Budget GPU for PL

andras.csore · June 16, 2026, 6:37pm

Small update: with PL 9.9 (may due “Optimise AI models” due the trick after 9.9 install) Intel Arc B570 w/ 10GB just shine. AI pre-defined masks works fine (and fast), Export times good, and so on. Results may around nVidia RTX 3060 level.

I update the thread with details (measurements) when i have time for longer tests. May a week from now.

andras.csore · June 21, 2026, 5:39pm

Okay, some results with Intel Arc B570 10GB GPU and PL 9.9.

TLDR: 20Mpix Oly raw: DP3: 5sec, DP3 XD: 15sec.
Masks tax with DP3: non-AI (6) add +4%, Manual Ai masks (4): +6%, Pre-defined AI masks (4): +16%, 6 Non-Ai + 4 Pre-Defined Ai: +25%, 12 Non-Ai + 8 Pre-Dedined Ai: + 28% .
Masks tax with DP3 XD → negliable
Intel Arc B570 10GB, PL9.9, AI masking fluid, rock stable.
Some conclusion in the end.

Test results:

PC config:
Windows 11 latest, CPU: AMD Ryzen 5 3600 (3.6 GHz, 6 core), Ram: 32GB, old (slow) SSD (not nVME) drives. GPU: Intel Arc B570 10GB GPU (driver: 32.0.101.8826 lates, Re-Bar enabled, PCIe 3.0 mode). PL version: 9.9 (latest as now)

Test notes:

Photos has some basic adjustment, like: Exposure, contrast, lens module, Lens sharpness, geometry, smart lighting.
‘1x CP, Brush, CL, Grad, Luma, Hue’ → manual standard mask: Control point, Brush (and NOT Auto-Brush), Control Line, Graduated, Luma, Hue. 1x → only one of them
‘2x CP, Brush, CL, Grad, Luma, Hue’ → 2x → Duplicated each of it.
Manual AI mask: “1x subj, backgrnd, hair, face” → Manual (‘selection’, ‘area’) mask of the subject, background, hair, face. → practically its the same mask as the AI pre-defined Subject, background, hair, face masks.
Manual AI and Pre-deined AI mask values (like Exposure, etc) is the same.
‘2x CP, Brush, CL, Grad, Luma, Hue’ + ‘2x Subject, Background, Hair, Face’ → 10 standard mask + 8 pre-defined AI mask → 18 mask overall.
‘Mpx / sec (median)’ → the median average of how many Mpx processed per second
‘DP3 vs DP3 XD’ → the same test method difference in DP3 vs DP3 XD
‘Mask ratio’ → how the different mask types and amount vs no mask at all.

Test photo

Selected as its easy to detect subject, background, hair, cloth, face.
Olympus 20Mpx RAW.

Testing method notes:

I execute multiple runs for each, from 2-8. Sometimes without any other apps used, sometimes web browser open, etc. Sometimes i quit from PL, sometimes i run different test multiple times in the row
I not do edit while export.
I not use Loupe or DeepPrimerendering.
Test use the same photo (copy for 88 times), and not Virtual Copy.
Test use photo, where AI masks like Subject, Face, Hair is easy to detect. And detect the same. So, in more complex situations times can be more longer.
Test not ideal, as its use the same photo, but as i see, its make no sense (seems no or minimal caching)
PC, GPU not Overclocked.
All version, OS, Drivers, PL is the latest.
‘Maximum perfomance’ mode. OpenCL enabled. Parallel export: 4
Why 88 photo? As one French PL8 test is use 88. I know its different photos. But looks nice at least.

