AMD turns off your GPU to maintain consumption, how do they do it?

Modern CPU and GPU architectures support numerous features to temporarily increase performance or reduce power consumption; some of these capabilities are advertised, and some are not, as is the case with the so-called Duty Cycle Scaling found in a Linux patch for AMD RDNA2 GPUs, a technology that literally shuts down the GPU cores to reduce power consumption. We tell you how it works below.

In theory, this technology can momentarily turn off the GPU graphics cores and then turn them on in an attempt to reduce power consumption and meet stringent TDP requirements, but how do they do it?

This is how Duty Cycle Scaling works on AMD graphics

Duty Cycle Scaling (DCS) is designed for low-end graphics and low TDP. DCS constantly monitors the current, power, and GPU temperature, and when something exceeds the limit under high workload, the cores turn off for a fraction of a second and then turn on again, once they are turned on again. reach the limits of temperature or power. Power on and off time is determined by firmware based on accumulated power credits.

DCS can shut down a graphics core for less than a millisecond. The patch says there are two types of DCS, asynchronous DCS that can turn off a graphics core based purely on thermal, current, and power readings, and frame-aligned DCS that is designed to turn off a graphics core right after rendering a frame, which which will obviously negatively affect performance. Asynchronous DCS is currently the only one supported, while the other is designed for full-screen VR and 3D workloads, but does not work for now.

In addition to Duty Cycle Scaling technology, it should be remembered that AMD GPUs also support the GFXOFF capability which is used to turn off graphics cores when idle, also to save power, and reduce TDP.

Will only work on low TDP charts

Before we mentioned that this technology is designed for low TDP graphics, and especially those where the TDP has very strict requirements such as laptops or mini PCs. Based on the information of the patches for Linux from AMD which is where it has Once all this information is out, the DCS capacity is currently compatible with the following GPUs:

• Sienna Cichlid (Navi 21, Big Navi).
• Dimgrey Cavefish.
• Navy Flounder (which are probably the RDNA 2's mid and basic range).

Neither high-end nor mid-range GPUs fall into this category of low-TDP GPUs, and therefore Duty Cycle Scaling will not be a major feature of these. However, if AMD or its partners decide to install one of these GPUs in environments with very high thermal restrictions, they can always enable DCS to ensure that the chips are not damaged under high workloads.

To work properly, this technology must be compatible with the GPU, its firmware, driver, and operating system; at present, it is unknown if it is something that really works and is enabled for some graphics subsystems, but it is something that AMD will have to clarify soon.