LLM planner - pick a rig for your use-case/model/budget, or pick models for your rig. 60+ builds, 50+ models, 130+ cited t/s sources, 150+ reviewer YouTube videos, idle+active watts, multi-region prices, regular updates.
Posted by totosse17@reddit | LocalLLaMA | View on Reddit | 18 comments
TL;DR: Sourced internet info into llm model/hardware choise guide. Two directions:
- "What rig should I buy for use-case / model /budget?"
- "I have a 3090 / M3 Max / DGX Spark / Strix Halo / R9700. What runs well on it?"
Plus a side-by-side compare mode for rigs and LLMs. Tokens/sec numbers cite a source; every build links the actual reviewer YouTube videos.
Why I built it: Needed to pick what I buy, 5090, spark or strix halo. Ended up with spark made by asus. I was building it to the point where I didnt need to exit the site and go google something.
What's actually in it:
- 60+ specific build configs of all sorts, plug and play on-off switch, datacentre on-off switch
- Decode tok/s + prompt-processing tok/s at Q2/Q4/Q5/Q8 per model
- 100K promt processing time to the first token
- Idle + active power draw in watts
- Used + new prices, multi-region
- 150+ reviewer YouTube videos linked across the builds (so you can watch the review to make an opinion)
- 130 cited sources across leaderboards, model cards, llama.cpp benchmark threads, Tom's Hardware, and this sub
- Reverse mode: paste hardware -> see open-weights that fit, ranked across chat/coding/agents/reasoning, with the closed-frontier four (Gemini 3.1 Pro, GPT-5.5, Sonnet 4.6, Opus 4.7) shown as ceiling reference
- Data is updated at least once a week.
What it does NOT do:
- Gives a link to a cheapest price in your region.
- Gives absolute best tps for can get for your hardware. Mileage may vary based on quant/software, patches and updates
Link: https://llmrequirements.com
All the data is exported into public repo
https://github.com/Trenin-Labs/LlmRequirements
There's link on the website to submit benchmark or report inaccuracies using issues on github for this public repo.
johnnydotexe@reddit
I'll just keep asking AI how to AI. It's worked out for me so far.
McZootyFace@reddit
This and asking AI to research this specific sub-reddit for configs.
johnnydotexe@reddit
Isn't that locked behind (paid?) APIs? I guess I've never looked in to it, but would be nice to have.
totosse17@reddit (OP)
You can ask AI agent to scope search just to this particular subreddit
johnnydotexe@reddit
Thanks, I'll give that a try. Already use brave-search. I guess I was mislead by Claude always refusing, which I just learned may be due to a lawsuit or something. TIL.
totosse17@reddit (OP)
Fair point. I did the same thing, but then had to point out that other build exist. So it is iterative process. The site has kinda internet summarized.
Obvious-Ad-2454@reddit
Here is some feedback (all these are just my opinion, i can be wrong) :
1) I feel like doing a single page for the picker, models and all builds and then using a new page for each of llm for hardware / compare is weird.
2) I couldn't find the submit benchmark page / form.
3) Could be interesting to estimate cost per million tokens for builds based on wattage (i know it isn't easy and not the most useful but i still find it interesting)
totosse17@reddit (OP)
First page with picker is a landing, but if navigation is unclear I can improve it with feedback
You need to select the build and than it is at the bottom.
I can add this since it can be calculated from tps+estimated time active + electricity price in your region. But it does not include prompt processing. I did such calculations, energy efficient builds usually break even after 6 years of 8 hours per day.
Obvious-Ad-2454@reddit
Maybe having one page per section is better ? (That's my feeling right now but maybe other people won't agree).
Do you think you could add a button that is easier to find to add builds ?
I know it seems easy to you but I feel like I had to search a long time to find this way of adding a build.
totosse17@reddit (OP)
Visibility of add benchmark, build I can definitely improve. I will add it somewhere visible.
It is already kinda one page per section Main is hardware for llm Next llm for hardware Next is compare
But since main is also landing it has below all the builds, models, sources for whose who decide to scroll beyond the hardware selection.
dorugamer@reddit
The watts and multi-region pricing matter way more than people think, because a cheap rig can get ugly once you run it 8 hours a day or factor in local electricity rates. I’d also want it to flag RAM bandwidth and context length, since t/s numbers without those two usually lie.
totosse17@reddit (OP)
RAM bandwidth is there. TTFT @100k context is also there. Context lengths is model param and it is also on the models values.
BevinMaster@reddit
Ngl in some of the hardware setups shown, I don’t think K quant and llama.cpp would be realistically used by users, same comment I made before I don’t think a single gpu mi3xx workstation exists, benches made where from using one of the 8 gpus of a server.
totosse17@reddit (OP)
Data center hardware is given mostly as a reference
Dany0@reddit
Slop
totosse17@reddit (OP)
I used AI to create it yes. But do you know a better solution?
Ashraf_mahdy@reddit
TPS is the new FPS
You can't change my mind
Hardware Unboxed in 5 years: The RTX 8090! benchmarked in 55 Local LLMs, you won't believe how fast it is for only 9999
totosse17@reddit (OP)
It is already like this. Review from llm side get more views on YouTube than gaming pov.