That's the theory but there seems like an obvious problem with it: The x64 desktop and server CPUs are basically the same designs as each other and they're selling plenty of servers, so they're going to keep designing new ones.
Meanwhile people who can't afford new systems because of high prices don't buy them, but that only creates pent up demand while people continue to use their existing systems. Then as soon as RAM prices get back to what they used to be, all the people who have been wanting a new PC can suddenly afford one again and instead of having below average sales they get catch-up demand.
> The theory behind capitalism only works if prices are transparent and goods are transferable.
The theory behind capitalism requires people to take into account what they know when making decisions.
Suppose you have a business where many customers expect to be able to try the product before committing to buy it so the cost of paying for shipping for "free returns" has to be incorporated into the advertised price. Then you notice that a subset of customers have a better idea of what they want and never trouble you with returns, so you want to give them a discount to try to get more of their business.
That's capitalism working the way it's supposed to. The customers who consume fewer resources get to pay lower prices. But it's the thing this bill prohibits, isn't it?
If capatilism was working the way it was supposed to, the customer could choose between paying more up front, but having the option of a "free" return, or paying less upfront byt having to pay for a return (or not be able to return it).
And for that matter, the customer would have enough information to know the quality of the product before purchasing, but that is often not possible.
Everybody knows the cows are not actually spheres. It's about how you deal with it.
If you try to sell "return insurance" then some customers don't buy it but end up wanting to return it anyway and then leave you a bad review for not having free returns. That costs you more than charging somewhat higher prices and having free returns, so that's what you do instead. But now efficiency requires some other mechanism of allowing the people who don't do excessive returns to pay a lower price.
Also, suppose you actually did sell return insurance. Then you notice that a subset of the customers who buy return insurance rarely use it, so you want to give them a discount to try to get more of their business.
The problem is that a small business in Florida or Massachusetts that does 95% of their business in their own state may have no idea that this Colorado law exists until someone sues them over it.
We don't really want small companies to have to start blocking people in other states by default. That's not great for interstate competition.
> This phase is the same thing derivatives markets looked like before the 2008 crisis and Dodd-Frank, and several other waves before that of crisis and reform (Securities Act, Market Reform Act).
Just because a rule was created after something bad happened doesn't mean that the rule is effective to prevent it from happening again. The most common result when they try to ban something without removing the incentive for it to happen is to cause it to happen less obviously. Then the rule (and all its unfortunate costs) gets credited with not observing the bad thing anymore, even though that's not the same as actually preventing it.
Notice that you can use the stock market in the same way as a prediction market. After that healthcare CEO got murdered the company's stock took a hit, as anyone could reasonably have predicted it would. That's a perverse incentive in line with betting that someone will kill the CEO. We don't really have a great way of preventing stock trading from creating that incentive, we mostly just rely on the fact that if you do the murder then murder is very illegal. But if that works for the stock market then why doesn't it work for prediction markets?
> Notice that you can use the stock market in the same way as a prediction market. After that healthcare CEO got murdered the company's stock took a hit, as anyone could reasonably have predicted it would. That's a perverse incentive in line with betting that someone will kill the CEO. We don't really have a great way of preventing stock trading from creating that incentive, we mostly just rely on the fact that if you do the murder then murder is very illegal. But if that works for the stock market then why doesn't it work for prediction markets?
This is true in theory, but in practice the impact of any regular individual's actions on a company is probably going to be small and uncertain enough that it's difficult to make a healthy and reliable profit from. Even the very extreme example of murdering the United Healthcare CEO seems to have caused the stock to drop ~16.5% (assuming the drop is entirely due to the murder). That's like placing a bet with ~1/6 odds. You'd need to short a lot of stock to make that worth the risk of murdering someone (leaving aside any moral issues obviously). You could use leverage to juice those returns but that is expensive and risky, too. If you can afford to deploy enough leverage to make it worth it, you can probably find ways to make money that don't carry a risk of the death penalty.
I guess viewed in this way a bet on a prediction market is like a very cheap, highly leveraged bet on a specific outcome. So the incentives are much stronger as the potential reward for the risk taken is greater.
> You'd need to short a lot of stock to make that worth the risk of murdering someone (leaving aside any moral issues obviously).
When they know exactly when something is going to happen, buying put options that are cheap because they're slightly out of the money seems like it would be pretty effective.
