Intel Announces Aurora GenAI, ChatGPT Competitor with 1T Parameters

https://wccftech.com/intel-aurora-genai-chatgpt-competitor-generative-ai-model-with-1-trillion-parameters/

by ksec • 3 years ago

108 86 3 years ago

jjcm• 3 years ago

Has there been any official confirmation of the number of parameters used in gpt4? Pages around the web seem to have conflicting reports on the number, ranging from 1T to 170T, but there doesn't seem to be anything official.

mchusma• 3 years ago

Just to expand your lower bar, some people suspect it is not much bigger than GPT 3.5 @ 175B parameters, but this is also speculation. I don't think there has been anything official.

rileyphone• 3 years ago

Given how PALM2 had less params than the first one (540B) I think that makes sense.

macrolocal• 3 years ago

1.4T according to Michael Douglas [1].

[1] https://youtu.be/B3Az2EONCHE

belter• 3 years ago

That is correct. The open science made at OpenAI managed to publish a 100 page GPT4 technical report without using the word parameters one single time :-) ( I checked ).

timhigins• 3 years ago

According to Intel's own press release, "Argonne is spearheading an international collaboration to advance the project, including Intel; HPE; Department of Energy laboratories; U.S. and international universities; nonprofits; and international partners, such as RIKEN." Maybe the headline should say "Argonne Labs announces GPT-4 competitor"

https://www.businesswire.com/news/home/20230522005289/en/Int...

alberth• 3 years ago

It’s interesting how it took OpenAI 7-years to produce GPT4, and in the matter of a few months Google/Intel/etc have released on par offerings.

I wonder how much of a competitive moat actually exists in this space.

gberger• 3 years ago

There is no moat, according to Google

https://news.ycombinator.com/item?id=35813322

Sai_5• 3 years ago

I think OpenAI is building a moat right in front of us through their plug-ins idea.

Executed well, plug-ins could become ChatGPT's App Store equivalent. Once that happens, OpenAI is undoubtedly going to convert a portion of search traffic into agent delegated work i.e., people will use ChatGPT to start and terminate their searches essentially delegating the work of collating data from multiple sources to ChatGPT.

FD - I've applied to release a ChatGPT plug-in and this space looks very interesting to me.

PS - Earlier today, I submitted to HN a link to my blog where I analyzed plugins. The HN post sank without a trace but the I wrote my blogpost using the Wolfram plugin in ChatGPT and it was a breeze. I genuinely feel I've seen the future.

nico2• 3 years ago

0, but

That’s in terms of technology

Their brands are still very powerful

Anyone can sell sugary drinks, but there is only one CocaCola

verdverm• 3 years ago

I don't think their brand is as strong as people believe. They are pejoratively referred to as "ClosedAI"

They have a lot of down time and change the model out from under people, only having 3 months support. It is very hard to build a product on OpenAI in this situation.

A close enough model that can run on your own hardware or cloud will get people to move. We are watching and waiting for this point for our own products. For now they are called experiments and demo applications.

m_0x• 3 years ago

I also thought it was curious. There is a wide range of possibilities for this.

My favorite hypothesis is that OpenAI was navigating the unknown open sea trying to answer the question "Is it possible to achieve this?" without knowing if they would ever find a new land.

The rest of them already know the answer is "Yes it is" so it's not an unknown open sea anymore with a dubious destination. They know it's possible, it's just a matter of money and time. They probably hired whatever engineer was available, and they paid tons of money to come up with a competitive product.

zeven7• 3 years ago

My own theory is Google had this earlier but couldn't figure out a way to make money from it. I think they felt threatened by it and didn't release what they had.

schwarzrules• 3 years ago

"matter of a few months" / "on par offerings"... My (very very) limited impression is both of those statements aren't fair. Google/intel/etc have been working on lots of AI-related projects, even LLM efforts over years. It's not like any of these companies are bootstrapping LLMs or just posted their first AI engineer job a few months ago.

I also don't think I'd call what they're offering "on par" with GPT-4. They're competitive, or impressive substitutes, but my personal tests of Bard at least aren't getting me as useful results as GPT-4.

Source: too lazy to look up any. this is just my lame impression. feel free to downvote.

danielmarkbruce2• 3 years ago

"on par" is wrong.

