Honestly good for them. US tech CEOs deserve to have their lunch eaten for ducking the industry into stagnation with their short sighted greed.
this post was submitted on 29 Jan 2025
148 points (96.8% liked)
Technology
61300 readers
2466 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 2 years ago
MODERATORS
In one story they're using PTX on Nvidia H800s. In another they're on Huawei chips.
Which is it? Are we all just hypothesising?
Not the best on AI/LLM terms, but I assume that training the models was done on Nvidia, while inference (using the model/getting the data from the model) is done on Huawei chips
To add: Training the model is a huge single-cost expense, while inference is a continuous expense.
Wait, so after you train, you don't need all those fancy Nvidia chips?
They should make one place where there is an overabundance of geo thermal energy and train all models there...
Yes, so R&D and finalizing the model weight is done on NVIDIA GPUs (I guess you need an excessive amount of VRAM).
Inference is probably gonna be offloaded to consumers in the end where the NPU is taking care of the inference cost (See Apple, Qualcomm etc)
Yes. You still need similar ones if you want to run the models really fast, but not nearly the same amount or cost. That's how people run LLMs on their laptops. You don't even need a GPU, a multi-core CPU is sufficient, just not very fast at it.
You do need great hardware, but it depends on your use case. If you want the full 671 billion parameter R1, you need to run it on specialized hardware that has enough RAM.
If you want to run R1 on a phone, you could get the 1.5B parameter R1 running as well. But the quality of results and the speed of response diminish significantly depending on the model and the hardware you use.
In Iceland they run their Bitcoin Mining facilities fully on geothermal energy. I wouldn’t be surprised to find the EU exploring there options regarding new data centers built on renewable energy for quite some time. For now it is a lot faster to train the models within existing data centers that already have the hardware while everyone is actively competing.
Meanwhile governments and corporations are trying to pull money out their ass (cutting important programs) to move mountains and create AGI, of which we have no evidence this is the way to accomplish that.
An unknown quantization of R1 is running on the 3rd iteration of outdated 7nm hardware taken from Sophgo's work with TSMC last year?
Is this meant to be impressive or alarming? Because I'm neither.