Intel: Our upcoming AI chip will be cheaper, run cooler than Nvidia, AMD options
- Thread starter JournalBot
- Start date
Upvote
51
(53
/
-2)
Considering the focus of marketing datacenters is on their power consumption, I suppose there might be a niche in being the one company marketing lower cooling costs and cheaper chips, but you're also making yourself uncompetitive in the space where dense, performant compute is being paid for at an eye-watering premium. Seems more like a compromise of necessity rather than a bet on there being a new market.
Upvote
20
(20
/
0)
Yeah, let's take something that'll cause regional electrical rates to skyrocket and then make them louder.
Upvote
61
(68
/
-7)
Seems to be that no matter how good or bad the chip is, being air-cooled alone will make it a runaway hit with the AI market.
Upvote
-10
(2
/
-12)
Then again if I was a business running a local model for something, being able to chuck in a dedicated inference card or two in a normal desktop case or server in the closet would be a massive game changer. After I got my models dialed in, it would be great for improving latency and preventing data leaks. And also save metric tons of money on not having to pay the Nvidia tax or the power bills of those crazy liquid cooled monstrosities.
Upvote
31
(34
/
-3)
Is there an inference chip market in China at all now? My impression is that thanks to the trade war, the CCP is pushing Chinese companies towards homegrown chip companies (e.g. Deepseek running on Huawei inference). The latter can be subsidized and/or coerced to have price points far lower than any Western company can stomach (e.g. Deepseek v4 is ~1% the price of Claude Opus for ~50% the intelligence).
Upvote
14
(14
/
0)
Yeah, because the liquid cooled CPUs get water straight from the river by the thousands of gallons, not like they are in a closed loop or anything, whereas an air-cooled cpu needs just air that then needs to be chilled by chillers that evaporate water
/s
Upvote
8
(17
/
-9)
Just because the chips are water cooled does not mean that the data center is also water cooled.
Many (most?) new datacenters proposed use heat exchangers, where the chips are water cooled, and then the water is cooled down. with air.
Just like the radiator in your car.
https://www.microsoft.com/en-us/mic...n-datacenters-consume-zero-water-for-cooling/
Upvote
22
(25
/
-3)
When you go with an AMD chip you’re paying margin to the chip designers (AMD) and margin to the fabricator (TSMC). With an Intel in-house chip you’re paying margin to Intel that incorporates both margin for the design and margin for the fabrication.
But the only way the Intel chip could be cheaper is if it’s making less margin than AMD plus TSMC. So either they’re planning on making less margin on their design services or less margin on their fabrication services or both. Taking less margin is not something you have to do when you’re making a desirable product!
Upvote
12
(12
/
0)
Your link is marketing material. From my understanding the majority of data centers use a combination of both closed and open loop cooling.
Open loop cooling is more efficient since you gain the evaporative effect as well.
Last edited:
Upvote
17
(17
/
0)
Pity. Intel had a great opportunity to step up and be the consumer based company that Nvidia and AMD are abandoning with their GPUs.
Upvote
13
(13
/
0)
Steam very much suggests otherwise though, almost all of Intel's share is casual gaming on non-gaming laptops. Intel Arc has like 0.3% marketshare, there are more 5090s gaming than all of Intel's discrete graphics cards combined. And then when Intel itself can't seem to decide if they want to commit or not I don't blame consumers for staying away.
Upvote
7
(9
/
-2)
Lower power usage and lower price tend to allow for increased density. Data center and maybe even rack level power consumption would be similar with these.
Upvote
1
(1
/
0)
Ruining the planet for >40% hallucination rate, it would be hilarious if it wasn't so disastrous. All for the sake of companies wanting endless growth and profits they really don't need as they are already have market caps in the trillions.
Not to mention all the data everyone is handing over to it, it's a companies wet dream.
Not to mention all the data everyone is handing over to it, it's a companies wet dream.
Upvote
7
(14
/
-7)
RickRoyLeonPrisZhoraRacha
Ars Scholae Palatinae
Can't wait for someone for find all the new flaws in the Intel AI chips (Meltdown, Spectre, ...). Because oh.. wait. Will Intel use AI to design itself? Oh boy...poison coder slop 2.0
Air cooling still needs chillers (think waterfalls with fans). So water extracts the heat from the processing, then is spent over panels that air cools via evaporation of the heated water. The water that remains is cooled and pumped back into the cycle. Any missing water through evaporation, is replaced from the municipal source. Building cooling also uses water on hot days to be sprayed on the HVAC cooling radiators to evaporate and cool them down. Lastly are the interior humidity systems that when its hot out, inside air is also mixed with outside air and to cool it, water is misted into the system. Hot, dry air is bad for components, as well as condensing air is bad for components.
Whether air cooled or water cooled, water is in demand.
No one asked for this. Silicon Valley created this, and now getting people addicted to the use of AI instead of using their brains. Humanity is at a precipice and technology is going to wipe us out in less than 100 years.