Intel have announced they are building a model. Anyone can announce they are building a model.

justinclift2• 3 years ago

... the target size for the new model is 1 trillion parameters.
Meanwhile, the target size for the free & public versions of ChatGPT is just 175 million in comparison. That's a 5.7x increase in number of the parameters.

5.7x seems off by several orders of magnitude, or do I need more caffeine?

mediaman• 3 years ago

It's a typo in the article. GPT-3.5 is 175bn parameters, not 175 million.

GPT-4 does not have publicly available parameter count, but is thought to have something similar to 1tn.

justinclift2• 3 years ago

Cool, thanks. :)

redeux• 3 years ago

There is an obvious typo in the article. ChatGPT (i.e. GPT 3.5) has 175 billion parameters, no 175 million.

simonw• 3 years ago

"Announced at today's ISC23 keynote, the Aurora genAI model is going to be trained on General text, Scientific texts, Scientific data, and codes related to the domain."

So this thing doesn't actually exist yet?

barthelomew• 3 years ago

Yes, it does seem like it's concept stage.

tpmx• 3 years ago

I can do a decent Daniel Gackle (dang) impression:

This looks like an announcement of an announcement, which isn't on topic for HN. https://hn.algolia.com/?dateRange=all&page=0&prefix=true&sor...

On HN, there's no harm in waiting for the actual thing: https://hn.algolia.com/?dateRange=all&page=0&prefix=false&so...

phkahler2• 3 years ago

Why train on scientific texts when half the papers are BS?

randomname93857• 3 years ago

it would be good to train LLMs to distinguish BS articles from good ones. but it's rather unrealistic expectations

yieldcrv• 3 years ago

“The 2 Exaflops Aurora supercomputer is a beast of a machine and the system will be used to power the Aurora genAI AI model.”

and instead of a data center with a distributed load of Nvidia cards, they’re going to do 1 supercomputer? Are these guys intentionally stuck in 2007 or am I missing something

whatshisface• 3 years ago

Racks full of GPU nodes are called supercomputers. Supercomputers that ran a single instance of an OS kernel are from even farther back, maybe the last century.

elcomet2• 3 years ago

It's a datacenter, not a single machine obviously. All "supercomputers" are datacenter today. They can use Nvidia or other processing units, it doesn't really matter.

mirekrusin2• 3 years ago

...which is equivalent to just 62x H100 GPUs from Nvidia?

godelski4• 3 years ago

and instead of a data center with a distributed load of Nvidia cards

You're missing something. Most models are trained on supercomputers. You still use DDP, but you also use Slurm. A supercomputer is just a many node machine, but one where you REALLY care about interconnect. It's why the fabric is one of the big points here. I/O is generally your bottleneck in these systems. Each node should be as powerful of a system as possible because of this.

A datacenter doesn't need a PB/s connection between machines. Your I/O probably isn't your bottleneck and so the tradeoff between PCIe and ethernet (infiniband or whatever) isn't a big deal. Usually because your processes are __independent__ (containers). In a supercomputer, your processes may be parallel, but that doesn't mean they're independent. You'll spend a lot of time writing code to be non-blocking but you're going to have to map reduce at some point.

HybridCurve• 3 years ago

Aurora Supercomputer:

- 200Gb dragonfly interconnect fabric

- 10,000 nodes with 2xCPU + 6xGPU with unified memory

- 10PB of agreggate RAM

- 230PB of storage

Scale matters in supercomputing. Keeping latency low between everything allows you to solve different classes of problems that gets no speed up benefit by simply offloading compute to nvidia cards. It makes a huge difference how everything is connected.

ftxbro3• 3 years ago

Not only is the model not yet trained, but the datacenter that will train it has not yet been built "the Aurora supercomputer is said to launch later this year"

godelski4• 3 years ago

Just an FYI, that's when it'll be completed and access opened up to a wider audience. Many parts are already built and being tested. We're talking about a DOE computer, not a AWS datacenter.

gigatexal• 3 years ago

so what's the api/tech powering the front end to this LLM api running on this super computer? and who pays for it?

techwizrd• 3 years ago

I'll be interested to see how this performs if/when it is released.

