News Chinese chipmaker launches 14nm AI processor that's 90% cheaper than GPUs — $140 chip's older node sidesteps US sanctions

[Admin]

Intellifusion has made it possible to sidestep sanctions placed on high-end GPUs in China by producing an ASIC with excellent AI performance in their stead.

Chinese chipmaker launches 14nm AI processor that's 90% cheaper than GPUs — $140 chip's older node sidesteps US sanctions : Read more

 

[The Historical Fidelity]

So they are offering discrete boxes with the performance of a mainstream Intel or AMD cpu with an NPU. This is the product the world has ever seen lol

 

[Notton]

If this could be slapped onto a PCIe card, and used in a multi-card setup, I guess it wouldn't be bad?

 

[The Historical Fidelity]

If this could be slapped onto a PCIe card, and used in a multi-card setup, I guess it wouldn't be bad?

I guess yeah if you must have Microsoft Co-Pilot but don’t want to upgrade CPU’s you could plop one into your computer.

 

[PEnns]

"Chinese manufacturer Intellifusion is introducing "DeepEyes""

Heh, the Chinese are publicizing the real names of Google, Fakebook and Microsoft!!

 

[Alvar "Miles" Udell]

Considering the term "AI" is used so generically that it can encompass anything that uses an algorithm, I think it's more than questionable that 80% of businesses will use "AI" and that more likely 100% of businesses will, since any business that tells "AI" to, say, "Examine last year's sales and revenue and generate charts and reports", something any business from an independent contractor to a Fortune 500 company will do, qualifies as "using AI".

 

[mistermcluvin]

Very interesting. As someone who is long in NVDA, I am always on the lookout on where their competition in the AI space will come from. Is AI like BTC mining where it started with multi purpose GPUs but will eventually end up in dedicated ASICs?

 

[The Historical Fidelity]

Considering the term "AI" is used so generically that it can encompass anything that uses an algorithm, I think it's more than questionable that 80% of businesses will use "AI" and that more likely 100% of businesses will, since any business that tells "AI" to, say, "Examine last year's sales and revenue and generate charts and reports", something any business from an independent contractor to a Fortune 500 company will do, qualifies as "using AI".

AI is so popular I’m surprised there is not a hip-hop artist named Algo-Rhythm

 

[RolandOlifant]

What's communist for "keep this the hell away from me"

 

[peachpuff]

What's communist for "keep this the hell away from me"

Tofu

 

[Findecanor]

The primary driver for AI in China is the PRC's surveillance machine.

The primary driver for AI in the West is fear over China having invested so much in AI, but without understanding why.

Otherwise it is gnomes collecting underpants all the way down.

 

[usertests]

The primary driver for AI in China is the PRC's surveillance machine.

The primary driver for AI in the West is fear over China having invested so much in AI, but without understanding why.

Otherwise it is gnomes collecting underpants all the way down.

BF16/INT8/INT4/etc. acceleration can be good to have. We're long past the point where this level of skepticism makes sense.

it should soon be suitable for many AI workloads, especially considering most existing NPUs are only as fast as 16 TOPS. However, Snapdragon's X Elite chips are set to boast 40 TOPS alongside industry-leading iGPU performance later this year.

Both Snapdragon X Elite and AMD's Strix Point should hit 45 TOPS (INT8), and that's just from a tiny bit of silicon in an APU. I think an RTX 3060 is 101 TOPS.

 

[slightnitpick]

Very interesting. As someone who is long in NVDA, I am always on the lookout on where their competition in the AI space will come from. Is AI like BTC mining where it started with multi purpose GPUs but will eventually end up in dedicated ASICs?

For training wouldn't a dedicated ASIC be locking yourself into an algorithm before figuring out whether it's best? I could see a dedicated ASIC being more efficient post-training, for things like dedicated task computers (ala Bitcoin mining). But for the training itself as general purpose as possible would be the most flexible and adaptable.

 

[Pierce2623]

While some don’t like to hear it, if you’re just doing inference there’s no reason to use a GPU over a small matrix math accelerator

 

[Pierce2623]

Very interesting. As someone who is long in NVDA, I am always on the lookout on where their competition in the AI space will come from. Is AI like BTC mining where it started with multi purpose GPUs but will eventually end up in dedicated ASICs?

It depends on whether you’re talking about training or inference. For training, GPUs will still be optimal, but for inference (just running a model like stable diffusion, LLAMA etc) a small group of matrix cores makes infinitely more sense than a big GPU.

 

[thullet]

Some price comparisons:
google coral: 4TOPS for $30 @ 2 TOPS / W
tenstorrent e150: 83TOPS FP8 for $800 @ ~0.4 TOPS / W (?)
this: 48TOPS with chip at $140 @ ? TOPS / W

ofc there's additional costs to make a full card, but this is looking good

 

[Mass Effect]

PC's are not the only thing getting onboard NPU's for AI performance. This is absolutely coming to the next generation of gaming consoles afterwards.

AI is the future of NPC's and personal assistants.

 

[Pierce2623]

Some price comparisons:
google coral: 4TOPS for $30 @ 2 TOPS / W
tenstorrent e150: 83TOPS FP8 for $800 @ ~0.4 TOPS / W (?)
this: 48TOPS with chip at $140 @ ? TOPS / W

ofc there's additional costs to make a full card, but this is looking good

The Google coral is almost never used as a solo unit. It generally comes as an add in card with 8-12

 

[Pierce2623]

PC's are not the only thing getting onboard NPU's for AI performance. This is absolutely coming to the next generation of gaming consoles afterwards.

AI is the future of NPC's and personal assistants.

The PS5 Pro is already getting an NPU to run ML based upscaling on.

 

[t3t4]

Well, good for them!