Air cooling still needs chillers (think waterfalls with fans). So water extracts the heat from the processing, then is spent over panels that air cools via evaporation of the heated water. The water that remains is cooled and pumped back into the cycle. Any missing water through evaporation, is replaced from the municipal source. Building cooling also uses water on hot days to be sprayed on the HVAC cooling radiators to evaporate and cool them down. Lastly are the interior humidity systems that when its hot out, inside air is also mixed with outside air and to cool it, water is misted into the system. Hot, dry air is bad for components, as well as condensing air is bad for components.
Whether air cooled or water cooled, water is in demand.
No one asked for this. Silicon Valley created this, and now getting people addicted to the use of AI instead of using their brains. Humanity is at a precipice and technology is going to wipe us out in less than 100 years.
Upvote
-3
(6
/
-9)
I really despise these shallow, credulous FT articles. They're nothing but click bait, no real analysis, nothing but regurgitation of a few quotes from "insiders" who usually just parrot a corporate line and no difficult questions. It's industry cheer leading at its worst.
That said, some of the commentators are kinda missing the point. Executives like to do two things when they can get away with it: check lists, and 'save' money. Check lists makes the sloppy real world shallow and abstract. They don't have to think about the real problems behind those check lists, hence the slate of unaddressed problems in corporate America: unlevel playing fields, growing numbers of disgruntled employees and contractors and activist hackers (you should read 'engaged security researcher' in layman's terms and context) turning on corporations that treat them poorly, abysmal physical and technical security, resigned consumers, so on.
Saving money is more of a creative endeavor, in that its interpretation is generally up to how the question is asked. If you can present a proposal to the middle and upper executives how you can 'save' them money in language they can understand, they're almost always going to listen to you, especially if they've already drunk the latest hype train koolaid and you can tell them how to do <it> cheaper.
That's one reason Intel's stock is rising, the current upper management at Intel is saying all the right things to the investor crowd who doesn't know enough about how much the technical details matter to the actual customers, like "Does it run my already established CUDA-based code at sufficient, provable performance to get me to buy?" Nvidia's hardware is fine, but that's not really what has customers not budging from the three incumbents, Nvidia, AMD, and Apple even on inferrence side... actually Apple is where the inference side shines with the unified memory architecture. Probably why new orders of Mac Mini/Studio have delivery dates into mid to late summer now while the iMacs have a lead time of a week or two (I checked yesterday for different reasons). The reasons they aren't budging is because their code bases are already dependent on the respective SDKs. If those SDKs don't explicitly support Intel's new product, they aren't going to get many sales (certification matters to corporate and government buyers).
That said, some of the commentators are kinda missing the point. Executives like to do two things when they can get away with it: check lists, and 'save' money. Check lists makes the sloppy real world shallow and abstract. They don't have to think about the real problems behind those check lists, hence the slate of unaddressed problems in corporate America: unlevel playing fields, growing numbers of disgruntled employees and contractors and activist hackers (you should read 'engaged security researcher' in layman's terms and context) turning on corporations that treat them poorly, abysmal physical and technical security, resigned consumers, so on.
Saving money is more of a creative endeavor, in that its interpretation is generally up to how the question is asked. If you can present a proposal to the middle and upper executives how you can 'save' them money in language they can understand, they're almost always going to listen to you, especially if they've already drunk the latest hype train koolaid and you can tell them how to do <it> cheaper.
That's one reason Intel's stock is rising, the current upper management at Intel is saying all the right things to the investor crowd who doesn't know enough about how much the technical details matter to the actual customers, like "Does it run my already established CUDA-based code at sufficient, provable performance to get me to buy?" Nvidia's hardware is fine, but that's not really what has customers not budging from the three incumbents, Nvidia, AMD, and Apple even on inferrence side... actually Apple is where the inference side shines with the unified memory architecture. Probably why new orders of Mac Mini/Studio have delivery dates into mid to late summer now while the iMacs have a lead time of a week or two (I checked yesterday for different reasons). The reasons they aren't budging is because their code bases are already dependent on the respective SDKs. If those SDKs don't explicitly support Intel's new product, they aren't going to get many sales (certification matters to corporate and government buyers).
Upvote
19
(20
/
-1)
Most new datacenters are still using evaporative towers. That's why they are using so much water to operate. They could use more advanced systems that use less water, but they are also more expensive.
Upvote
14
(14
/
0)
Saying that margin would have to be reduced to be cheaper assumes that the costs are exactly the same. There are many aspects of a design that could be tweaked to allow for higher yields which would lower the costs. And since they are designing this chip to use cheaper memory and be air cooled tends to point to designing something that is more focused on manufacturability.
Upvote
10
(10
/
0)
A side effect of this being successful would be convincing RAM manufacturers to switch some HBM capacity to LPDDR5 which could unstick non-server PC sales.