I'm not convinced that simply increasing the number of parameters improves the models, and I don't think we should be putting out press releases "spec-chasing" increases in the number of parameters. I've also found in my research that larger models often do not perform better, but this is becoming more difficult to explain to non-technical folks due to irresponsible marketing. If we aren't careful about the bold marketing claims, we'll disillusion many people and reduce the potential for future growth—we don't want another AI Winter.

There's a significant environmental cost to training these large models, and it's harder to understand, use, and control these very large models (even as a researcher).

EGreg• 3 years ago

What happened to “the scale hypothesis” LOL. Is scale really is all you need? Maybe this model can invent time travel and improve itself in the past..

londons_explore3• 3 years ago

I'm not convinced that simply increasing the number of parameters improves the models

At every size people say this... And then someone comes out with a model 10x larger which outperforms the previous state of the art.

Sure, you need to still take a little care with architecture choice and hyperparameter selection, but size really does seem to be king.

weatherlite• 3 years ago

we don't want another AI Winter

Speak for yourself...

Der_Einzige• 3 years ago

Parameter count is strictly better IF the number of tokens (and ideally better quality tokens) trained on increases, and if the training is done for longer (most LLMs are way undertrained)

Most of the huuuuuge models failed on most or all of these fronts and that's why they suck compared to Llama or Alpaca or Vicuna

elcomet2• 3 years ago

That's not true. For the same number of training tokens, bigger is better. And for the same size, more tokens is better. So obviously more tokens and bigger is better.

rvz• 3 years ago

There's a significant environmental cost to training these large models, and it's harder to understand, use, and control these very large models (even as a researcher).

Exactly. The worse part is there is NO viable efficient way of training, fine tuning etc and inference with these massive deep learning models for more than a decade since deep neural networks used GPUs for training, and it still requires a substantial amount of compute power and energy that is incinerating the planet and the result is untrustworthy large black-box models that cannot be trusted or transparently explain their decisions, especially with safety-critical or high risk tasks.

Crypto at least has an alternative to the wasteful proof-of-work system, and Ethereum which was formerly PoW has shown it is possible to switch to a greener alternative consensus method.

Deep Learning however, still has not shown such a viable switch and still needs to burn the planet with more data-centers of GPUs, ASICs and FPGAs to create hallucinating models that have shown to break on a single pixel or to confidently regurgitate nonsense as the truth with little understanding in reasoning and also answer with demonstrably false information.

LLMs like this one is still essentially snake-oil BS generators hiding behind regurgitation and sophistry to pretend to show signs of 'intelligence'.

EDIT: It is all true. [0] There is no amount of green-washing to hide the problem of deep learning systems wasting essential resources like thousands of running taps of water. Literally.

[0] https://gizmodo.com/chatgpt-ai-water-185000-gallons-training...

sebzim45002• 3 years ago

I'm not convinced that simply increasing the number of parameters improves the models

How so? Llama 33B is clearly worse than llama-64B and that's 'only' a 2x increase.

Are you aware of any example where a larger model fails to outperform a smaller one when all else is equal (tokens, architecture, data quality, etc.)?

Obviously for a fixed amount of training compute more parameters can be bad, but there's a trade off where more parameters means you train on fewer tokens.

londons_explore3• 3 years ago

This kind of news isn't of interest until the model is made and tested side-by-side with GPT-4 and shown to at least be in the same ballpark of performance.

So far, nobody has claimed to exceed GPT-4's performance (either by academic metrics or by youtubers subjective evaluations)

sylware3• 3 years ago

I still don't know what are neural net "parameters".

inputs+outputs? Number of neurons+synapes?

phkahler2• 3 years ago

Synapse weights.

sylware3• 3 years ago

How are they distributed? How many neurons levels? And still, how many inputs and outputs has their neural net?

Well, how are they "connected"? Their topology?

immichaelwang• 3 years ago

Isn't there more and more research coming out that at a certain point (200B~), parameters have significantly decreasing returns and it's better to just then do some supervised learning ontop of the base model?

gmadsen• 3 years ago

can you point to any of that? my understanding is that we haven't reached any dimishing returns with transformer models and scaling yet

CuriouslyC• 3 years ago

Parameters don't have diminishing returns so much as we don't have enough (distinct) data to train models to use that many parameters efficiently.

mirekrusin2• 3 years ago

It smells like somebody forgot to flush PR from it.

