Arm juices mobile GPUs with neural tech for better graphics
(2025/08/12)
- Reference: 1755005408
- News link: https://www.theregister.co.uk/2025/08/12/arm_bringing_neural_acceleration_to/
- Source link:
Chip designer Arm is bringing dedicated neural accelerator hardware to its GPU blueprints used in phones. It expects this to deliver higher quality visuals while boosting AI performance.
[1]
The UK-based tech biz says that its mobile graphics processor designs, scheduled to be launched in 2026, will feature neural tech that can cut GPU workloads in half and allow for other AI capabilities in future. Arm also announced a developer kit so programmers can start getting to grips with it today, plus ML Extensions for the Vulkan Graphics API.
Arm initially sees neural acceleration being used for upscaling graphics to a higher resolution, without affecting performance. Other envisioned uses are doubling the frame rate using interpolation, and improving image quality by enabling real-time path tracing on mobile devices using fewer rays per pixel.
"As AI increasingly merges with real time graphics, we need GPU-based AI that's tightly integrated, performant and most importantly, power efficient. And making AI on GPUs easy for developers is what's driven the technical innovations that we're talking about here," said Geraint North, Arm Fellow in AI and Developer Platforms.
Arm declined to share any detailed technical information on the neural accelerators before the Mali GPUs that will include them are announced.
[2]
However, it will fit inside the shader cores in each GPU, thus the neural performance should scale in line with the number of shader cores in a particular GPU implementation. Arm's [3]5th Gen designs could scale from five cores or fewer up to a maximum of 16.
[4]
[5]
Last year, the chip designer announced its current upscaling technology, Arm Accuracy Super Resolution ( [6]Arm ASR ). This allows a game to render a lower resolution image and apply an algorithm to upscale, reducing the processing cost of a frame while preserving quality.
To improve on this, Arm is punting Neural Super Sampling (NSS) powered by its accelerator hardware, which enables upscaling from 540p resolution to 1080p in 4ms per frame, saving up to 50 percent of the GPU workload compared with rendering the full frame, the firm claims.
[7]
"It's real time AI driven rendering. It's faster, it's sharper, it's more power efficient. So NSS can produce the equivalent quality output with lower quality inputs, or it can produce better quality output with the same inputs," North said.
Arm is also working on other applications, as mentioned above, including Neural Frame Rate Upscaling (NFRU) and Neural Super Sampling and Denoising (NSSD). The first is designed to improve frame rates by taking a pair of consecutive frames and generating an intermediate frame.
"The neural network is also tightly coupled with new hardware that we'll be adding to our GPUs to accelerate the generation of motion vectors, which track how the pixels are moving between frames. And this will allow content running at 30 FPS to be really cheaply upscaled to 60 FPS," North added.
[8]
NSSD is intended to produce image quality expected from full path ray tracing, which is far too computationally expensive even on desktop systems, according to North, "but when you couple path tracing with a neural network, you can actually cast only a handful of rays per pixel into the scene, and you can use neural technology to add back in those missing details. So the neural network is able to extrapolate data, not only from adjacent pixels, but also from previous frames."
Enabling all this is a neural graphics development kit, which Arm says it is making available today so game developers can get a head start on integrating AI-powered graphics before the hardware is available.
The kit comprises plugins for Unreal Engine that allow devs to incorporate neural super sampling into a game "with just a few clicks," with the models available in open formats via GitHub and Hugging Face. There is also a full PC emulation of the Arm ML Vulkan extensions, allowing developers to run the entire stack ahead of the mobile hardware being ready.
[9]Arm plots move up the stack with push into end-to-end silicon
[10]AI don't know: Enterprises slow to pick up on Copilot+ PCs
[11]Alphawave Semi swallowed in Qualcomm's $2.4B connectivity conquest
[12]Trump seeing green as he weighs deal to allow Nvidia Blackwell GPU sales to China
This isn't the first time neural tech has been infused into phone chips. It is already widely used to power camera features, and Arm licensee Qualcomm has been pushing the AI capabilities of its smartphone platforms thanks to neural processing units (NPUs) it embedded into its processors.
At MWC last year, Qualy showed off a [13]7 billion parameter large language model running on an Android phone , and unveiled its AI Hub for developers.
When asked if its latest tech could be used for other purposes, North said that programming model (Vulkan) is different and is graphics-first, but not exclusively graphic-focused and could theoretically be used for some inferencing.
"We're super excited to see what people do with it," he said.
The question is, will others share that enthusiasm? ®
Get our [14]Tech Resources
[1] https://regmedia.co.uk/2025/08/12/arm_neural2.jpg
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aJtlFNEybkErEIMKXX6anwAAAQY&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://www.theregister.com/2023/05/29/arm_tcs23/
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aJtlFNEybkErEIMKXX6anwAAAQY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aJtlFNEybkErEIMKXX6anwAAAQY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://community.arm.com/arm-community-blogs/b/mobile-graphics-and-gaming-blog/posts/introducing-arm-accuracy-super-resolution
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aJtlFNEybkErEIMKXX6anwAAAQY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aJtlFNEybkErEIMKXX6anwAAAQY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[9] https://www.theregister.com/2025/07/31/arm_q1_2026/
[10] https://www.theregister.com/2025/07/28/copilot_pc_sales_grow_slowly/
[11] https://www.theregister.com/2025/06/09/qualcomm_alphawave_acquisition/
[12] https://www.theregister.com/2025/08/11/trump_seeing_green_as_he/
[13] https://www.theregister.com/2024/02/26/qualcomm_puts_genai_into_smartphones/
[14] https://whitepapers.theregister.com/
[1]
The UK-based tech biz says that its mobile graphics processor designs, scheduled to be launched in 2026, will feature neural tech that can cut GPU workloads in half and allow for other AI capabilities in future. Arm also announced a developer kit so programmers can start getting to grips with it today, plus ML Extensions for the Vulkan Graphics API.
