The equation for designers of mobile graphics processing units (GPUs) is an impossible one to solve. They are trying to bring the power of desktop gaming to a backpack or a laptop of a given size, with a given amount of battery life, and within a given weight. The physical laws which govern the behavior of matter do not permit this to be achieved. There will be compromises of various kinds.
Thermal constraint reality
But again, many mobile systems do not even have the room for a very large cooler, so a Chip Architect has a very difficult job trying to keep a mobile GPU within the thermal design power (TDP). For a desktop GPU, this can be in the hundreds of watts. For a mobile system, a TDP of 30-70 watts is typical. The TDP is the maximum amount of heat that can be dissipated. If the GPU produces more heat than that, then it must be throttled to prevent damage. Of course, the amount of heat that is produced depends on many things including the clock rate, the amount of parallel processing, the number of processing cores, and more. In summary, the Chip Architect of a mobile GPU has to make some tough decisions to keep the GPU cool and perform well.
There is always room for higher frequency and more features, but the problem is that higher frequency for longer periods of time generates more heat, and as a result it reduces the performance due to throttling. The boost is designed to allow short periods of time of maximum performance, but for longer period of time it reduces the frequency in order to keep the temperature under control and maintain stable operation.
The Boost algorithm enables high short-term peak performance, matching the high performance of a desktop GPU for short time frames of activity, while at the same time the chip remains within safe temperatures for long time frames of activity such as long gaming sessions. This makes the chip a great solution for gamers who require high levels of performance, while also maintaining long battery life for non gamers.
Process node advantages and limitations
Process Node size does matter for higher Performance per Watt: for instance moving from 8nm to 5nm (as done for Volcanic X) allows for up to 20% higher Performance per Watt due to the increased number of transistors on the same die size, all running at the same power.
The increase in performance per watt from 28nm to 14nm was around 7.7x greater than the increase from 28nm to 20nm. However since then the increases in performance per watt have lessened. 7nm is only 1.4x more performance per watt than 14nm while 5nm is only 1.2x more performance per watt than 7nm. (As
For instance, the recently introduced NVIDIA RTX 4060 mobile chipset, incorporated into mobile designs, supports ray tracing amongst other functions and also brings desktop-like gaming experiences to gamers while on the move.
Architectural compromises in mobile designs
In many cases the functionality of a Mobile GPU is reduced for one reason or another. The memory bus for example is in many cases reduced to a narrower configuration to save space. The same applies to the number of shader units, the size of the cache etc. to save die space as well as power consumption. The reduced size of the cache will affect the GPU’s performance because it now has to access the main memory even more often.
Additionally, by shrinking the amount of cache on a GPU, it can decrease die size, which also decreases power consumption. However, the more cache a GPU has, the slower it becomes as the amount of memory on the GPU increases. To try to minimize performance in areas where decrease in performance has the least impact on users, such as high-resolution gaming that requires a lot of memory bandwidth to achieve high frame rates, a GPU designer might decrease the amount of memory on the GPU, and decrease the number of shader processors to try to keep frame rate up in compute-intensive applications. Dynamic frequency scaling, in which a GPU operates at highest possible frequency in idle (light load) and lowers frequency as load increases, is also very common to try to keep battery life up as long as possible for light loads of operation, while maintaining highest possible performance for applications that require it.
However, there are also large trade-offs between battery life and gaming performance, where a GPU’s dynamic frequency scaling (DFS) between different workloads is critical. To extend the battery life of a GPU, it will typically reduce frequency and power consumption in lighter workloads in order to achieve the same level of performance, but in a gaming workload, it will increase frequency and power consumption in order to reach higher levels of performance.
Battery life versus gaming performance
Since there is a hard limit to battery capacity (as determined by the physics of energy storage in a given volume or weight), there is also a hard limit to the amount of energy that can be spent by a given piece of hardware.
In order to extend the battery life of a mobile system the power that the GPU consumes can be managed in a number of ways. During idle time the GPU can power down certain sections of the die in order to reduce the amount of power that the GPU consumes. In addition the GPU can also implement variable refresh rates, for example during low intensity gaming or non gaming tasks the GPU can reduce the frame rate in order to save power.
To keep the cost of GPU’s down while increasing performance some chips today have a very low power GPU core to handle normal system tasks like email, web browsing, etc. on the low power GPU core and then use the higher performance GPU core for the high performance gaming applications. In some cases the very low power GPU core is shared with other system chips on the board. In these cases the core can be completely powered down when not in use.
Most modern gaming laptops also come with pre-configured performance modes as well as user-configurable power profiles that allow users to fine-tune the laptop’s power consumption while running on battery versus plugged into the power supply.
Memory bandwidth and power efficiency
However, even for GPUs high bandwidth also means high power consumption and a large surface area on the board. Therefore there are also for the memory interface of a GPU in mobile design many challenges.
When video memory is used on a GPU for rendering images, it is typically in the form of GDDR6. Although it delivers high bandwidth for images being processed by the GPU, its use in mobile designs creates problems due to the amount of power the memory draws, the amount of additional real estate on the PCB that is required for the additional memory, and the cost of that memory. Thus, in order to save on power, reduce cost and minimize the amount of real estate required on the PCB for memory, many designs today use the system memory (shared with the CPU) for rendering. This presents a couple of additional challenges. First, the shared memory between the GPU and CPU creates conflicts for bandwidth between the two components. Second, the GPU must now achieve the necessary bandwidth required by images being processed by the GPU, when the memory is not exclusively dedicated to the GPU.
There are also Power Optimized Memory Systems such as GPU shared memory and video memory which deliver high bandwidth while low power. Many modern GPUs also support Memory Compression which can deliver 20% to 40% effective bandwidth in order to address narrow memory interfaces.
The optimization balancing act
Also data on how people use their mobile phones is plentiful and for the purposes of gaming on mobiles this can be crucial in designing a GPU to be optimised for usage. Rather than designing a GPU to play benchmarks for hours on end, they are designed to play games for short periods of time. They must have the ability to start up quickly and play at a high frame rate. The vast majority of people play for short periods of time, so starting up a game in a couple of seconds is just as important as running at a high frame rate.
Most modern GPUs are tuned and optimized for typical use case patterns. For the gaming market, this generally means short to medium length sessions of gaming in bursts (interspersed with other activities). As a result, the thermal management and performance algorithms in modern mobile GPUs are generally tuned for sustained performance at a lower peak rather than short-lived high peaks followed by a drop off.
New releases of GPU’s are trying to bring the desktop experience to mobile platforms. However battery technology is not evolving at the same pace as GPU’s and as such there is a huge focus in the industry to ensure future releases of mobile graphics processing units are as power efficient as possible.
Ava Mitchell turns celebrity facts into smart, scroll-worthy stories at Star Ledger Pro, backed by 3 years of experience.
