All posts
By Pravit Gandhi··8 min read

Your grading performance problem is not your Mac

Is the M3 Pro worse than the M1 Pro for color grading? Usually the chip is fine. The real causes are decode, media engine tier and heat.

If your new machine grades worse than your old one, the chip is almost never too weak. Four things produce that experience and none is raw compute: the codec you are feeding it, which media engine your chip tier got, how long the chassis holds its clock, and whether the silicon knows your codec at all. A 2021 laptop plays 31 simultaneous streams of 4K ProRes by Apple's own testing. Whatever is stuttering on your timeline is not a shortage of horsepower.

The assumption behind every one of these threads

Someone buys a premium machine, opens Resolve, and finds playback worse than expected. The replies are almost always about hardware: more GPU, more RAM, a desktop.

The assumption underneath is that grading is a compute problem, so a faster part must fix it. The evidence against it is in who is complaining. One of the most upvoted recent posts says an M3 Pro laptop is worse than an M1 Pro for color grading. Another comes from someone whose two thousand dollar laptop cannot run basic video editing software. A third calls Resolve on a brand new laptop almost completely unusable. These people did not buy too little computer. They bought a computer that is excellent at the wrong thing for the file they are holding.

What actually happens when you scrub a timeline

Camera H.264 and H.265 are long-GOP formats. Most frames in them are not stored as pictures. They are instructions describing how a frame differs from other frames, some earlier in the sequence and some later. To show you one arbitrary frame, the decoder finds the last complete frame, decodes forward through everything in between, and frequently decodes a future frame too.

Apple's ProRes white paper describes the opposite design in one sentence: all ProRes codecs are frame-independent, or intra-frame, codecs, meaning that each frame is encoded and decoded independently of any other frame. Apple adds that this provides the greatest editing performance and flexibility.

Then Apple publishes what that buys. Testing conducted in December 2021 on a preproduction 16-inch MacBook Pro with M1 Max, a 10-core CPU, a 32-core GPU and 64GB of RAM played back 33 simultaneous 4K streams of ProRes 422 Proxy, 31 streams of ProRes 422, 30 streams of ProRes 422 HQ and 21 streams of ProRes 4444 XQ, all at 4096 x 2160 and 23.98 frames per second.

Thirty-one streams, on a machine some people now call outdated. Small camera files are small precisely because the encoder pushed work onto the decoder.

Media engines differ across a generation, not just between them

People compare chip generations and assume the video hardware scales with the name. It scales with the tier.

Apple's current MacBook Pro specifications list the media engine per chip. For the M5 and M5 Pro that means hardware-accelerated H.264, HEVC, ProRes and ProRes RAW, one video decode engine, one video encode engine, one ProRes encode and decode engine, and AV1 decode. For the M5 Max the list is identical except that it carries two video encode engines and two ProRes encode and decode engines.

Apple's 2021 announcement said the same of that generation: M1 Pro included dedicated acceleration for ProRes, and M1 Max went further with two ProRes accelerators and up to twice the video encoding speed of M1 Pro.

A Max is not simply a bigger Pro. On the encode and ProRes path it has twice the hardware. Move from a Max in one generation to a Pro in a later one and you can lose video engine capacity while gaining CPU cores, and general purpose benchmarks will not have shown it.

Bandwidth does not rise with the model number

The other axis that moves against expectation is memory bandwidth, and it matters because the GPU, the CPU and the media engine all draw on one pool.

Apple's own figures do not form a tidy ladder. The M1 Pro was announced in 2021 with up to 200GB/s and the M1 Max with up to 400GB/s. Apple lists the M4 at 120GB/s and the M4 Pro at 273GB/s. It lists the M5 at 153GB/s, the M5 Pro at 307GB/s and the M5 Max at 460GB/s or 614GB/s depending on GPU configuration.

Read the base tiers against that 2021 Pro chip. The M1 Pro's 200GB/s is higher than the M4's 120GB/s and higher than the M5's 153GB/s. A newer, faster, more efficient chip can sit below a four year old one on the axis a heavily nodded grade leans on. The tiers are not comparable across generations by name alone. Compare the tier you are leaving to the tier you are buying on the media engine list and the bandwidth figure, not on the generation number.

