Korea AI Chip - DEEPX NPU . Price? Under 50$ . Better that GPU?

[-]

LumpyWelds@reddit

This seems to be their version of the Hailo-8.

Hailo-8 provides 26TOPS, DX-M1 provides 25TOPS.

And while DX-M1 is way cheaper, it's not consumer ready. Hailo-8 is around $199 in an M.2 format compared to DX-M1's $50 as a bare chip.

They need to actually come out with a usable consumer version instead of just a bare chip.

Until they do, Hailo-8 is better deal.

[-]

Much_Screen7100@reddit

I think Radxa has a model, maybe 5B+, or something like that, with it incorporated.

[-]

LumpyWelds@reddit

Oooohh..

The 5B+ aint it. It is rated at 6 TOPS

But might be this one with 30 TOPS:

https://radxa.com/products/orion/o6/#techspec

For me, I still like the idea of an M.2 mount. Theirs a PCIe card that uses that to host 8 Hailo-8s

[-]

notAllBits@reddit

I believe real competition for NVidia will be coming from designs as this "analog computation chip" by https://vaire.co/ probably not for 50 bucks though

[-]

BloodSoil1066@reddit

I was watching a vid on that last night: https://www.youtube.com/watch?v=2CijJaNEh_Q

was surprised to see something new in electronics

[-]

MrTubby1@reddit

This presenter is so embarrassing. He's asking such complicated roundabout questions. I'm a native speaker and I barely understand what he's trying to ask.

[-]

BloodSoil1066@reddit

I expected him to suddenly start talking about a flat earth

[-]

BloodSoil1066@reddit

Chip is $50, but the dev kit is still $2000 (but that does come with support apparently)

It might be OK, there have been a bunch of cheap single digit TOPS NPUs before, but they all died a death because their support was awful and not even experts could get anything working

[-]

MayorWolf@reddit

CES Snake Oil. Their primary example shown is basically an image classifier. You can run Segment Anything on the cpu. You can run it on a raspberry pi.

[-]

notAllBits@reddit

I believe real competition for NVidia will be coming from designs as this "analog computation chip" by https://vaire.co/ probably not for 50 bucks though

[-]

lothariusdark@reddit

Where are all these likes coming from?

None of the chips mentioned on the website are for generative AI.

They are focused on segmentation and object detection(yolov, etc).

This is pointless here, LLMs are memory limited, these chips dont help at all.

[-]

bi4key@reddit (OP)

I logged in app, my post have ONLY 2 UP vote.

When I copy link (to my post) and paste to browser (where I not logged), now show 14 up vote?

How Reddit count ratio UP / DOWN vote? Strange.

[-]

T-Loy@reddit

AFAIK Reddit obfuscates likes sometimes to make it more difficult for bots to game the system.

[-]

uti24@reddit

Well, presentation makes no sense.

First of all, we know that AI needs memory and there is no way that thing has a lot of memory for 50$

Second, if you look closely, on their screen you can see that "thing" has 27 fps/TOPS and GPU has 1 fps/TOPS, so really, we are not comparing raw power here, but rather algorithms.

But really, we don't have enough information here, only hopes and promises.

[-]

HiddenoO@reddit

A demo like that is frankly useless if what they're comparing to isn't setup by a reputable source (like Nvidia themselves).

It's easy to make any chip look bad if your code is bad and/or unoptimized, and they don't exactly have an incentive to optimize the code for something they're comparing their product to.

[-]

Pedalnomica@reddit

Trustworthy?... Nvidia is well known for inflating the performance gains of new generations with similar tricks.

[-]

By trustworthy, I don't mean trustworthy in general, I mean with respect to that specific benchmark being set up in a way it doesn't leave any performance on the table. Obviously, you'd still want to validate the results.

[-]

gomezer1180@reddit

You’re missing the point. NVIDIA is never going to be a trustworthy actor when comparing their product with a competitor. No one does that, from Apple to Intel to whatever, you don’t trust the competitors numbers because they are always biased.

It has to be a 3rd party doing the testing and not a 3rd party like Linus tech tips, because they are known to give higher rating to whoever gives them the most money, or who they have a business relationship with.

[-]

HiddenoO@reddit

Did you even read my comment?

You're not supposed to trust their numbers. You're supposed to take their benchmark, validate their results, and then compare yours against that.

I literally just clarified that "trustworthy" is only meant in the context of "trust that they won't mess up setting up their own hardware", not "trust that they won't try to trick you" - the latter is irrelevant here because you're validating their results first.

[-]

uti24@reddit

Also true, I guess what they want to say, that for extremely narrow task their "chip" could be more effective.

[-]

DataPhreak@reddit

NVidia has a similar product in the Orin series of SOCs. the 8gb card is currently priced at $250. The thing is, the Orin has CUDA cores. (Worth noting that their 32gb board is like 3 grand)

I'm assuming that they can scale ram and this $50 card is a base model. Maybe even as low as 4gb, but is on ARM and could therefore scale to 64gb since NPUs don't require VRAM. Maybe higher, depends on the board. Assuming you are paying close to at cost for the ram, a 64gb card could theoretically be less than $1000. This is all just theory crafting.

However, 27 TOPs isn't amazing.

I think it's a nothingburger.

[-]

-Mainiac-@reddit

also he mentions that their chip uses "integer 8bit" instead of "floating point 32bit" (8:55)

[-]

DataPhreak@reddit

INT8 is the primary quantization for all NPUs. This is normal.

[-]

_thedeveloper@reddit

True, but precision of floating point wouldn't matter if you reach the same conclusion. I ain't advocating for them but looks like this could be another industry emerging.

[-]

_thedeveloper@reddit

Yeah, looks like what you said is true. But the presentation is done with no previous context. They seem to be working on providing accurate results of embodied AI (robots, robo-dogs and others). They do provide value for as long as the embodied AI needs vision processing. $50 is actually a good price to get there.

[-]

LaOnionLaUnion@reddit

Korean researchers have a habit of promising something game changing to generate interest and investment even though they don’t have results that are peer reviewed and replicated in other labs.

I trust this even less. Like it could be true, but I suspect is really just trying to generate investment

[-]

yoomiii@reddit

Why do so many people type that instead of than? Is it a typo? Or is it actual bad grammar?

[-]

brahh85@reddit

back in my days intel was 25 times bigger than nvidia , now nvidia is 38 times bigger than intel.

One day a company designs something so good that it makes you look prehistoric, because you used your dominance on the market to cash out everyone until the last dollar, instead of developing cutting-edge technologies that surpasses by far your previous products.

In 2007 nokia had the 50% of the mobile phone market... that year apple introduced the iphone... where is nokia and where is apple now.

The same happened to IBM. Think on apple, microsoft, alphabet, amazon and nvidia being together a company, IBM was bigger than that, dominating every sector(hardware, software, enterprises, research...)... and now IBM size is like 6% of microsoft.

[-]

bi4key@reddit (OP)

Thx! Nice comparison.

Now we have another company. Groq (not X from Musk) and they LPU: https://groq.com/about-us/

The Groq Language Processing Unit, the LPU, is the technology that meets this moment. The LPU delivers instant speed, unparalleled affordability, and energy efficiency at scale. Fundamentally different from the GPU – originally designed for graphics processing – the LPU was designed for AI inference and language.

And they chat: https://chat.groq.com/

[-]

"Any headline that ends in a question mark can be answered by the word no." - Betteridge's law of headlines

[-]