> I guess viewed in this way a bet on a prediction market is like a very cheap, highly leveraged bet on a specific outcome. So the incentives are much stronger as the potential reward for the risk taken is greater.
You seem to be trying to make this about leverage as if that's a thing that isn't available anywhere else.
Let's try another example. Some group breaks into the systems of some publicly traded company and gets access to everything. Now they're in a position to publicly disclose their trade secrets to competitors, publish internal documents that will cause scandals for the company, vaporize the primary and backup systems at the same time, etc. Anything that allows them to place a bet against the company gives them the incentive to do this; the disincentive is that the thing itself is illegal. Leverage gives them a larger incentive, but there are plenty of wages to place a leveraged bet in the stock market.
> When they know exactly when something is going to happen, buying put options that are cheap because they're slightly out of the money seems like it would be pretty effective.
But you don't know exactly what would happen. You know what you will do, but not how it will affect the company's stock price. Maybe it will go down a little, maybe it will go down a lot. Maybe you kill the CEO on the same day as good news is published about the company, which offsets the drop. Or maybe the market just decides the guy wasn't that good a CEO anyway. So you bought a bunch of cheap puts with a strike price of 100, but the stock only drops to 101, and you lose everything. You can buy puts with a higher strike but they will be more expensive.
> Leverage gives them a larger incentive, but there are plenty of wages to place a leveraged bet in the stock market.
Yes, but they are expensive, is my point.
Generally, the disincentive outweighs the incentive. You can increase the incentive through leverage. But that also increases the costs, which increases the disincentive.
There may well be situations where the incentive outweighs the disincentive. But in the context of traditional financial markets I think those situations are likely very rare due to the risks and costs, whereas with a predictions market the risks and costs could be reduced, so it is more likely to happen.
> But you don't know exactly what would happen. You know what you will do, but not how it will affect the company's stock price. Maybe it will go down a little, maybe it will go down a lot. Maybe you kill the CEO on the same day as good news is published about the company, which offsets the drop.
You never know exactly what would happen. You know what you will do, but not if the CEO is going to catch the flu and not show up that day, or have better security than you were expecting, or have a great surgeon, or a spouse who is willing to keep them on life support until after your prediction market contract expires.
> Yes, but they are expensive, is my point.
Only they're not. There are many ways to bet all or nothing on something people generally expect to have a <1% chance of happening, so that you either lose $1000 or make $100,000. Under normal circumstances you could make that bet 100 times in a row and lose $100,000 and the counterparty is happy to take all your money, but if you're able to do something to change the outcome yourself then it's different, which is why it's the same.
> Ethernet and audio have been standard integrated onto the motherboard itself for decades
Unless the one it comes with isn't as fast as the one you want, or they didn't integrate one at all, or you need more than one.
> Across all desktop PCs, the most common number of slots filled is one (a single GPU), and the average is surely less than one (systems using zero slots and relying on integrated graphics must greatly outnumber systems using more than one slot).
There is an advantage in having an empty slot because then you can put something in it.
Your SSD gets full, do you want to buy one which is twice as big and then pay twice as much and screw around transferring everything, or do you want to just add a second one? But then you need an empty slot.
You bought a machine with an iGPU and the CPU is fine but the iGPU isn't cutting it anymore. Easy to add a discrete GPU if you have somewhere to put it.
The time has come to replace your machine. Now you have to transfer your 10TB of junk once. You don't need 100Gbps ethernet 99% of the time, but using the builtin gigabit ethernet for this is more than 24 hours of waiting. A pair of 100Gbps cards cuts that >24 hours down to ~15 minutes. If the old and new machines have an empty slot.
My motherboard has 3 16x PCIe slots, but realistically only one is used for the GPU as the other two are under the mastodon of a cooler needed by the GPU. Can't use a 100G network card if I can't fit it under the GPU. Can't not use the GPU as I don't have an iGPU in my CPU.
He's not advocating from removing PCIe slots, but in practice, it's needed by way less consumers than before. There's probably more computers being sold right now without any PCIe slot than there are with more than 1.
> My motherboard has 3 16x PCIe slots, but realistically only one is used for the GPU as the other two are under the mastodon of a cooler needed by the GPU.
