New variety in the chip market, and trouble for Intel (2017) | Hacker News Reader

New variety in the chip market, and trouble for Intel (2017)

economist.com

107 pointstosh73 comments

(2017), see also previous discussion from a year ago:

https://news.ycombinator.com/item?id=13736763 (55 comments)

It's interesting that AMD isn't even mentioned in the article.

The article was published a week before AMD's Ryzen processors started shipping. At that point, it was uncertain whether AMD would be able to recapture any relevance or threatening position against Intel.

Nice catch, didn't noticed the date. Strange also that cryptocurrencies aren't mentioned at all in the article, or maybe it's simply lack of research from the journalist.

If you look around in the HW forums, the blame for current GPUs prices is strong demand from cryptocurrencies mining operations, but yeah, could also be hearsay repeating.

The article date predates the Bitcoin spike of the end of the 2017. Furthermore, aren't cryptocurrencies more of a niche hobby than an actual growing field in the industry?

Yes but it would be worth mentioning when explaining Nvidia’s surge in income rather than attributing it solely to AI/ML.

But yes the article predates that.

They're a growing niche right now, and are impacting gpu availability quite strongly, but AMD/Nvidia are both quite rightly worried. If ethereum drops, or someone cracks ASIC mining (nobody mines bit coin w/ gpus anymore, but some currencies were designed with ASIC resistant algorithms) then demand could plummet overnight and literal tons of cards would show up used. Sure, ML has been a growing field, and I don't see that changing overnight, but basing a long-term business and trillion dollar fabs on crypto would be insanity.

Although I wouldn't call mining a niche hobby, there are some fairly sizeable commercial operators.

Just had a great idea for NVidia:

Pay per usage graphics cards.

Card is cheaper but you pay a monthly fee based on usage.

Gets extra revenue from the Cryptos, plus ensures a revenue base if cards become cheap again.

Just because its physical doesn't mean you can't go the way of azure, aws etc.

Yes, they will have to make do with their huge profit margins, huge free cash flow, dominant market position, and strong balance sheet. I hope they'll be OK!

Exactly. It's like when people says apple/google/ms is in trouble.

But even if they suddenly made no profit for the new 50 years, they would survive with the money they already have in bank anyway.

It’s like saying IBM is in trouble in 1980. Of course it’s not, it’s a very strong company. But...

IBM is a perfect example. The trouble was really not trouble at all. They felt the tremor, but that's it.

IBM is a withered husk of it's former self.

Adjusting for inflation IBM is worth more now than it was in 1980 and it's been kicking off a lot of dividends. Calling it a husk is misleading. Sure it failed to keep up with the growing size of the PC market, but that's a long way from failure.

The pie has become gigantic but their share of the pie has reduced to pretty much zero% from close to 100%.

They never got close to 100% and they still make ~1/4 what google does so that's far from nothing.

I've heard on the grapevine that they've been selling off real estate and other capital to stay in the black.

I don't see how that's generating 10+ billion per year in profits. Their P/E is 12 and they are worth ~130 billion. It would be one thing if they where nominally profitable, but that's a lot of money.

If you stop R&D and coast you can have outside profits for a while until your share erodes completely.

I don't think IBM has positive ROI for their R&D costs. They are mostly just a consulting shop so cutting R&D is probably a long term net gain.

IBM is not the same company it was in the 1980’s. Lou Gerstner laid off 100,000 people in the early 1990’s. Now most of its employees are non-US based:

https://www.nytimes.com/2017/09/28/technology/ibm-india.html

I think that was exactly the point :)

I agree some people seem to say these things with an unrealistically short timeframe in mind. But I think when most people say a very large company is in trouble, they're talking much more meta, because they understand that 50+ billion dollars takes decades to burn through.

Alan Mulally, arguably one of the most effective modern day leaders, has a great quote about his [former] company: "We have been going out of business for 40 years." This is what most people are talking about.

Was he talking about Boeing, Ford, or both?