Upvote
7
(7
/
0)
I'm just glad to see a few foundries staying in America. They don't have to be the AAA++ top-tiered, but good enough for commercial and consumer use. Sure, go ahead and market the questionable data centers to your stock holders. When those fail, we'll still at least have some infrastructure and employment to make technology. We lost so much skill and knowledge when we exported all of that overseas, so I'm glad to see at least some attempt to build it back up, even if it's a little misguided.
Upvote
5
(6
/
-1)
We'll see. They're talking the right talk, but Intel has wanted into the vector market for a long long time and so far no success.
I am glad they're keeping it vertical.
I am glad they're keeping it vertical.
Upvote
1
(1
/
0)
??? More focused on manufacturability than AMD, NVIDIA and TSMC who notoriously don’t give a shit about manufacturability? Because that doesn’t sound right to me. Literally everyone in the industry cares about manufacturability. The idea that Intel has some sort of silver bullet on that front is nonsense.
Upvote
-5
(1
/
-6)
You are missing though that not every company has the same goals or constraints in what they are designing. Memory interfaces are notoriously tricky to design and a very susceptible to flaws. So it is very likely that the design choice to utilize HBM memory makes those chips significantly harder to get higher yields on.
Also sure Intel isn't at its peak but that doesn't mean that they don't have vast institutional knowledge in the semiconductor space.
Upvote
4
(5
/
-1)
Actually, it would have the opposite effect. It would relieve some minor pressure on the HBM package capacity, and lay it even more heavily on DDR5 packages. Consumer grade hardware packaging would lose out to the enterprise bulk buyers. The packaging facilities have finite output. They're already out of capacity this year for all packages to be delivered early next. This wouldn't change any of that capacity shortage because all of that will still go to the big bulk buyers first. You and me and other small time buyers will still miss the train because we can't afford the ticket. It's not exactly supply and demand, because there are thumbs on the scales at most levels, but when X, Y & Z are willing to pay $1000 for a ticket, the person that only has $200 is going to be walking.
Earlier I pointed out that Apple's Mac Minis have a lead time of 4-6 weeks or even more. That's partly because it's become well known the Minis make excellent inference platforms with the built in unified compute platform, CPU/GPU/NPU and a unified memory architecture that doesn't distinguish between VRAM and system RAM. Makes them highly flexible in comparison to the discrete GPUS in PCs. On top of that, one Mini as a package deal is the same price as just a low end GPU on PCs (like the RT 9070 XT with 16 GB VRAM is ~ $700 USD last I checked, which is the price of a 16 GB Mac Mini).
Last edited:
Upvote
1
(2
/
-1)
Are they?
Project Jupiter in New Mexico is designed as closed loop with air cooling.
EdgeCore in Phoenix is also closed loop with air cooling.
Meta in Louisiana is closed loop with air cooling.
A half dozen other ones I looked at that were announced were all closed loop with air cooling.
It's basically the "new" standard for buildout. It's why Nvidia is releasing chips that can run at a hotter temperature; a temperature that can be more readily met with closed loop cooling.
Upvote
4
(4
/
0)
Yes, open loop cooling is more efficient.
And, yes, the majority of data centers use open loop cooling.
I was talking about data centers planning to be built; most of those are closed loop cooling with none or very little evaporative cooling.
I can't think of a proposed data center in the Southwestern US that isn't closed loop cooling.
The thing is that if we want to retire the old data centers that evaporate a ton of water, we actually do have to allow new non-evaporative data centers to get built.
Last edited:
Upvote
2
(2
/
0)
Not necessarily.
It depends on cost and performance. They've spent an insane amount of money to get things to where they are now, and they'll have to spend slightly less insane amounts of money to replace what we have now with what's newer and cheaper.
In looking at the financials, which is to say the bottom line, the questions arise: When will AI make a profit? How much will that profit be? How long will it be before investors see a penny in returns? How long will it then be for investors to even RECOVER what they invested (no interest, no profits above that)?
There's been a lot of promise with absolutely no hard cash being delivered on those investments.
Considering that roughly two trillion dollars have gone into this still highly unprofitable endeavor, my thought about this kind of bullshit is "too little, way too fucking late".
What they're PROBABLY looking at is token cost. If air-cooling is going to be cheaper, and then help get the cost per token into the black, then yeah, that's their sole motivation. But that's not a guarantee, especially with the extra capital they need to replace the hotter chips and all that infrastrucure with new ones - or just build new data centers from scratch and apply the design in the build.
This sounds like an accommodation to a problem that the AI companies probably can't afford to do without more VC funding. And that lake is getting drier and drier as time goes on.
Upvote
0
(0
/
0)
ITT: people who have seemingly been unaware that data centers have existed for decades and yet are now suddenly experts on how they're run
Upvote
-5
(0
/
-5)
I wonder how that guy from WSB is doing with his grandma's inheritance dumped into Intel stock.
Upvote
0
(0
/
0)