78alpha / auto1111-fork Goto Github PK

This test is to determine if CPU single core performance actually matters. A generation test will be provided between all CPUs at their optimal settings. In addition, 1 CPU will be downclocked to frequency ranges of 25%, 50%, 75%, and 100%.

The second test cannot be performed on the Ryzen 7 5800X3D as it will likely kill the chip.

Testing to see the maximum performance of the available cards and what settings are needed to hit those top speeds. Direct code alterations are not permitted. This is meant to test end user speed.

Commit: 0cc0ee1

Standards:

• V1-5-pruned-emaonly
• 512 x 512
• cfg 7
• Euler a
• Steps 20

Links will direct to peak performance settings

• RTX 3070 Single | Batch8
• RTX 3090 Single | Batch8
• RTX 2060
• Tesla P40 Single | Batch8

This test will be conducted purely on Windows. This is Windows native and WSL, as installing pure Linux for this one tool is beyond reasonable for an end user. The Linux specific optimizations will be Triton, DeepSpeed, AITemplate (If possible), and TensorRT (If possible).

WEBUI: 0cc0ee1
DreamBooth: 5be87ba63f62c228cf135425e21577f70c4e3351

Users have been mentioning performance regressions over time. It's time to find them. Testing will be conducted on the end commit of each month.

0cc0ee1
17b24e4
3f401cd
0fddb4a
5b0f624
3f41756
3e41035

GPU: RTX 3070
CPU: Ryzen 7 3700X

Testing Bat file:

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--no-half-vae --autolaunch --disable-nan-check --disable-safe-unpickle --ckpt-dir A:\\CKPT\\SDModels --vae-dir A:\\CKPT\\VAE


call webui.bat

This test is to determine the maximum resolution that a card can push before it hits OOM. On Pascal series and below, they will have a lower cap due to lack of proper FP16 support.

Commit: 0cc0ee1