# P100 users, what kind of performance do you get? — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/5e37cc2a834a981d

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-09-21 14:22 ingested 2026-09-22 02:40

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1wme7y0/p100_users_what_kind_of_performance_do_you_get/)

## retained text



```
Already got one coming in the mail, figured at $80 it's cheap enough to give a go regardless, but I am curious what kind of performance people get with different models. Just picked up a new PSU yesterday to support it, 3D printer for the fan shroud and a 92mm fan coming from amazon,  a noctua unit.

  
I'm hoping to get somewhat similar performance, at least on the token generation side, to my existing setup. 

  
The machine I currently have has an RX6600XT (8GB) and 32GB of DDR4 3600 MHZ. Technically right now it's 48 GB at 3200, but I just threw some extra ram in to test something a few weeks ago. Numbers I'm giving are from the 32GB config. 

Model I'm intending to use is unsloth Qwen3.6 35B\_A3B UD\_Q4\_K\_XL. On my current setup, I get 30 t/s generation prose, 48-50 decode, both with MTP. Context is -fitc 64k full precision, run with --cpu-moe. Context rarely gets quite that full, rarely getting past 20k.

Prefill I've managed to get up to 800@0, 700 by 10k.

  
I'm *hoping* I can at least match the decode numbers. Prefill I'm less worried about; the agent it's going to be powering is a homegrown harness heavily optimized for KV cache reuse while keeping context size limited, while big enough for some longer tasks. Not coding, though the harness has all of it's tooling run through code execution rather than dedicated tools for every function of the agent. 

Prefill I'm less optimistic about. It's an older architecture, has more cores, but a slower clock speed. I'll be playing with -b and -ub to try and take advantage of the higher core count, but we'll see what it can pull. I know there are some llama.cpp forks that are focused on this particular card, and show pretty promising results on the Q2\_K\_XL version of this model, which fits fully resident in it's VRAM. I will still be offloaded (for now, until another \~$80 frees itself up), since I'm not running a Q2 model.

  
Realistically this card will be going into a dedicated machine specifically to free up the current system, which currently is serving double duty as gaming PC and inference box, reducing the times that my agent can actually be local. It runs fast enough on the existing system, so if I can get close to my existing baseline with this card I'd be happy. Plus more vram gives some flexibility I simply don't have. Swapping in other models, multiple smaller guys, or especially when I get a second or third one, this guy + a small gemma4 e4b quant for really quick, minor tasks. 
```



document fetched 2026-09-22 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1wme7y0/p100_users_what_kind_of_performance_do_you_get
