Will an open sourced SOTA LLM be trained on Intel hardware by 2024?

190Ṁ359

Dec 31

14%

chance

ALL

In this market, we define

Open sourced as in the weights are released to the public to be downloaded

SOTA as in better performing than every other open sourced model in HF's open sourced LLM leaderboard (https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Intel hardware as in Intel designed hardware, including Intel CPUs, GPUs and XPUs

If all of these above conditions are filfilled, this market will be marked as resolved.

Get

1,000

to start trading!

People are also trading

Will a SOTA model be trained with Kolmogorov-Arnold Networks by 2029?

8% chance

By 2026, will it be standard practice to sandbox SOTA LLMs?

28% chance

Will an open-source LLM under 10B parameters surpass Claude 3.5 Haiku by EOY 2025?

99% chance

Will the transformer architecture be replaced in SOTA LLMs by 2028?

65% chance

Will a SOTA open-sourced LLM forecasting system make major use of quasilinguistic neural reps (QNRs) before 2027?

19% chance

What will be true of the SOTA AI on the FrontierMath benchmark, before 2026?

Will we see a public GPU compute sharing pool for LLM model training or inference before 2026 ?

86% chance

Will Intel's photonics CPUs be available before 2027?

30% chance

What will be true of the SOTA AI on the FrontierMath benchmark, before 2028?

When will a non-Transformer model become the top open source LLM?

6 Comments

7 Holders

21 Trades

Sort by:

I recall Facebook was experimenting with Habanas in their servers a few years ago, but I think they settled on Nvidia by now. OpenAI was using Graphcore(I think?) & Nvidia, so not going to happen there. Most others are sticking with Nvidia. There's a niche business for FPGAs for deep learning, and Intel is one of the two manufacturers of cutting-edge FPGAs; but those are usually more for inference than training.

So this isn't impossible, but does seem unlikely.

Intel GPUs aren't good enough (they only have lower end options) and they won't be for a while, and the maker would have to disclose that they used Intel. Really unlikely.

predictedYES

@ShadowyZephyr intel's HPC gpus are very good. PVC has more VRAM than anything avaliable atm, i wouldnt discount raja koduri. Remember, he basically was responsible for GCN which is now CDNA.

predictedNO

@GiftedGummyBee It has the same amount of VRAM as AMD's offerings from what I could find. Thank you for making me aware of this though, I didn't realize they had HPC offerings at the moment.

predictedYES

@ShadowyZephyr That is very true. However, two things to consider. The first is that AMD has ROCm which is not working very well. OneAPI already has been proven to work again and again. Two, intel has been in the deep learning game for much longer, as seen with how they already had habana in 2020 and earlier. I would bet more on Intel than AMD because of those reasons.

predictedNO

@GiftedGummyBee Hmm, interesting. I didn't know Intel's Habana beat the A100 so early on. Although I think availability is still an issue regarding Intel (easy for a consumer to buy an A100, intel habana, not so much), does seem more likely than I first thought.