github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #9306] Ollama performing better with fewer threads #6071

New Issue
Open
opened 2026-04-12 17:24:03 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-04-12 17:24:03 -05:00
Owner
Copy Link

Originally created by @AncientMystic on GitHub (Feb 24, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/9306

i benchmarked my ram with this rambenchmark tool

image

https://github.com/rsusik/rambenchmark

And i noticed 7 threads hit higher than any other, so i switched ollama to 7 threads instead of 16, which is one below the total core count.

Results:
16 threads = 0.83t/s
7 threads = 6.56t/s

Maybe this will help someone else too, i was wondering why my token rate was so low and there it is, apparently less threads and less CPU usage equals a 155% improvement in performance, as i guess setting the full number of threads is over allocating the CPU.

This was specifically testing phi4 14b Q4_K_M (8.4GB) that consumed 10gb ram and offloads 36% to CPU

Testing other models that exceed vram:
M-MOE-4X7B-Dark-MultiVerse-UC-E32-24B-D_AU-max-cpu-IQ4_XS (12.5gb) consuming 15gb ram/vram 58% cpu/42% gpu:

2.7 t/s

DeepSeek-R1-Distill-Qwen-14B-IQ4_XS (7.6gb) using 10gb 38% cpu/62% gpu:

4.93t/s

DeepSeek-R1-Distill-Llama-70B-abliterated.IQ4_XS (35.6gb) using 41gb 84% cpu/16% gpu:

0.93 t/s

(I didnt even dare load 70b before since it was dipping to 0.x tokens on 14-24b)

Testing models that fit into vram, it seems to have absolutely zero effect positive or negative on token speed although ollama feels a bit snappier and seems to load them into ram/vram faster.

Specs:
OS: Ubuntu 22.04 kernel v6.0.0 in a vm running on Proxmox 8.3.4
CPU: i7-7820X 8c/16t
Ram: 96gb (8 sticks) 42gb allocated to VM
GPU: Nvidia tesla p4 8gb with 7gb vGPU profile in vm
Ollama: 0.5.11 (flash attention enabled, kv cache = q8_0 )

Originally created by @AncientMystic on GitHub (Feb 24, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/9306 i benchmarked my ram with this rambenchmark tool ![image](https://github.com/user-attachments/assets/9d540967-7ec7-4d01-8483-23566a82ca15) https://github.com/rsusik/rambenchmark And i noticed 7 threads hit higher than any other, so i switched ollama to 7 threads instead of 16, which is one below the total core count. Results: 16 threads = 0.83t/s 7 threads = 6.56t/s Maybe this will help someone else too, i was wondering why my token rate was so low and there it is, apparently less threads and less CPU usage equals a 155% improvement in performance, as i guess setting the full number of threads is over allocating the CPU. This was specifically testing phi4 14b Q4_K_M (8.4GB) that consumed 10gb ram and offloads 36% to CPU Testing other models that exceed vram: M-MOE-4X7B-Dark-MultiVerse-UC-E32-24B-D_AU-max-cpu-IQ4_XS (12.5gb) consuming 15gb ram/vram 58% cpu/42% gpu: 2.7 t/s DeepSeek-R1-Distill-Qwen-14B-IQ4_XS (7.6gb) using 10gb 38% cpu/62% gpu: 4.93t/s DeepSeek-R1-Distill-Llama-70B-abliterated.IQ4_XS (35.6gb) using 41gb 84% cpu/16% gpu: 0.93 t/s (I didnt even dare load 70b before since it was dipping to 0.x tokens on 14-24b) Testing models that fit into vram, it seems to have absolutely zero effect positive or negative on token speed although ollama feels a bit snappier and seems to load them into ram/vram faster. Specs: OS: Ubuntu 22.04 kernel v6.0.0 in a vm running on Proxmox 8.3.4 CPU: i7-7820X 8c/16t Ram: 96gb (8 sticks) 42gb allocated to VM GPU: Nvidia tesla p4 8gb with 7gb vGPU profile in vm Ollama: 0.5.11 (flash attention enabled, kv cache = q8_0 )
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#6071
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?