Discrete GPUs generally consume two PCI slots, not three, and even the mATX form factor allows for four PCI slots (ATX is seven), which gives board makers an obvious thing to do. Put one x16 slot at the top and the other(s) lower down and use the space immediately under the top x16 slot for an x1 slot which is less inconvenient to block or M.2 slot which can be used even if there is a GPU hanging over it. This configuration is currently very common.
It also makes sense to put one of the x16 slots at the very bottom because it can either be used for a fast single height card (>1Gb network or storage controller) or a GPU in a chassis with space below the board (e.g. mATX board in ATX chassis) without blocking another slot.
> An Ultra class chip has like 16 memory channels, which even in a 1-DIMM per channel routing would have trace lengths long enough to bottleneck operating frequency.
CAMM fixes this, right?
> Actually, it's a detriment, because price-discovery-enforcing scalpers can rip RAM out of perfectly working computers and resell the RAM. It's way harder to scalp RAM that's soldered on the board.
Scalping isn't a thing unless you were selling below the market price to begin with which, even with the higher prices, Apple isn't doing and would have no real reason to do.
Notice that in real life it only really happens with concert tickets and that's because of scam sandwich that is Ticketmaster.
Ticketmaster is a reputation management company. Their true purpose is to take the reputation hit for charging market value for limited availability event tickets. Artists do not want to take this reputation hit themselves because it impacts their brand too much.
Which is why it's quite appropriate for their reputation to be absolute shit and for members of the public to make sure the stink spreads to anyone who chooses to do business with them as a disincentive to doing it.
Ticketmaster is owned by Live Nation which owns at least 338 major concert venues [1]. Their market power in the venue business allows them to force artists to use Ticketmaster for ticket sales. The artists don't mind though, as they can tell their fans they have no other choice but to use Ticketmaster. Ticketmaster absorbs all of the reputational stink and the artists likely earn more money than they otherwise would have if they were forced to sell tickets at the low prices their fans want.
Except that they don't absorb all of the reputational stink because "Live Nation owns at least 338 major concert venues" is clearly a BS excuse when there are more than 10,000 concert venues in the US, and then the fans still blame the artists for using Ticketmaster.
> A huge reason for this is Apple needs unified memory to keep their money maker (laptops) profitable and performant
None of the things people care about really get much out of "unified memory". GPUs need a lot of memory bandwidth, but CPUs generally don't and it's rare to find something which is memory bandwidth bound on a CPU that doesn't run better on a GPU to begin with. Not having to copy data between the CPU and GPU is nice on paper but again there isn't much in the way of workloads where that was a significant bottleneck.
The "weird" thing Apple is doing is using normal DDR5 with a wider-than-normal memory bus to feed their GPUs instead of using GDDR or HBM. The disadvantage of this is that it has less memory bandwidth than GDDR for the same width of the memory bus. The advantage is that normal RAM costs less than GDDR. Combined with the discrete GPU market using "amount of VRAM" as the big feature for market segmentation, a Mac with >32GB of "VRAM" ended up being interesting even if it only had half as much memory bandwidth, because it still had more than a typical PC iGPU.
The sad part is that DDR5 is the thing that doesn't need to be soldered, unlike GDDR. But then Apple solders it anyway.
> None of the things people care about really get much out of "unified memory". GPUs need a lot of memory bandwidth, but CPUs generally don't and it's rare to find something which is memory bandwidth bound on a CPU that doesn't run better on a GPU to begin with. Not having to copy data between the CPU and GPU is nice on paper but again there isn't much in the way of workloads where that was a significant bottleneck.
the bottleneck in lots of database workloads is memory bandwidth. for example, hash join performance with a build side table that doesn't fit in L2 cache. if you analyze this workload with perf, assuming you have a well written hash join implementation, you will see something like 0.1 instructions per cycle, and the memory bandwidth will be completely maxed out.
similarly, while there have been some attempts at GPU accelerated databases, they have mostly failed exactly because the cost of moving data from the CPU to the GPU is too high to be worth it.
i wish aws and the other cloud providers would offer arm servers with apple m-series levels of memory bandwidth per core, it would be a game changer for analytical databases. i also wish they would offer local NVMe drives with reasonable bandwidth - the current offerings are terrible (https://databasearchitects.blogspot.com/2024/02/ssds-have-be...)
> the bottleneck in lots of database workloads is memory bandwidth.
It can be depending on the operation and the system, but database workloads also tend to run on servers that have significantly more memory bandwidth:
> i wish aws and the other cloud providers would offer arm servers with apple m-series levels of memory bandwidth per core, it would be a game changer for analytical databases.
There are x64 systems with that. Socket SP5 (Epyc) has ~600GB/s per socket and allows two-socket systems, Intel has systems with up to 8 sockets. Apple Silicon maxes out at ~800GB/s (M3 Ultra) with 28-32 cores (20-24 P-cores) and one "socket". If you drop a pair of 8-core CPUs in a dual socket x64 system you would have ~1200GB/s and 16 cores (if you're trying to maximize memory bandwidth per core).
The "problem" is that system would take up the same amount of rack space as the same system configured with 128-core CPUs or similar, so most of the cloud providers will use the higher core count systems for virtual servers, and then they have the same memory bandwidth per socket and correspondingly less per core. You could probably find one that offers the thing you want if you look around (maybe Hetzner dedicated servers?) but you can expect it to be more expensive per core for the same reason.
>The sad part is that DDR5 is the thing that doesn't need to be soldered, unlike GDDR. But then Apple solders it anyway.
Apple needs to solder it because they are attaching it directly to the SOC to
minimize lead length and that is part of how they are able to get that bandwidth.
Except they don't use DDR5. LPDDR5 is always soldered. LPDDR5 requires short point-to-point connections to give you good SI at high speeds and low voltages. To get the same with DDR5 DIMMs, you'd have something physically much bigger, with way worse SI, with higher power, and with higher latency. That would be a much worse solution. GDDR is much higher power, the solution would end up bigger. Plus it's useless for system memory so now you need two memory types. LPDDR5 is the only sensible choice.
CAMM2 is new and most of the PC companies aren't using it yet but it's exactly the sort of thing Apple used to be an early adopter of when they wanted to be.
And it's called "CAMM2" because it's not even the first version. Apple could have been working with the other OEMs on this since 2022 and been among the first to adopt it instead of the last:
> Not having to copy data between the CPU and GPU is nice on paper but again there isn't much in the way of workloads where that was a significant bottleneck.
Isn't that also because that's world we have optimized workloads for?
If the common hardware had unified memory, software would have exploited that I imagine. Hardware and software is in a co-evolutionary loop.
Part of the problem is that there is actually a reason for the distinction, because GPUs need faster memory but faster memory is more expensive, so then it makes sense to have e.g. 8GB of GDDR for the GPU and 32GB of DDR for the CPU, because that costs way less than 40GB of GDDR. So there is an incentive for many systems to exist that do it that way, and therefore a disincentive to write anything that assumes copying between them is free because it would run like trash on too large a proportion of systems even if some large plurality of them had unified memory.
A sensible way of doing this is to use a cache hierarchy. You put e.g. 8GB of expensive GDDR/HBM on the APU package (which can still be upgraded by replacing the APU) and then 32GB of less expensive DDR in slots on the system board. Then you have "unified memory" without needing to buy 40GB of GDDR. The first 8GB is faster and the CPU and GPU both have access to both. It's kind of surprising that this configuration isn't more common. Probably the main thing you'd need is for the APU to have a direct power connector like a GPU so you're not trying to deliver most of a kilowatt through the socket in high end configurations, but that doesn't explain why e.g. there is no 65W CPU + 100W GPU with a bit of GDDR to be put in the existing 170W AM5 socket.
However, even if that was everywhere, it's still doesn't necessarily imply there are a lot of things that could do much with it. You would need something that simultaneously requires more single-thread performance than you can get from a GPU, more parallel computation than you can get from a high-end CPU, and requires a large amount of data to be repeatedly shared between those subsets of the computation. Such things probably exist but it's not obvious that they're very common.
> If you don't think this is right, you are literally going to empty the bank account of my dumb ass grandma who can't stop installing malware, and in every way is better served by a flip phone from the early 2000's.
Then why are you demanding that everyone else's mobile computers have to be locked down instead of demanding that somebody make a mobile phone that only makes phone calls?
Meanwhile people who can't afford new systems because of high prices don't buy them, but that only creates pent up demand while people continue to use their existing systems. Then as soon as RAM prices get back to what they used to be, all the people who have been wanting a new PC can suddenly afford one again and instead of having below average sales they get catch-up demand.
reply