News AMD, Intel, and Nvidia Reportedly Slash Orders with TSMC

[sitehostplus]

GPUs did not cause a recession.


It's not even clear they "shot themselves in the foot", because the RTX 4090 indeed seems to be selling rather well. The reason they would slash future production targets is that they expect to sell fewer GPUs across the entire range, but potentially what it could mean is that they see demand clustering more towards the lower end of the scale, and those GPU dies are smaller.

It's also not clear what you'd have them do instead. I gather you want them to sell the same GPUs for less $$$, but what if they actually can't? If you want GPUs that perform like the RTX 4090 and 4080, they need to be a certain size and we know that new manufacturing nodes are more expensive. That puts them in a certain ballpark on pricing, which limits sales. Lots of people would love it if they sold these at cost or even a loss, but that would be devastating for them.

So, about the only thing they could've done differently is to make the 4080 be the top product. That would've put them at risk of losing the performance crown, but we only know in hindsight that AMD couldn't match it. They couldn't have known that, at the time.

Nvidia's CEO just took a pay cut.

Trouble in team green land. 😉

 

[Endymio]

Auto chips are too simple to make, and do not generate much profit as a result.

So, no, I don't think TSMC makes those at all. More likely, they come from China.

Actually, TSMC not only makes automotive chips, it has entire process nodes devoted specifically to the automotive industry.

Trouble in team green land. 😉

Why would an adult take pleasure in expressing schadenfreude towards a corporation, especially a gamer who is presumably intelligent enough to realize that any downturn for the industry, while it may result in short-term price decreases, is ultimately harmful to supplying them with newer and more powerful GPUs?

 

[InvalidError]

Auto chips are too simple to make, and do not generate much profit as a result.

So, no, I don't think TSMC makes those at all. More likely, they come from China.

Modern cars have hundreds chips in them. If your car has lane assist features, those likely run on a 7nm SoC similar if not the same as what is found in smartphones, tablets, chromebooks, etc. likely fabbed at TSMC. The infotainment / control center likely runs on another similar SoC. Secondary screens likely also have their own SoC to drive them from CAN bus data.

There are pretty good chances your car has dozens of TSMC-fabbed 5-16nm chips in them.

 

[kjfatl]

There are multiple types of chips used in Automobiles. The computing engines used for self-driving, collision avoidance and the like are high end chips made by the likes of TSMC. Think high end GPU based system that makes your gaming PC look simple.

There are small microcontroller that might do something like control everything in a door. These can be made using 10 year old technology.

There are 'simple' chips that control power, but must handle tens or hundreds of amps of current at hundreds of volts. Often these are made using Galium Nitrite or another non-silicon material. Battery powered cars use many of these.

 

[sitehostplus]

Why would an adult take pleasure in expressing schadenfreude towards a corporation, especially a gamer who is presumably intelligent enough to realize that any downturn for the industry, while it may result in short-term price decreases, is ultimately harmful to supplying them with newer and more powerful GPUs?

Now why would nVidia lower prices, when all they have to do is cut production?

And no, I don't take pleasure in this (unless I want to buy their stock for cheap) it's just reinforcing the position I've had all along that nVidia isn't the profit-making juggernaut people think it is.

 

[InvalidError]

Now why would nVidia lower prices, when all they have to do is cut production?

Profit = unit margins * volume - costs

In a healthy market, companies would aim for the optimal volume-margin point to maximize profit. Also, the more units you amortize your R&D and other costs over, the leaner those costs become.

Then there is also the long-term cost of alienating your market: if too many people bail out of PC gaming because there are no decent GPUs within their price range, that is that many fewer potential customers to sell stuff to in the future.

Also, everyone and their dog is jumping into the AI accelerator pool. Nvidia may not have many years left of being the most notable source of AI acceleration hardware. What will Nvidia do in 2-3 years when companies ditch Nvidia in favor of other AI accelerators capable of doing the same jobs for 1/10th the price?

Nvidia may be about to 3dfx itself.

 

[Deleted member 14196]

Wow, I believe that Nvidia is doing just that as well—reminding me of 3dFX

 

[sitehostplus]

Profit = unit margins * volume - costs

In a healthy market, companies would aim for the optimal volume-margin point to maximize profit. Also, the more units you amortize your R&D and other costs over, the leaner those costs become.

Then there is also the long-term cost of alienating your market: if too many people bail out of PC gaming because there are no decent GPUs within their price range, that is that many fewer potential customers to sell stuff to in the future.

Also, everyone and their dog is jumping into the AI accelerator pool. Nvidia may not have many years left of being the most notable source of AI acceleration hardware. What will Nvidia do in 2-3 years when companies ditch Nvidia in favor of other AI accelerators capable of doing the same jobs for 1/10th the price?

Nvidia may be about to 3dfx itself.

If you cut production, you can clear out unsold inventory without lowering prices, plus net yourself profits in the quarters ahead.

The only time cutting the price makes sense, is if a newer model of your product is about to hit the shelves.

I don't get the 3dfx comparison. The thing that killed them was ending all of their partnerships and making their own graphics cards out of Mexico. Is nVidia doing that kind of stupid move?

If anything, there is a better comparison: Atari. At one point, the 2600 machines were so hot you needed ove gloves to handle them. The next year, they were burying game cartridges in landfills because demand was so soft.

They have to keep innovating, or that will happen to them too.

 

[InvalidError]

If you cut production, you can clear out unsold inventory without lowering prices, plus net yourself profits in the quarters ahead.

If people aren't buying your GPUs at current prices and you cut production, your old stock still isn't selling much and that is how you get at 64% sales decline like the one in Nvidia's last financial statement

I don't get the 3dfx comparison. The thing that killed them was ending all of their partnerships and making their own graphics cards out of Mexico. Is nVidia doing that kind of stupid move?

