Samsung Patents a Chip That Loads the Next Filter While the Current Image Renders
Every time your phone processes a photo or video frame, the chip has to fetch the math instructions it needs before it can do any work. Samsung's new patent proposes keeping the next set of instructions already loaded and waiting, so there's no gap between frames.
How Samsung's two-filter memory setup speeds up image processing
When a chip processes video, it applies a mathematical recipe called a filter to each frame, one after another. The problem is that fetching the next recipe from memory takes time, which creates tiny stalls between frames.
Samsung's patent describes a design that keeps two filter memories running at once. While the chip is busy processing the current frame using the first memory, the second memory is already holding the filter for the next frame. When the chip finishes, it doesn't have to wait. The next filter is right there.
Think of it like a relay race where the next runner is already at the baton-handoff point instead of jogging over from the sideline. The result is a tighter, more continuous pipeline for processing images or video frames on a device.
… a first filter memory storing a first convolution filter used in the convolution operation in a current frame; a second filter memory storing a second convolution filter used in the convolution operation in a next frame …
Translation: The device uses two separate memory banks to hold the current image filter and the next one simultaneously.
How the dual filter memories pass work frame to frame
The patent describes an electronic device built around two dedicated filter memories: one holding the convolution filter for the current frame, one pre-loaded with the filter for the next frame.
A convolution filter is a small grid of numbers applied across an image to achieve effects like sharpening, blurring, or detecting edges. It's the core mathematical operation behind most on-device image processing and many neural network inference tasks. Fetching that filter from main memory each frame adds overhead that compounds across thousands of frames per second.
The device works like this:
- The processor reads the current-frame filter from the first memory and runs the convolution on the input image.
- While that work is happening, the second memory already holds the next frame's filter.
- When the current frame finishes, the processor immediately picks up the next filter without a separate fetch cycle.
The claim is deliberately hardware-level: two physically separate filter memories, not a software scheduling trick. The patent doesn't specify a particular image type or application domain, so the design could apply to anything from camera pipelines to on-device AI inference that leans on convolution layers.
… obtain the first convolution filter from the first filter memory in the current frame, perform a convolution operation by using the first convolution filter with respect to an input image, obtain a second convolution filter from the second filter memory in the next frame, and perform the convolution operation by using the second convolution filter …
Translation: The processor switches between two different filters in sequence to process the image frames one after another.
What this means for Samsung's on-device image processing chips
On-device image processing is one of the most compute-intensive things a phone chip does, and even microseconds of stall time matter when you're handling high-frame-rate video or real-time camera effects. A hardware solution that pre-loads the next operation rather than waiting for a software request is a straightforward way to close that gap without asking the processor to work faster.
Samsung sits in an unusual position: it designs its own Exynos chips and also builds image sensors and camera systems, so a tighter convolution pipeline has potential value across multiple product lines at once. The latest Big Tech patents in image-processing silicon show a consistent push toward reducing the memory-access bottleneck, and this filing fits squarely in that current.
That makes this Samsung's third filing we've tracked in the GPU rendering race since July, adding to their two-model 3D lighting patent and light-per-cube coloring patent.
Memory delays between frames are a real and recurring cost in any device that processes fast-moving images, from phone cameras to video filters. Every time a new frame arrives, the processor has to fetch the instructions for how to analyze it, and that fetch takes time. At high frame rates, those small delays stack up and can make the difference between a fluid experience and a stuttering one.
Keeping a second set of instructions pre-loaded and ready addresses that fetch delay directly. It is a focused fix, not a sweeping redesign, and its value lives or dies inside the tight timing constraints of camera hardware.
What remains open is how much of the total slowdown this actually solves. Moving the instructions is a small part of the work; moving the image data itself is far larger, and that is untouched here. Matching the scale of the fix to the scale of the problem is the right question to ask.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
16 drawing sheets from US 2026/0253173 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →