github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 08:02:14 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #5745] Upgraded hardware no change in performance #29338

New Issue
Closed
opened 2026-04-22 08:06:32 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-04-22 08:06:32 -05:00
Owner
Copy Link

Originally created by @AncientMystic on GitHub (Jul 17, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/5745

Wondering if this is a bug in ollama, i just upgraded my core hardware and was hoping to see a slight improvement in token/s

From
i7-6700 (avx2 8k benchmark score)
DDR4 64gb 2133mhz (17GB/s)

To
I7-7820x (8c/16t avx512 17k benchmark score)
DDR4 96GB quad channel 2600mhz (56GB/s)

Everything about the new hardware is at least 2x faster plus avx-512 which should be able to deliver up to 5-6x the performance

But testing models i seem to be getting the exact same token/s performance, literally no change at all not even 1 token/s higher, many of the larger models are still 0.9-2 t/s and 5 for the slightly smaller, 24 t/s (at most) for the gpu accelerated models that will fit in vram.

It seems unexpected to see no changes at all in that big of a performance leap, the first one was also running in a vm the motherboard wasn't handling properly with no resizable bar or any other performance improving features that the new system has, that alone should be 10-20% improvement

Is there anything i can add to environment variables or something to maybe fix this?

Originally created by @AncientMystic on GitHub (Jul 17, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/5745 Wondering if this is a bug in ollama, i just upgraded my core hardware and was hoping to see a slight improvement in token/s From i7-6700 (avx2 8k benchmark score) DDR4 64gb 2133mhz (17GB/s) To I7-7820x (8c/16t avx512 17k benchmark score) DDR4 96GB quad channel 2600mhz (56GB/s) Everything about the new hardware is at least 2x faster plus avx-512 which should be able to deliver up to 5-6x the performance But testing models i seem to be getting the exact same token/s performance, literally no change at all not even 1 token/s higher, many of the larger models are still 0.9-2 t/s and 5 for the slightly smaller, 24 t/s (at most) for the gpu accelerated models that will fit in vram. It seems unexpected to see no changes at all in that big of a performance leap, the first one was also running in a vm the motherboard wasn't handling properly with no resizable bar or any other performance improving features that the new system has, that alone should be 10-20% improvement Is there anything i can add to environment variables or something to maybe fix this?
GiteaMirror commented 2026-04-22 08:06:33 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jul 17, 2024):

More information would be needed to diagnose the issue. Model? Version of ollama? GPU type? Output of nvidia-smi or other monitoring? CPU load during inference? Output of vmstat or iostat or other system statistics?

<!-- gh-comment-id:2233452535 --> @rick-github commented on GitHub (Jul 17, 2024): More information would be needed to diagnose the issue. Model? Version of ollama? GPU type? Output of `nvidia-smi` or other monitoring? CPU load during inference? Output of `vmstat` or `iostat` or other system statistics?
GiteaMirror commented 2026-04-22 08:06:33 -05:00
Author
Owner
Copy Link

@dhiltgen commented on GitHub (Jul 23, 2024):

I believe those two CPUs are from the same generation, and their clock speeds are ~close. You mention benchmarks but not which ones, so I'm guessing they're not specifically vector math performance comparisons.

which should be able to deliver up to 5-6x the performance

I'm not sure where you got that data. In my testing, AVX512 didn't outperform AVX2, which is why we haven't compiled a specific runner for that vector feature. We might offer a AVX512 runner in the future, and this is tracked via #2205

Where you'll really see big performance improvements is using a GPU. I'm not aware of any silver bullet to get significant CPU performance benefits, and even over multiple generations of Intel (or AMD) CPUs, I only see incremental changes in performance.

Some examples with llama3 and the cpu_avx2 runner

  • 13th gen i9-13900K: 12.39 tokens/s
  • 13th gen i7-13700KF: 9.27 tokens/s
  • 12th gen i5-12400F: 11.53 tokens/s
<!-- gh-comment-id:2244024216 --> @dhiltgen commented on GitHub (Jul 23, 2024): I believe those two CPUs are from the same generation, and their clock speeds are ~close. You mention benchmarks but not which ones, so I'm guessing they're not specifically vector math performance comparisons. > which should be able to deliver up to 5-6x the performance I'm not sure where you got that data. In my testing, AVX512 didn't outperform AVX2, which is why we haven't compiled a specific runner for that vector feature. We might offer a AVX512 runner in the future, and this is tracked via #2205 Where you'll really see big performance improvements is using a GPU. I'm not aware of any silver bullet to get significant CPU performance benefits, and even over multiple generations of Intel (or AMD) CPUs, I only see incremental changes in performance. Some examples with llama3 and the `cpu_avx2` runner - 13th gen i9-13900K: 12.39 tokens/s - 13th gen i7-13700KF: 9.27 tokens/s - 12th gen i5-12400F: 11.53 tokens/s
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#29338
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?