Honestly this is very very fresh, I was tinkering with hosting some models and wanted to optimize costs, tried few inference engines. Just want to collaborate on organizing data.
Agree, we will add a MUI table very soon. Also some charts.
I genuinely want someone to roast the way I did my benchmark process described there. Want something good enough yet easy to run.
For throughput data, well, you need to actually run prompts to gather the data which racks up costs fast and performance can vary based on input prompt lengths. The two sources I use are OpenRouter's provider breakdown [1] and Unify's runtime benchmarks [2].
Yeah we want to do exactly this, benchmark and add more data from differnt gpus/cloud providers, will appreciate your help a lot!
There are many inference engines which can be tested and updated to find best inference methods
Goodluck, companies would love that. Don't get depressed unlike my tool I think you should charge, that might keep you motivated to keep doing the work.
It's a lot of work, your target users is companies that use Runpod and AWS/GCP/Azure, not Fireworks and Together, they are in the game of selling tokens, you are selling the cost of running seconds on GPUs.
This is true especially if you are deploying custom or fine-tuned models. Infact, for my company i also ran benchmark tests where we tested cold-starts, performance consistency, scalability, and cost-effectiveness for models like Llama2 7Bn & Stable Diffusion across different providers - https://www.inferless.com/learn/the-state-of-serverless-gpus... Can save months of evaluation time. Do give it a read.
Thanks for bringing llmprices.dev to my attention. I have also a comparison page for models hosted on OpenRouter (https://minthemiddle.github.io/openrouter-model-comparison/), I do comparison via regex (so "claude-3-haiku(?!:beta)|flash" will show you haiku, but not haiku-beta vs flash.
I wish that OpenRouter would also expose the amount of output tokens via API as this is also an important criteria.
The problem with these Open weights LLMs hosted by these provider is that we don't what's the precision of the LLM, that makes a huge difference in the speed and cost (compute).
I think Together recently introduced a different price tier based on precision but otherwise it is usually dark.
and it changes the dynamics of the generative AI space completely ! absolutely exciting to watch. I am bullish on generative AI even if I think scaling laws will generate diminishing returns going forward.
Everyone should know that every single model runs differently on every system and so to decide which is best literally requires you to go through the painstaking process of running inference with each provider and then deciding. The price of inference is not sufficient to decide where to run your models.
33 comments
[ 4.0 ms ] story [ 75.2 ms ] threadhow about adding more models and providers?
and making a sortable table?
Agree, we will add a MUI table very soon. Also some charts.
I genuinely want someone to roast the way I did my benchmark process described there. Want something good enough yet easy to run.
https://github.com/BerriAI/litellm/blob/main/model_prices_an...
Simple UI to search:
https://models.litellm.ai/?q=llama3
[1]: https://openrouter.ai/models/meta-llama/llama-3.1-70b-instru...
[2]: https://unify.ai/benchmarks/llama-3.1-70b-chat
I like your charting, many have taken this task and then lose interest.
similar other tools for inspiration https://llmprices.dev/ https://www.llmpricing.app/
What no one is doing is focusing on GPUs, what is the cost of running L3-8B on an A100 or H100 per second.
It's a lot of work, your target users is companies that use Runpod and AWS/GCP/Azure, not Fireworks and Together, they are in the game of selling tokens, you are selling the cost of running seconds on GPUs.
P.S: I am from Inferless.
What sets them apart is that they have speed and latency as well.
I wish that OpenRouter would also expose the amount of output tokens via API as this is also an important criteria.
I think Together recently introduced a different price tier based on precision but otherwise it is usually dark.
Email me at ed at shadeform dot ai if we can help
Then I realized, that I changed the provider. And the new one quantized Llama 3.1 with fp8.
Then I tried Hyperbolic [2], because they offer the model in different quantizations. As result, Llama 3.1 was better than Llama 3 or at least on par.
[1] https://github.com/s-macke/AdventureAI
[2] https://app.hyperbolic.xyz/models
Without any quantization our current price is 30cts ingest and 50cts output per million tokens. [1]
1: https://withexxa.com/#pricing
[0] https://chat.lmsys.org/?leaderboard
[1] https://llmcompare.net