Notes:

i know Intel Arc B570 vs B580 price difference not big, but i go to find ‘most budget GPU what you can buy as new’ and not ‘second hand’. However, i test some second hand AMD, NVIDIA, even far older generations - its runs just fine, and ‘second-hand’ cards price is approx the same (at least in my country)
PCIe 3.0 vs PCIe 4.0 → i know it has a difference in bandwidth. However, as i read, the real world difference (if not High end GPU) is just a few percent.
GeekBench (GPU OpenCL) and GeekBench AI test can show you some realistic value for performance, however for ‘rule of thumb’ the generic benchmark results, like Techpowerup is just okay.
B570 is a relatively budged stuff, seems performance somewhere nVidia RTX 3060 or more.
No crash, rock stable. AI masking ‘fluid’.

Some other conclusion (as i see, in my opinion):

I think, PL can use all (any) GPU. At least i not see any reason to not works with any.
I think, even integrated GPU (iGPU) can be just fine → IF memory amount allocation is possible for 6-8GB (or auto allocation works fine and can do that → if you have enough system memory, probably with more than 16GB its can work fine)
With 6GB more-or-less everyone okay (as DxO write in minimum specs)
With 4GB only Manual AI masks works and may DP3 export also works.
For ‘fluid’ AI masking performance may 35-50% performance of B570 is just enough (as i see, its use (GUI Ai-masking) like 30%-50% or less → so, slower second-hand GPU can be fine, if you less care about Export performance.
Even ‘budget’ new or 1-2 generation behind GPU just okay.
Budget (second-hand) can be okay, what can be the ‘worst’, may export not 5sec, but 10sec. And what… Just okay (for generic usage).
Parallel export performance: 2-6 not different too much, difference is few % only. 2: 4.43 sec, 3: 4.29 sec, 4: 4.25 sec, 5: 4.4 sec, 6: 4.44 sec
Masking (calculation the effect on masks) and Mask flattening (merging the different layers) seems more CPU sensitive.
If you use DP3 XD for export → as the main time is the noise reduction → how many masks, how many AI masks you use → impact only few % (1-5%)
If you use DP3 for export (and NOT XD) → as GPU usage is ‘shorter’ → Maskin impact a bit, but still not so much: Non-AI masks like non-AI add +4%, Manual AI masks: +6%, Pre-defined AI masks: +16%, 6 Non-AI + 4 Pre-Dedined AI +25%, 12 Non-AI + 8 Pre-Defined Ai: + 28% (DP3). With faster CPU this may smaller.
Smart–lighting ‘Spot-weighted’ in ‘Auto’ (face detect) mode seems impact performance.
More photos you export → performance is going better, example: 88 photo vs 380 photo: 4.4 sec → 4sec.
Seems once you use one Manual AI mask, than the other manual AI masks not tax too much. Same for Pre-defined masks.
Interesting: Standard masks (+4%, 5 mask) + Pre-defined masks (+16%, 4 mask)) → if we sum its: 20%. In the measurements its +25%. I think its about ‘matting’ (flattening) mask numbers increment (5+4 → 9).
As it was expected, the most taxing mask type is the AI pre-defined masks.

So,that’s all in the nutshell. Overall im happy. As i see, ‘just good enough’ GPU is just good enough (even better than ‘just enough’).

I not do (yet) Memory usage for different tests. May i do once. ut definitely the AI Pre-defined and DP3 XD is the most taxing.

Misc:
Parallel export GPU usage with 1: (spikes larger than 18%, i not remember, but like 100%)

With 3:

With ‘Standard’ NR vs DP3 saving is negligible, like: 3.5-4 sec vs 5 sec. GPU around 25%

andras.csore · June 23, 2026, 3:18pm

In the meantime i do some ‘budget gpu test’ and realize what you describe → seems RAW decoding performance and/or Noise reduction performance more ‘heavy depend’ on RAW/Camera type. I get for example a few differente RAW: 40Mpx Fuji X-E5, 25Mpx Panasonic GH7, 20Mpx Canon R6 mk1, my 20Mpx Oly EM5Mk3 raws. DP3 export average results like: 8s, 8s, 6s, 5s. I do just ‘light measurements’, but anyhow, Whoppp… How the 40Mpx export done same time as the 25Mpx?