He said this about Ford[0].

[0] https://www.bloomberg.com/news/articles/2007-06-03/the-new-h...

Are we talking about blackberry or Nokia here?

The point of the economist article is that those reliably huge profit margins and free cash flows may not be so reliable in the future. Nvidia is doing a great job and has released some amazing technology.

AMD has already significantly disrupted and diluted Intel's hold on the market. They have literally forced the market from anemic 2 core chips on laptops held in place without movement for 5 years by Intel.

And on the desktop to 4 cores as a minimum and 8 cores for performance at a fraction of the price Intel demands. There cannot be more disruption to Intel than this.

Intel's days of micro gains and controlled markets as progress and with it oversize profits are over.

>There cannot be more disruption to Intel than this

Yes there is and it is called server chips.

The AMD chips in question are also being sold into the server market, packaged as four dies per socket. Server product cycles and validation cycles are slow. HP and Dell and several ODMs are selling servers with AMD EPYC CPUs, they can be rented on Microsoft Azure, they're being deployed by Baidu. They're having about as much impact as could be expected from a CPU platform that's less than a year old.

I hope AMD are able to stay afloat. My main concern is that their pattern seems to heavily revolve around hiring Jim Keller repeatedly, putting the hurt on a stagnating Intel for one generation, and then fading into mediocrity as Intel are able to leverage their billions of dollars and larger, parallel tick-tock US/Israel teams to snap back. If AMD can secure a coup like Zen or K8 without him I could see this being a sustainable approach, but I'm skeptical.

Not to mention AMD's commitment to long term support for their flagship socket, AM4. While Intel might squeeze two generations out of a socket before forcing the market into yet another revision and retooling, AMD seems likely to go four or more generations of Zen processors for their desktop socket.

This may not mean much to the average user buying off the shelf complete PCs and laptops, but for those who want to maximize the lifetime they get out of a custom build, it paints a prettier picture than sticking with Intel and buying a new motherboard every CPU revision or two.

For those of us who only occasionally upgrade our hardware, its not a huge advantage either way.

In the early 2000s it was really nice being able to replace a dead Socket A motherboard (bad capacitors) with one that was several years newer, not only bringing a system back to life but adding new capabilities and improving performance, and for competitive retail prices instead the premiums you pay for new-in-box EOL hardware.

Yep, on the desktop Intel managed to only offer 4 core mainstream chips all the way from the Core 2 days up until Kaby Lake I think? Finally they've had to start releasing 6 core parts, that's a 50% bump.

I think it is all about future and whether Intel will be able to adequately respond to new challenges in order to avoid the fate of Nokia or Kodak.

Well Intel has faced Nokia and Kodak moments literally 6-7 times in the past and survived. Granted not always because their tech was better. But survival has nothing to do with just cheaper prices and better tech. Evidence of that is all over the place.

> Well Intel has faced Nokia and Kodak moments literally 6-7 times in the past and survived.

It's ironic, Kodak invented the Kodak moment but in the end the real Kodak moment was not the positive kind of thing they envisioned when they coined that term in their marketing.

Intel has mostly just had issues with a specific design, the core approach of better process = better chips is failing for the first time. If being a 18 months ahead means their chip is 5% faster that's not going to maintain huge margins by default. Without those margins they can't stay 18 months ahead.

Which means they need to compete on design for the first time in decades and I have plenty of doubts they can do so.

I don't think that's a good summary of Intel's past approach. The Pentium 4 to Core and Core 2 by way of Pentium M for example - Intel abandoned raw clock counts (what process bought them before) and started targeting IPC, which requires more engineering than pure process improvement.

Yes, there are still gains from getting more transistors for your buck, but you still have to put those transistors to work, and Intel hasn't been a slouch here. The process improvements didn't give them their advantage over AMD.

A large part of why Core/Core 2 worked was very large caches which take extreme transistor counts, but not much else. If they had been forced to cut them in half their performance would have been significantly worse.

