# Generative Remove runs on the CPU instead of the GPU since version 27.9.1

**Issue summary**
Since updating to Photoshop 27.9.1, the Remove tool in Generative mode loads about 6 GB of data into the graphics card memory, then performs the actual work on the processor instead of the graphics card. The graphics card stays at idle power for the whole operation. A removal that used to complete quickly now takes several minutes.

**Product and version**
Photoshop desktop version 27.9.1 (build 20260803.r.1 65900e1)

**Last version where it worked correctly**
Photoshop 27.8.0 (build 20260610.r.13 4081857). The problem started immediately after updating on 6 August 2026.

**Operating system**
Windows 11 Pro, version 26200

**Hardware**
Graphics card: NVIDIA RTX PRO 4000 Blackwell, 24 GB, driver 32.0.15.9672
Processor: Intel i9-13900K (32 logical cores)
Memory: 128 GB

**Steps to reproduce**
1. Open any photo.
2. Select the Remove tool.
3. In the options bar, make sure the generative mode is turned on.
4. Paint over a small area of the image.
5. Watch the Task Manager Performance tab, or any graphics card monitoring tool, while the operation runs.

**Expected result**
The removal is processed on the graphics card and completes quickly, as it did in version 27.8.0.

**Actual result**
The removal is processed on the processor. It takes several minutes for a small area. During the operation:

- The processor runs at 74 percent of 32 cores, peaking at 97 percent.
- The graphics card draws 12.8 watts, which is its exact idle level. It draws over 100 watts when it is genuinely working.
- Photoshop does not appear in the per-application graphics usage counters at all.
- Photoshop uses up to 50 GB of system memory.

The graphics card memory is still filled at the start: selecting the tool raises it from 5.6 GB to 11.7 GB. That memory is then released partway through the operation, without ever having been used for the work.

**Measurements**

Recorded once per second using nvidia-smi for graphics card power, utilisation and memory, and the Windows performance counters for per-application graphics memory and usage. Test conditions: a single small selection, all other programs using the graphics card shut down, and 18.9 GB of graphics memory free before starting.

| Phase | Graphics memory | Graphics card power | Photoshop processor use |
| :-- | --: | --: | --: |
| Idle | 5.6 GB | 13.4 W | 0 % |
| Tool selected | 5.6 to 11.7 GB | peak 54.8 W | 0 % |
| Operation running | 11.7 GB held | 12.8 W | 74 %, peak 97 % |
| Partway through | released to 7.3 GB | 12.8 W | still 74 % |

Out of 169 measurements, only 3 exceeded 40 watts, and all 3 occurred while the data was being loaded, none while the work was being done.

**Workaround**

Turning off the generative mode and using the classic Remove is instantaneous. Only the generative path is affected. This may help narrow down where the problem is.

**Things I checked first, so they can be ruled out**

- Not a shortage of graphics memory. 18.9 GB was free during the test, and the shared system memory figure stayed flat at its normal level, so nothing overflowed either.
- Not a disabled graphics card in preferences. Use Graphics Processor and OpenCL are both switched on, and the sniffer crash counter is zero.
- Not a detection problem. The sniffer log ends normally and reports the RTX PRO 4000 as the preferred device, with an OpenCL compute score of 19,138 against 339 for the built-in Intel graphics.
- Not caused by other software. All other programs using the graphics card were shut down for the final test.
- Not related to selection size or to processing several areas at once. The same result occurs on a single small selection and on a batch of twenty.

**Impact**

I bought this graphics card one month ago, for 2,000 euros, specifically so that the local generative Remove model would run on my own machine. It worked exactly as intended until this update. Since 27.9.1 the card is no longer used for this work, and the tool is too slow to use in my retouching workflow.

Loading 6 GB into the graphics card and then not using it is the worst of both outcomes: the memory cost is paid in full, and so is the processing time. If the generative path cannot run on this hardware, the data should not be loaded onto the card in the first place. If it can run, it should be used.

**What I am asking**

1. Please confirm whether this change is intentional. If generative Remove now runs on the processor on some hardware, it would help to know which.
2. If it is a regression, please fix it. At the very least, stop filling the graphics card memory when the work will not run there.
3. Please consider an option to use the graphics card for generative tasks, so that people with suitable hardware can actually use it.

I can provide the full measurement files or repeat any test that would help.