS parameters in order to achieve the best possible performance while minimizing power consumption. This entails trying its best to hit the nearest DVFS states on the CPUs, GPU, and memory controllers that still allow for hitting 60fps, yet without going to any higher state than is necessary (in other words, minimizing performance headroom). The end result is that for every unit of work that the CPU/GPU/DRAM has to do or manage, the corresponding hardware block has the perfectly optimized amount of power needed. This has a knock-on effect for both performance and power consumption, but mostly in the latter.

The resulting model is then included in the firmware for devices that support GPU Turbo. Each title has a specific network model for each smartphone, as the workload varies with the title and the resources available vary with the phone model. As far as we understand the technology, on the device itself there appears to be an interception layer between the application and GPU driver which monitors render calls. These serve as inputs to the neural network model.  Because the network model was trained to output the DVFS settings that would be most optimal for a given scene, the GPU Turbo mechanism can apply this immediately to the hardware and adjust the DVFS accordingly.

For SoCs that have them, the inferencing (execution) of the network model is accelerated by the SoC’s own NPU. Where GPU Turbo is introduced in SoCs that don’t sport an NPU, a CPU software fall-back is used. This allows for extremely fast prediction. One thing that I do have to wonder is just how much rendering latency this induces, however it can’t be that much and Huawei says they focus a lot on this area of the implementation. Huawei confirmed that these models are all 16-bit floating point (FP16), which means that for future devices like the Kirin 980, further optimization might occur through using INT8 models based on the new NPU support.

Essentially, because GPU Turbo is in effect a DVFS mechanism that works in conjunction with the rendering pipeline and with a much finer granularity, it’s able to predict the hardware requirements for the coming frame and adjust accordingly. This is how GPU Turbo in particular is able to make claims of much reduced performance jitter versus more conventional "reactive" DVFS drivers, which just monitor GPU utilization rate via hardware counters and adapt after-the-fact.

Thoughts After A More Detailed Explanation

What Huawei has done here is certainly an interesting approach with the clear potential for real-world benefits. We can see how distributing resources optimally across available hardware within a limited power budget will help the performance, the efficiency, and the power consumption, all of which is already a careful balancing act in smartphones. So the detailed explanation makes a lot of technical sense, and we have no issues with this at all. It’s a very impressive feat that could have ramifications in a much wider technology space, eventually including PCs.

The downside to the technology is the per-device & per-game nature of it. Huawei did not go into detail about long it took to train a single game: the first version of GPU Turbo supports PUBG and a Chinese game called Mobile Legends: Bang Bang. The second version, coming with the Mate 20, includes NBA 2K18, Rules of Survival, Arena of Valor, and Vainglory.

Technically the granularity is per-SoC rather than per-device, although different devices will have different limits in thermal performance or memory performance. But it is obvious that while Huawei is very proud of the technology, it is a slow per-game roll out. There is no silver bullet here – while an ideal goal would be a single optimized network to deal with every game in the market, we have to rely on default mechanisms to get the job done.

Huawei is going after its core gaming market first with GPU Turbo, which means plenty of Battle Royale and MOBA action, like PUBG and Arena of Valor, as well as tie-ins with companies like EA/Tencent for NBA 2K18. I suspect on the back of this realization, some companies will want to get in contact with Huawei to add their title to the list of games to be optimized. Our only request is that you also include tools so we can benchmark the game and output frame-time data, please!

On the next page, we go into our analysis on GPU Turbo with devices on hand. We also come across an issue with how Arm’s Mali GPU (used in Huawei Kirin SoCs) renders games differently to Huawei’s competitor devices.

The Claimed Benefits of GPU Turbo: Huawei’s Figures The Difficulty in Analyzing GPU Turbo
POST A COMMENT

64 Comments

View All Comments

  • jjj - Tuesday, September 04, 2018 - link

    Good first step but could be expanded well beyond the GPU and the entire system built around it.
    Been thinking about this for some years and could lead to very different hardware if you have an NPU manage everything.
    Reply
  • ZolaIII - Tuesday, September 04, 2018 - link

    NNPU didn't even earn to eat hire. No one will do it better than your self. Switching off two big cores does much more than all this fuss about GPU turbo. Reply
  • mode_13h - Tuesday, September 04, 2018 - link

    Not to agree or disagree with your point, but I think you mean "here" instead of "hire". Reply
  • mode_13h - Tuesday, September 04, 2018 - link

    You mean like this?

    https://www.phoronix.com/scan.php?page=news_item&a...
    Reply
  • mode_13h - Tuesday, September 04, 2018 - link

    Anyway, I'm thinking a deep learning model could probably do a better job at managing core clock speeds and perhaps even deciding whether to schedule certain tasks on big vs. little cores. Reply
  • ZolaIII - Wednesday, September 05, 2018 - link

    We (not AI) need to evolve scheduler logic and add SMP affinity flags to processes. Only then game's can begin. Here goes the rain again. By the way thanks for that. Reply
  • mr_tawan - Tuesday, September 04, 2018 - link

    Does it raytrace? Reply
  • Manch - Wednesday, September 05, 2018 - link

    I think the question that should be asked is:

    Can it raytrace Crysis?
    Reply
  • Lord of the Bored - Wednesday, September 05, 2018 - link

    I'd settle for raytracing FEAR. Reply
  • sing_electric - Tuesday, September 04, 2018 - link

    This presents a quandary for benchmarking/reviewing: In the past, if a company released drivers that changed the performance of the device while running a specific app, we'd all call it "cheating" if that app was a benchmark. However, by extending the "cheats" to other apps, the user sees real benefits, even though it's the same behavior.

    It also means that performance numbers have to be taken with even more salt, because the performance on a popular app which has been "Turboed" by Huawei might not be indicative of the performance you see if you only say, play less popular games that Huawei hasn't profiled.
    Reply

Log in

Don't have an account? Sign up now