P4 Willamette was 42M transisters December 2000, i7 2600k from January 2011 (almost exactly 10 years) is arguably the peak of their Core dominance as progress dropped to a standstill over the last 7 years had 8 MiB of cache. That's 4 cores for 1.16 billion transistors or 290 Million each.

After 1Ghz it's really just a latency game, faster ram does not do much so CPU's without massive caches just starve. Which is why the P4 eventually moved to 130 Million transistors and a relatively large cache.

I'm optimistic that they'll be able to almost solely based on their poaching of Jim Keller. If there's a major revolution in instruction sets or architecture to support ML or energy efficient HPC, he's one of the few people with the technical experience and political clout to change the future of Intel.

https://en.wikipedia.org/wiki/Jim_Keller_(engineer)

It is not a good article for a couple of reasons and disappointing for the Economist. While Intel is being challenged both in the mainstream (by ARM) and by applications like AI, the driver of NVidia profits for the last few years had been crypto coin mining.

While AI is great and all, it doesn't consume the sheer number of chips like the blockchain craze does. If history is any guide, once the block chain thing burns out or blows over, the number of available GPUs on the aftermarket is going to be pretty incredible.

I agree on the assessment of the article, however not with the conclusion that blockchain is driving this market alone - its certainly there, but I don't know any ML practitioner who answers 'how many gpu's' with anything other than 'how many can I have/afford ?' We're starting to get more and more projects reporting their processing time in GPU/years - the singularity is bull, but ML ain't going away:-) That said, if people start dumping their mining rigs, I'm on ebay looking for them :-D

I always figured that the singularity idea might work in theory but in practice it would likely take an ungodly amount of time for it to work.

> the driver of NVidia profits for the last few years had been crypto coin mining.

Do you have a source for this statement? Crypto currency demand is not the "driver" of Nvidia profits.

“We’re sold out of many of our high-end SKUs, and so it’s a real challenge keeping [graphic cards] in the marketplace for games,” he said, adding “At the highest level the way to think about that is because of the philosophy of cryptocurrency — which is really about taking advantage of distributed high-performance computing — there are supercomputers in the hands of almost everybody in the world so that no singular force or entity that can control the currency.” -- https://techcrunch.com/2018/03/27/nvidia-ceo-comments-on-gpu...

It is the mining of crypto currency that anyone can do in their garage and maybe mine a bitcoin worth thousands or some other coin of the day.

Again, Nvidia only became a real option for serious crypto miners in the last 18 months.

Until then anyone doing cost/profit analysis would have gone with AMD cards.

When Bitcoin was worth mining on GPU (ie 2011), AMD cards were much better suited than Nvidia cards.

It was about 2x difference for cards costing the same amount.

Which vendors will be hurt/helped by the GPU/blockchain hangover?

While Nvidia has benefitted from the crypto craze in the last 18 months, it is debatable that crypto mining was the primary driver.

First of all, AMD cards have been historically better suited for crypto mining(I had multiple AMD rigs in 2011) and yet still it did not drive the profits for AMD until 2017.

Nvidia's cash cow has been the workstation cards(Quattro) and Tesla computing cards(used just for processing not video) which carry much higher margins than the gaming chips while carrying the same architecture underneath.

None of those things matter if you are competing in a shrinking market.

I don't really understand what these new chips accelerate - are they useful for anything else when this current machine learning fad wears off?

Hmm, while talking about ML might be a fad it's going to be part of the plumbing now - like SQL.

ML actually works for real applications like speech recognition, face recognition, image captioning, etc. While it might be over-hyped it's not a fad that will "wear off". Google aren't going to stop using deep learning for speech recognition and go back to the old hidden markov model technique that barely worked.

Thanks - that makes sense. Guess it's probably best phrased the other way, that these chips are more useful than say a physics accelerator or a cryptocurrency mining ASIC.