Arm initially sees neural acceleration being used for upscaling graphics to a higher resolution, without affecting performance. Other envisioned uses are doubling the frame rate using interpolation, and improving image quality by enabling real-time path tracing on mobile devices using fewer rays per pixel.
"As AI increasingly merges with real time graphics, we need GPU-based AI that's tightly integrated, performant and most importantly, power efficient. And making AI on GPUs easy for developers is what's driven the technical innovations that we're talking about here," said Geraint North, Arm Fellow in AI and Developer Platforms.
Arm declined to share any detailed technical information on the neural accelerators before the Mali GPUs that will include them are announced.
[2]
However, it will fit inside the shader cores in each GPU, thus the neural performance should scale in line with the number of shader cores in a particular GPU implementation. Arm's [3]5th Gen designs could scale from five cores or fewer up to a maximum of 16.
[4]
[5]
Last year, the chip designer announced its current upscaling technology, Arm Accuracy Super Resolution ( [6]Arm ASR ). This allows a game to render a lower resolution image and apply an algorithm to upscale, reducing the processing cost of a frame while preserving quality.
To improve on this, Arm is punting Neural Super Sampling (NSS) powered by its accelerator hardware, which enables upscaling from 540p resolution to 1080p in 4ms per frame, saving up to 50 percent of the GPU workload compared with rendering the full frame, the firm claims.
[7]
"It's real time AI driven rendering. It's faster, it's sharper, it's more power efficient. So NSS can produce the equivalent quality output with lower quality inputs, or it can produce better quality output with the same inputs," North said.
Arm is also working on other applications, as mentioned above, including Neural Frame Rate Upscaling (NFRU) and Neural Super Sampling and Denoising (NSSD). The first is designed to improve frame rates by taking a pair of consecutive frames and generating an intermediate frame.
"The neural network is also tightly coupled with new hardware that we'll be adding to our GPUs to accelerate the generation of motion vectors, which track how the pixels are moving between frames. And this will allow content running at 30 FPS to be really cheaply upscaled to 60 FPS," North added.
[8]
NSSD is intended to produce image quality expected from full path ray tracing, which is far too computationally expensive even on desktop systems, according to North, "but when you couple path tracing with a neural network, you can actually cast only a handful of rays per pixel into the scene, and you can use neural technology to add back in those missing details. So the neural network is able to extrapolate data, not only from adjacent pixels, but also from previous frames."
Enabling all this is a neural graphics development kit, which Arm says it is making available today so game developers can get a head start on integrating AI-powered graphics before the hardware is available.
The kit comprises plugins for Unreal Engine that allow devs to incorporate neural super sampling into a game "with just a few clicks," with the models available in open formats via GitHub and Hugging Face. There is also a full PC emulation of the Arm ML Vulkan extensions, allowing developers to run the entire stack ahead of the mobile hardware being ready.
[9]Arm plots move up the stack with push into end-to-end silicon
[10]AI don't know: Enterprises slow to pick up on Copilot+ PCs
[11]Alphawave Semi swallowed in Qualcomm's $2.4B connectivity conquest
[12]Trump seeing green as he weighs deal to allow Nvidia Blackwell GPU sales to China
This isn't the first time neural tech has been infused into phone chips. It is already widely used to power camera features, and Arm licensee Qualcomm has been pushing the AI capabilities of its smartphone platforms thanks to neural processing units (NPUs) it embedded into its processors.
At MWC last year, Qualy showed off a [13]7 billion parameter large language model running on an Android phone , and unveiled its AI Hub for developers.
When asked if its latest tech could be used for other purposes, North said that programming model (Vulkan) is different and is graphics-first, but not exclusively graphic-focused and could theoretically be used for some inferencing.
"We're super excited to see what people do with it," he said.
The question is, will others share that enthusiasm? ®
Get our [14]Tech Resources
[1] https://regmedia.co.uk/2025/08/12/arm_neural2.jpg
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aJtlFNEybkErEIMKXX6anwAAAQY&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://www.theregister.com/2023/05/29/arm_tcs23/
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aJtlFNEybkErEIMKXX6anwAAAQY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aJtlFNEybkErEIMKXX6anwAAAQY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://community.arm.com/arm-community-blogs/b/mobile-graphics-and-gaming-blog/posts/introducing-arm-accuracy-super-resolution
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aJtlFNEybkErEIMKXX6anwAAAQY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aJtlFNEybkErEIMKXX6anwAAAQY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[9] https://www.theregister.com/2025/07/31/arm_q1_2026/
[10] https://www.theregister.com/2025/07/28/copilot_pc_sales_grow_slowly/
[11] https://www.theregister.com/2025/06/09/qualcomm_alphawave_acquisition/
[12] https://www.theregister.com/2025/08/11/trump_seeing_green_as_he/
[13] https://www.theregister.com/2024/02/26/qualcomm_puts_genai_into_smartphones/
[14] https://whitepapers.theregister.com/
More Bollocks
NewModelArmy
Article states :
"Chip designer Arm is bringing dedicated neural accelerator hardware to its GPU blueprints used in phones. It expects this to deliver higher quality visuals while boosting AI performance."
Boosting AI for people is akin to giving laxatives to someone with dysentery.
People don't want dysentery, and they certainly don't want it boosted.
Fake resolution.
Fake frames.
Fake fun.