A thin chassis cannot hold the clock it starts at

If playback is fine for the first few minutes and degrades after that, you are watching heat rather than capability.

Puget Systems describes the mechanism plainly: most CPUs have a maximum allowable temperature between 95C and 110C, though the maximum is usually 100C, and the automatic reduction in power that prevents additional heat also results in lower clock speeds and, as a consequence, slower computation.

That guide is not about laptops specifically, but the mechanism explains why laptops show it worst. One cooling system serves everything, the volume for moving air is small, and a color grade is a sustained load. Benchmarks measure the first minute. Grading measures the third hour. Consistently bad from the first frame is decode. Fine then worsening is thermal.

Your codec may be newer than your silicon

The last cause gets misread as a defective machine: the file uses a format the hardware has no block for, so decode falls back to general purpose compute.

Apple lists AV1 decode in the media engine for its M4 and M5 chips. Its 2021 M1 Pro and M1 Max announcement describes the media engine and ProRes acceleration and does not mention AV1. Puget Systems notes for DaVinci Resolve that hardware decoding support can change depending on the capability of the hardware in your system, so two files from the same camera can take different paths on bit depth and chroma subsampling alone.

There is a licensing layer on top. Blackmagic's own product page puts multiple GPU support for real time playback of professional 10-bit formats, and accelerated H.264 and H.265 hardware decoding, in DaVinci Resolve Studio. Temporal and AI spatial noise reduction are Studio features too. Someone on the free version with excellent hardware can be watching software decode with nothing on screen to say so.

Ten minutes to find out which one you have

Do these in order and stop when one of them answers.

  1. Transcode one problem clip to ProRes 422 and put it on a fresh timeline. If it plays cleanly, your problem is decode and the rest of this page is a distraction.
  2. Watch the clock. Bad immediately is decode or missing hardware support. Bad only after ten to twenty minutes is thermal.
  3. Look up your chip's media engine line and bandwidth figure on the manufacturer's specification page, then do the same for the machine you are comparing against.
  4. Check your licence, and confirm which acceleration a free version excludes before blaming the hardware.
  5. Look at the node tree. Temporal noise reduction, optical flow retiming and motion estimation need multiple frames per output frame, which multiplies against long-GOP source rather than adding to it.

The case where the machine really is fine

Sometimes all five come back clean, and the honest answer is that the hardware is not the problem and the workflow is.

Grading every clip individually on a long timeline is a workflow decision rather than a hardware requirement, and it is the most common way people turn a manageable job into one their machine cannot sustain. We covered that boundary in do you have to color grade every clip. Rendering from long-GOP originals twice, once for a master and once for delivery, is another, and why your 4K render is taking hours covers the export side. If your log footage fights you on import too, Premiere is applying a LUT to my log footage is the companion piece.

Leumos AI, our browser-based grading studio, sits at the far end of that argument: the decode, the grade and the 4K ProRes or H.265 render happen in the cloud rather than on your chassis, so media engine tiers and thermal ceilings stop being your variable. It grades and renders rather than edits, and it is in closed beta with a waitlist at leumos.ai. The general version of that trade is in grading log footage without DaVinci Resolve.

Frequently asked questions

Is the M3 Pro worse than the M1 Pro for color grading?

For some workloads it can be, and not because the chip is slower in general. Media engine hardware and memory bandwidth are set per tier rather than rising with the generation number, and Apple's figures show base tier chips landing below the 2021 M1 Pro's 200GB/s. Compare the media engine line and bandwidth figure for the exact chips rather than the generation.

Why is DaVinci Resolve so bad at playback for H.264?

Because H.264 is long-GOP, so most frames have to be reconstructed from neighbouring frames before they can be displayed. Apple's ProRes documentation describes the opposite design, where each frame is decoded independently, which is why transcoding a problem clip to ProRes is the fastest way to confirm decode is your bottleneck.

My high-spec laptop cannot run basic video editing software. What is wrong?

Almost always decode, thermals or a missing hardware path rather than insufficient compute. Test in that order: transcode a clip to an intraframe codec, watch whether performance degrades over time, then check your chip's media engine list against the codec you shoot.

Do I need a better GPU to color grade?

Not usually. Long-GOP decode, media engine tier, thermal headroom and codec support all bite before shader performance does, so diagnose those first.

Sources