Training methods and architectures keep getting more efficient by leaps and bounds and scaling up was well into the realm of diminishing returns last I checked. The necessity of exceeding 100B seems questionable. Just because you can get some benefits by piling ever more data on doesn't necessarily mean you have to.
Also keep in mind we aren't talking about a small company wanting to do competitive R&D on a frontier model. We're talking about a world superpower that operates nuclear reactors and built something the size of the three gorges dam deciding that a thing is strategically necessary. If they were willing to spend the money I am absolutely certain that they could pull it off.
Has the Ukraine situation not shown that the EU has relegated itself to second fiddle?
It’s too old, too complacent, and too broke. Even compared to the US and our level of discord, there’s no unity across divisions.
The US absurdly threatens Greenland, but Denmark/EU’s response is “Sanction US tech or kick out US military bases on Europe”, rather than be able to rattle a saber back and show some credible backbone.
Yeah, I just don’t know that there’s the will to blow up the world economy for which flag flies over Taiwan.
China absorbing Taiwan (especially to Americans) just doesn’t seem like a radical, terrifying concept.
A Hong Kong style negotiated transfer might be best for the world - Taiwanese that want to leave can, the US can build up a parallel source of semiconductors, China gets Taiwan without firing a shot.
Is it better than the alternative? Do you think TSMC wants to see a Dongfeng or ATACMS headed for their fab, if the alternative is a negotiated handover?
There's some intersection point between long term decreasing in China's ability (demographic collapse) and long term increase in China's ability (their current build up of military hardware in air, land, and sea that is currently outpacing America's). Maybe somewhere in 10-20 years where their regional military power is much higher than America can project across the Atlantic but they still have a lot of military aged men.
Atlantic? IDK if China even has aspirations to play World Police like the US. Military protection of things like their interests and the stability of Belt and Road, sure, but I don’t see China trying something like the Gulf War or OEF.
It’s very possible that they will be able to dominate South China Sea and their zone of the Pacific, even now, given the proximity advantages and ship/missile production; and I think that would be satisfactory to them.
20 years from now, China’s sphere and America’s sphere are separate, with China having a lead in competing for Africa, and Europe in a very weird place socially, economically, demographically, and WRT Russia/US competition.
My point is that China can sustain a naval blockade of Taiwan nearly indefinitely, and at some point Taiwan will have to decide whether they want to live under siege forever (poor, cold, getting everything via scarce and expensive air freight), or give up come to a political solution.
I'm not like, rooting for this, I'm just trying to be realistic.
That's exactly what the USA has been doing to Cuba since 1959 and they're still (barely) hanging around. If we go by that example, it'll only end with with an actual invasion (which is what will happen to Cuba within one to two years).
The US has an embargo that doesn't impact other countries that want to trade with Cuba. China is going to put an actual cordon around Taiwan.
Also, the US has no historical reason for claiming Cuba and has no real domestic pressure to do so (nobody in either party is asking for it). China has been very clear they see Taiwan as a part of China and will reunite with it not for economic or strategic reasons, but for nationalistic ones.
The key here is Intel is expanding the idea of operating their fab for an external customer (foundry services). What they’re doing with specific fabs or processes is less important relative to their bigger emphasis on working for a client like Apple.
That is the traditional textbook yield curve logic, if I'm not wrong? Smaller area = higher probability of a surviving die on a dirty wafer.
But I wonder if the sheer margin on AI silicon basically breaks that rule? If Nvidia can sell a reticle-sized package for 25k-30k USD, they might be perfectly happy paying for a wafer that only yields 30-40% good dies.
Apple OTOH operates at consumer electronics price points. They need mature yields (>90%) to make the unit economics of an iPhone work. There's also the binning factor I am curious about. Nvidia can disable 10% of the cores on a defective GPU and sell it as a lower SKU. Does Apple have that same flexibility with a mobile SoC where the thermal or power envelope is so tightly coupled to the battery size?
I am curious about the binning factor too since in the past, AMD and Intel have both made use of defect binning to still sell usable chips by disabling cores. Perhaps Apple is able to do the same with their SoCs? It's not likely to be as granular as Nvidia who can disable much smaller areas of the silicon for each of their cores. On the other hand, the specifics of the silicon and the layout of the individual cores, not to mention the spread of defects over the die might mitigate that advantage.
They do bin their chips. Across the range (A- and M-series) they have the same chip with fewer / disabled cpu and gpu cores. You pray a premium for ones with more cores. Unsure about the chip frequencies - Apple doesn’t disclose those openly from what I know.
> They need mature yields (>90%) to make the unit economics of an iPhone work.
Sauce on the number?
iPhones are luxury goods with margins nowhere near typical for consumer electronics. Apple can easily stomach some short term price hikes / yield drops.
With current AI pricing for silicon, I think the math’s gone out the window.
For Apple, they have binning flexibility, with Pro/Max/Ultra, all the way down to iPads - and that’s after the node yields have been improved via the gazillion iPhone SoC dies.
NVIDIAs flexibility came from using some of those binned dies for GeForce cards, but the VRAM situation is clearly making that less important, as they’re cutting some of those SKUs for being too vram heavy relative to MSRP.
> For Apple, they have binning flexibility, with Pro/Max/Ultra, all the way down to iPads
The Pro and Max chips are different dies, and the Ultra currently isn't even the same generation as the Max. And the iPads have never used any of those larger dies.