Intel is too bureaucratic, slow and margin-oriented for the modern world (just like IBM). They tried to produce 10nm processors using an outdated technology and they struggle even now to make it work. Competitors, on the other side, using the newer and more expensive EUV technology, because they don't need super-huge margins. TSMC is already working on 5nm technology, while Intel is trying to fix their 10nm...

Where's competing processors?

I think the comment was about competing foundries, which are mostly using their leading-edge processes to manufacture chips that aren't going head to head with Intel's desktop and server CPUs. Instead, they're making smartphone SoCs and GPUs.

Intel produces 10nm mobile CPUs which corresponds to 7nm process of other manufacturers. And their 14nm desktop or server CPUs are still the best ones (in terms of pure performance). I'm sure that they are working on next generation CPUs. I don't see them being behind.

> Intel produces 10nm mobile CPUs

Those aren't mobile CPUs in the sense of being anywhere close to something that could power a smartphone or even a tablet. They're barely-functional desktop chips with large portions of the chip turned off and the rest down-clocked to small laptop power levels. And it's only one SKU so far, with extremely limited availability; mass production is currently scheduled for next year.

Intel's actual low-power microarchitecture is still on 14nm and can barely get its foot in the door for the tablet market. Meanwhile, two generations of smartphones have shipped using TSMC and Samsung 10nm SoCs, and TSMC's 7nm has started volume production.

And an N/A price on ARK. It's a disaster for only a 70mm^2 sized die. The no iGPU for a laptop CPU is almost a meme, and 2.2GHz (with only 3.1 turbo) dual-core at 15W just makes it even worse.

>"Intel produces 10nm mobile CPUs which corresponds to 7nm process of other manufacturers."

Can you say what's the reason for this differential?

From reading these threads previously, the answer is generally 'marketing'.

The node sizes listed aren't consistent. There's no standard, so 10nm for Intel isn't 10nm for TSMC. It doesn't mean that the transistors are 10nm,just the smallest feature. Since everyone measures this differently, it's effectively just a marketing number.

Surely the managers of Intel have read about margins and mini-mills?

10nm isn’t the same across the industry. There are no standards to defining a process node. Please watch Mark Bohr’s presentation on YouTube. The way you make your claims is kind of off-putting because of sheer confidence in your statements without properly understanding the fundamentals.

I think you've missed the point. Strip out the product names (which is all those node numbers are, as you say) and understand that the comment is still entirely correct - they are struggling with their outdated manufacturing method and now losing ground.

Intel is struggling with their manufacturing method, but they aren't wed to it. Everyone is struggling with EUV.

ASML - the guys who make the stuff that all the fab houses use, said that they only shipped 10 EUV machines in 2017. Intel's public statements on EUV is that they will use it when its ready.

TSMC and GF both manufacture their newest large scale production nodes using the same basic technique as Intel (hilarious amounts of exposures and layers at 193nm).

Everyone except Intel has said that they only plan to start rolling out EUV built chips in the later half of this year into next year.

No manufacturer has gained ground with EUV in their products yet.

Which sounds like 'everyone except Intel' has gained ground. You don't plan to go into production with something next year if you have not made it work.

I mean, you can announce your plans all you want. Intel did this with their 10nm stuff and now has significant egg on their face for their inability to deliver. Maybe Intel is just playing a tighter PR game given their recent mess up.

But really, EUV is a whole new ball game and its entirely possible for Intel to lose whatever is left of their traditional advantages in the transition.

I just don't think its wise reading too too much into everyone's PR and marketing at this time.

Semiconductor fab timelines are notoriously overoptimistic. It is undeniable that Intel's fab has been struggling for a few years, but you can't take a Fab at it's word on when it will begin production. More critical is when it reaches high volume maturity at a given node.

Fabs are all about margins, Intel has probably invested more in EUV than any other company at this point. Intel practically saved EUV at ASML by investing in ASML.

Deployment of EUV technology was initially expected in 2007. They had probably at least some valid reasons to stay with known technology. This is probably a mistake, but every company do mistakes.

Reply on news.ycombinator.com