Interesting, may i go some testing on that.
I know in Lr a small (but measurable) performance difference between Canon and Nikon RAW’s (at general), but not so much (as i remember, may 5%?).

randomroyalty · June 23, 2026, 5:57pm

The G9M2 / GH7 files are 16 bit. Fuji are 14.

there is also compression. So if the Fuji camera is set to uncompressed it might be faster but I am speculating (I haven’t tested).

andras.csore · June 24, 2026, 3:19pm

Yep. Its can be a difference.

But its also means → may some of the ‘my system export under 5 sec’ sometimes may ‘not comparable’ with others. And quite interesting the whole thing.

I plan to do some test about:

Memory usage (exact as possible) for each Masking, Exporting, etc type. However, i think PL use “limiter” now with “maximum performance” mode. In the past i do this with 4GB GPU, but obviously cant test with Pre-defined masks, etc.
Decoding an Noise Reduction difference between RAW formats. may i go thru some sample files from DpReview camera tests. Previously i expect its more or less linear (Megapixel / second based on GPU performance), so its more or less predictable (the GPU). But now the Fuji files put a difference (but may Fuji files themself still linear performance vs Mpx).

andras.csore · July 4, 2026, 11:17pm

Some real world export (DP3, 20Mpx Oly):
82 photo - multiple AI masks (like 4-6 Pre-defined + 4-6 Manual): 07:44 → 5.65s / photo
54 photo - multiple manual AI masks (4-6): 04:33 → 5s / photo
Seems overall the AI mask types: Manual vs Pre-Defined don’t tax too much.
Measured values approx same as previously measured.

Next (coming weeks):

GPU VRAM Memory usage for AI masking, Loupe/DeepPrime, DP3 vs DP3 XD.
RAW (Camera body) test. Seems PL faster with Fuji raw’s

andras.csore · July 7, 2026, 9:00pm

Update: RAW (different cameras) Exporting (DP3) Megapixel / second.

Mpx / sec column → that really matter.

As its show the difference between RAW processing percentage
As value may can be interpolate to different GPU’s

Difference vs the “fastest” (in this test the Fuji X-T30 III X-Trans)

Mpx / sec in graph

Observations:

Different RAWs has different performance in Export. Most of it more-or-less the same, but even ‘near same’ models has a difference, example: Canon R6 I vs R6 III → 3.5 vs 4.2 Mpx /sec, so the R6 III is 20% faster.
Fuji X-Trans sensor RAW s is the “fastest”
Mpx / sec not correlated vs total Mpx
Export tests comparison need to take care of that.
Export only performance tests may best if running on mixed camera/Raws.
I wonder on a few point: Canon EOS R6 I: 3.5 R6 III: 4.2. Why? Okay, RAW internal data/structure is different, but still looks a bit strange. Why X-Trans is so ‘fast’? May it has less color, easier to decode (less compressed) and manipulate. But still, why?

Notes:

All measurements based on my system (see previously), GPU: Intel ARC B570 10GB.
All photos has some basic adjustment, geometry, lens sharpness, etc.
NO mask at all
NR: DP3 (and not DP3 XD)
GPU VRAM amount not mater → no masking, DP3 use like few GB only.
Cameras picked by my opinion
Test running on 40 photos (not 39 x VC ). Parallel export: 4
Various test: export with all body with the same (body related) photo (copied 39 times), for some raw with different photos again 40, for some camera 54 totally different photos. All test export has at least one time PL9 re-start.
As i see, its doesn’t really (not) matter the photo itself (how many color in there, etc.). Export time differences negligible: for one photo is 11.8 sec, for another 12sec.
Test photos from DpReview site (Camera tests)
Median value used.
Min/Max values example: Median: 4.43; min: 4.15, max: 4,9 → -6% / +5%, but usually smaller.
After PL start, the 1st export usually a bit slower, like 5-12% (but not all case)

andras.csore · July 7, 2026, 9:17pm

A made some measurements, may interesting for you. See previously.