For starters future tense is embarrassing in this area. By the time they finish it, it may very likely be outdated.

Also number of parameters is not all you need. It's enabler, yes, but you need tons of other things to fit together.

londons_explore3• 3 years ago

Intel could just make every CPU/GPU they manufacture have a 'burn in test' for a week or so where it runs a bunch of self tests and then trains a massive network for them.

If sales slow down, instead of having the excess stock in a box in a warehouse, have it training their neural nets.

GaggiX• 3 years ago

I don't see any reason why someone should train a 1T model, they probably don't have a big enough dataset to make it worth it, so with the same compute or less you can get a better model by training a smaller one (the chinchilla scale).

davidkunz2• 3 years ago

the target size for the new model is 1 trillion parameters. Meanwhile, the target size for the free & public versions of ChatGPT is just 175 million in comparison. That's a 5.7x increase in number of the parameters.

Did nobody notice?

Daffodils• 3 years ago

Aren't you missing a 1000 there ? 10^6 / 175 = 5.7 x 10^3.

davidkunz2• 3 years ago

They mixed up million with billion.

htrp• 3 years ago

The foundations of the Intel Aurora genAI model are Megatron and DeepSpeed.

Intel is PLANNING to train a genAI model with a target of 1bn parameters... using an architecture from 2019

godelski4• 3 years ago

There's questions about why this announcement matters if it isn't trained. There's a few:

1) This is going to be using a publicly funded computer, which is the most powerful in the world AND was also announced today (btw, the specs are better than what was initially planned). The program gives justification for the machine and that public money (though in comparison to many things governments spend money on, this is very cheap).

2) THAT'S A LOT OF GPUS. This will, as far as I'm aware, be the biggest training every. GPT4 was supposedly trained on 10k A100s. Remember that Summit, the #3 computer, has >27k GPUs (V100s), but Aurora has almost 64k. This is going to be a big engineering feat in of itself. Now it won't make training 6x faster, but it will definitely be much better. It does say that the US government is taking this very seriously.

That's the point of this kind of announcement: marketing, propaganda, and cool engineering project.

ftxbro3• 3 years ago

something to keep in mind is this datacenter will be doing so many other projects as well. it's not a dedicated LLM maker. Maybe the LLM training won't have access to all those GPUs, or maybe it will have only time sharing access split with the other projects.

godelski4• 3 years ago

Not really a datacenter, it is a supercomputer. And yes, it is important to keep in mind that there aren't just single users. But with almost 64k nodes it isn't unreasonable for multiple teams to get 10k+ nodes at a time. I've seen runs on Summit use >50% capacity. Let's be clear, the teams aren't just going to start off using all the nodes at once. You step up.

ppaattrriicckk• 3 years ago

I love bashing Intel as much as the other person, but looking at the number of tokens LLaMA[1] was trained on (1,400B tokens), the number of parameters of GPT-3[2] vs. the number of tokens it was trained on (175B params vs. 300B), it's not like the announced 1,000B param model is unreasonable. Taking 175/300*1400 yields 816B parameters - which is fairly close to 1,000B.

Not outright the most efficient utilization of data, but as others have mentioned, there might still be something left to gain (from not solely optimizing for compute). E.g. look at Figure 9 in [3]. Although the model trained on the largest model obviously utilized the data most efficiently, it's not perfectly clear - to me at least - whether some over-parameterization will necessarily lead to a decrease of test/out-of-sample performance.

Of course, LLaMA was trained on way fewer parameters relative to data set size. I've just mentioned LLaMA as a point of reference to the largest dataset known to me.

[1] https://arxiv.org/pdf/2302.13971.pdf

[2] https://arxiv.org/pdf/2005.14165.pdf

[3] https://arxiv.org/pdf/2001.08361.pdf

EDIT: Formatting of references

segmondy• 3 years ago

intel can't even in hardware, and somehow they think they can do so with software? Right! This is just for the clueless investors.

Intel Announces Aurora GenAI, ChatGPT Competitor with 1T Parameters

Related Stories