PowerColor Puts NPU Into Radeon GPUs, Nets Additional Power Savings
While most mentions of AI in this day and age has been very superficial (marketing is to blame), sometimes you can find genuinely useful cases where AI brings visible improvements to existing hardware. In this case, PowerColor – one of the leading AMD GPU makers – has found an innovative way of utilizing the NPU, with measurable results.
PowerColor’s AI-Powered Power-Saving Solution
As discovered by @harukaze5719 (via Videocardz), PowerColor has demonstrated a prototype GPU with NPU (neural processing unit) attached, which gets compared to an identical GPU sans NPU. Based on the results shown, the GPU+NPU prototype saves more power compared to the stock AMD Power Saving profile via AMD Software.

Take Cyberpunk 2077 for example – in its stock form, the GPU consumes 263.2 watts; AMD Power Saving knocks that number down to 214.5 watts (a 18.5% reduction), while the onboard NPU has further slashed its power draw to 205.3 watts, a 22% reduction from default settings. Final Fantasy XV showed similar results, with NPU shaving additional power draw over AMD’s power saving feature.
The chip that powers this process is from a US-based chipmaker Kneron, which mainly serves the AI and edge computing markets. According to its press release, TUL (parent company of PowerColor) will be implementing these NPUs into its Hellhound and Red Devil line of GPUs.
The NPU offers two working modes – first, the ECO mode works by “intelligently detect GPU temperature changes and dynamically control fan speed”; another mode, called Boost mode, is designed to “dynamically adjust GPU power consumption and heat dissipation, strictly managing power consumption to ensure high performance and stability.” Kneron says its AI model is responsible for adjusting the GPU’s power output while maintaining its performance.
Based on live testing done by Japanese media ITMedia on PowerColor’s Computex booth, the NPU does bring significant power savings, dropping the power draw from 338W to just 261W. However, there’s a few minor tradeoffs – one, the framerate is slightly lower (118 FPS vs 107 FPS), the GPU is slightly hotter (51°C vs 61°C), though this is due to the fan speeds being significantly reduced (1918 RPM vs 1037 RPM).
This proof of concept does look fairly promising, though some may argue that similar power savings should be doable via software as long as there is a suitable AI model for the GPU to work on. Perhaps having a discrete NPU helps with reducing that software overhead, though it may increase the overall cost of the card as a second chip is needed to complement the main chip.
Pokdepinion: A pretty significant differentiator among the sea of samey AIB GPUs.