> NVIDIAs flexibility came from using some of those binned dies for GeForce cards
NVIDIA's datacenter chips don't even have display outputs, and have little to no fixed-function graphics hardware (raster and raytracing units), and entirely different memory PHYs (none of NVIDIA's consumer cards have ever used HBM).
The desktop 4090 uses the AD102 die, the laptop 4090 and desktop 4080 use the AD103 die, and the laptop 4080 uses the AD104 die. I'm not at all denying that binning is a thing, but you and other commenters are exaggerating the extent of it and underestimating how many separate dies are designed to span a wide product line like GPUs or Apple's computers/tablets/phones.
"Ultra" isn't even binned - it's just 2x "Max" chips connected together.
Otherwise, yes, if a chip doesn't make M4 Max, it can make M4 Pro. If not, M4. If not, A18 Pro. If not that, A18.
And even all of the above mentioned marketing names come in different core configurations. M4 Max can be 14 CPU Cores / 32 GPU cores, and it can also be 16 CPU cores and 40 GPU cores.
So yeah, I'd agree that Apple has _extreme_ binning flexibility. It's likely also the reason why we got A19 / A19 Pro / M5 first, and we still don't have M5 Pro or M5 Max yet. Yields not high enough for M5 Max yet.
Unfortunately I don't think they bin down even lower (say, to S chips used in Apple Watches), but maybe in the future they will.
In retrospect, Apple ditching Intel was truly a gamechanging move. They didn't even have to troll everyone by putting an Intel i9 into a chassis that couldn't even cool an i7 to boost the comparison figures, but I guess they had to hedge their bet.
> yes, if a chip doesn't make M4 Max, it can make M4 Pro. If not, M4. If not, A18 Pro. If not that, A18.
No, that's entirely wrong. All of those are different dies. The larger chips wouldn't even fit in phones, or most iPad motherboards, and I'm not sure a M4 Max or M4 Pro SoC package could even fit in a MacBook Air.
As a general rule, if you think a company might ever be selling a piece of silicon with more than half of it disabled, you're probably wrong and need to re-check your facts and assumptions.
There are two levels of Max Chip, but think of a Max as two pros on die (this is simplification, you can also think of as pro as being two normal cores tied together), so a bad max can't be binned into a pro. But a high-spec Max can be binned into a low-spec Max.
Datacenter GPU dies cannot be binned for Geforce because they lack fixed function graphics features. Raytracing acceleration in particular must be non-trivial area that you wouldn't want to spend on a datacenter die. Not to mention the data fabric is probably pretty different.
The A40, L40S and Blackwell 6000 Pro Server have RT cores. 3 datacenter GPUs.
If you want binning in action, the RTX ones other than the top ones, are it. Look for the A30 too, of which I was surprised there was no successor. Either they had better yields on Hopper or they didn't get enough from the A30...
On the flip side, it massively increases your BOM and maintenance considerations, as now you have to have all the bits of an electric powertrain and most of the bits of a gas powertrain. All for the few times that you’re driving more than 300ish miles in a stretch?
You don’t have any savings in the bank, right? That money you’re hoarding could be buying mosquito nets to save lives - you’re killing people by not donating everything you have.
There is no moral requirement for me to impoverish myself in response to an idiot cutting government/public spending on critical assistance to those in need.
There might be other moral imperatives which indicate that I ought to cash out the 401(k) and give it to people who need support, but this guy and his fucked up "DOGE" bullshit ain't it.
Note the same, maybe if my bank account was equivalent to Elon Musk's it would be a fair argument but hardly the same to expect a shitposter to be equivalent to a man who is a billionaire.
Totally the same. $1 is $1, and it can go to saving those very lives you’re talking about. Put your money where your mouth is - otherwise you just want to virtue signal with my tax dollars.
I do, through my tax dollars. And the amount of money that was DOGEd was literally couch cushion change on the scale of the federal budget. And not only did those cuts directly lead to deaths but weakened US soft power all around the world, letting China step in.
China’s soft power play yielded returns: ports, minerals, oil, factories, customers for their exports.
The only Americans benefiting from the existing aid scheme are the network of lobbyists and NGOs.
As I said upthread, if you’re that motivated, donate $500 to a high-impact charity and you’ll do far more good for people on the ground than what your taxes were doing at USAID.
In the US, two moderate incomes see a similar federal tax bill to a single person, with things actually getting worse at higher incomes for the married couple. Is the UK tax code really that different?
Huh? In the US the married filing jointly tax brackets are exactly double the single tax brackets for every rate except the top 37% rate. A single person making 100k definitely pays a lot more in tax than than a married couple making 100k together. It's generally advantageous to be married filing jointly unless you're at the absolute top 37% rate, at the very bottom (where means tested benefits phase out), or both spouses make roughly equal incomes (in which case MFJ vs two single filers works out around the same).
In your $100K scenario, that single person pays about $6K more in taxes, but has $36K more in take home pay per person, so that additional tax bill seems reasonable in light of their ability to pay it and pay for their cost of living.
You’re moving the goalposts. The point is the tax bill, not “ability to pay per person.” On 100k total, MFJ owes thousands less in federal income tax than a single filer at 100k because the standard deduction and brackets are basically doubled. That’s exactly why marriage is usually a bonus at moderate incomes. The “marriage penalty” shows up mainly when both spouses have high, similar incomes/phaseouts, not in the $100k total case.
Realistically, to train a frontier model you’d need quite a lot of compute. GPT4, which is old news, was supposedly trained on 25,000 A100s.
There’s just no reasonable way of catching modern hardware with old hardware+time/electricity.
reply