Nvidia has its FE cards which undercut pretty much everyone else. It makes its partners buy parts from it before anyone know what it intends to set the MSRP at, which tends to screw AIBs over when the prices are lower than initially thought. EVGA bailed out of its Nvidia partnership because of all the BS Nvidia puts AIBs through and I bet there will be more in the future as Nvidia forces AIBs to make boards they effectively cannot sell.

 

[Endymio]

Nvidia makes its partners buy parts from it before anyone know what it intends to set the MSRP at... EVGA bailed out of its Nvidia partnership because of all the BS Nvidia puts AIBs through...

This just isn't correct at all; it's not even what EVGA claimed. EVGA stated NVidia was publicly announcing MSRPs before privately informing boardmakers of the wholesale costs of chips -- not making them purchase those chips, then announcing a price.

Furthermore, it's extremely telling that EVGA didn't dump NVidia for AMD or Intel -- they just exited the videocard market entirely, a tacit admission they couldn't turn a profit with any vendor.

 

[InvalidError]

This just isn't correct at all; it's not even what EVGA claimed. EVGA stated NVidia was publicly announcing MSRPs before privately informing boardmakers of the wholesale costs of chips -- not making them purchase those chips, then announcing a price.

There have been instances where Nvdia didn't announce the MSRP before launch or changed it at the last minute where boards are already assembled and shipped out to retailers. Then you have more recent history where boards simply aren't selling at MSRP due to how horrible their value-per-dollar is with AIBs having next to no margin to drop prices on and Nvidia refusing to budge on pricing or rebates on its end.

Next week, we'll have Nvidia attempting to foist 8GB GPUs onto the public for ~$400 when 8GB doesn't belong on anything over $250 IMO. AMD is slightly less bad with tentative pricing being $300 on the RX7600.

Kind of hard for AIBs to make ends meet when they are being over-charged 50+% by AMD and Nvidia at a time where most buyers aren't buying into their over-inflated prices anymore.

 

[Deleted member 14196]

Like the Fonzie used to say, “sit on it Potsie”

They will sit on unsold hardware until they lower prices. There’s no magic and this is how it works every time. Jeep is just learning this again for the millionth time with their huge levels of inventories, they’re begging dealers to take cars off their hands, but they’re not reducing prices so they’re not selling. Seems so simple. Add an average price of $60,000 normal people just can’t afford that

I stopped buying GPU a long time ago because of prices. Looks like everyone else is starting to do the same. Now all I buy are integrated graphics and I game on the console

 

[Endymio]

There have been instances where Nvdia didn't announce the MSRP before launch ...Then you have more recent history where boards simply aren't selling at MSRP due to how horrible their value-per-dollar is with AIBs having next to no margin to drop prices...Next week, we'll have Nvidia attempting to foist 8GB GPUs onto the public...Kind of hard for AIBs to make ends meet when they are being over-charged 50+%

You're laboring under several misconceptions.

1. The "s" in MSRP stands for suggested. NVidia suggests a price, but AIB makers are free to charge whatever they wish. More importantly, the entire MSRP concept exists not to benefit the manufacturer, or even the consumer, but the middle-man vendor (in this case board-makers). And if NVidia was mistreating board makers so badly, they'd all be fleeing to Intel. Do you see that happening?
2. Board memory is at the discretion of the board maker. The 16GB version of the 4060 is already announced, and there's nothing to stop a boardmaker from selling a 12GB version as well. However, I'll bet you dollars to doughnuts their $299 8GB version (not $399 as you claimed) will be their best seller, and will fly off store shelves like hotcakes.
3. The performance value-per-dollar of current boards is the best it's ever been in history. Sales have declined because (a) Covid's over, and (b) nearly everyone who wants a board at these prices has already purchased one. It'd be nice to be back in the Moore's Law world of 2015 or even 2005 ... but that's not happening.
4. No business in the world will choose to manufacture products at a loss. NVidia could reduce its profit margin to zero, and board prices wouldn't drop by anywhere near 50%. That situation won't change until TSMC adjusts its own rates. Your anger is misdirected.

 

[InvalidError]

4. No business in the world will choose to manufacture products at a loss. NVidia could reduce its profit margin to zero, and board prices wouldn't drop by anywhere near 50%. That situation won't change until TSMC adjusts its own rates. Your anger is misdirected.

TSMC's prices have very little to do with it: 5N costs about $25 per 100sqmm, so a 200sqmm chip only costs about $50. That only justifies a ~$30 price increase over the GTX1050Ti and the extra 4GB may justify another $15. Anything above that is entirely AMD and Nvidia greed.

 

[Endymio]

TSMC's prices have very little to do with it: 5N costs about $25 per 100sqmm, so a 200sqmm chip only costs about $50.

You've made several errors. First, you appear to be using a widely-reported wafer cost of $17K from 2020 -- but TSMC upped prices dramatically since then. You're also assuming zero waste and zero failure rate. And wafer cost alone is far from the total cost of a chip: it must be sliced, packaged, and tested.

Taking all these into account, AD106 likely has a net cost of around $100/chip. Given NVidia's current gross margins of 65%, it's selling these to AIB makers for $165 each, or slightly over half the board's MSRP. Now, you can complain about that $65 markup if you wish-- but chips don't design themselves. NVidia's net margins are at 16% now, which works out to all of $17.00 for each of those 4060s.

 

[InvalidError]

You've made several errors. First, you appear to be using a widely-reported wafer cost of $17K from 2020 -- but TSMC upped prices dramatically since then.

Not really: in November 2022, N3 was quoted at around 20k$ per wafer. Lack of demand does wonderful things for pricing stability.

 

[helper800]

You're laboring under several misconceptions.

2. Board memory is at the discretion of the board maker. The 16GB version of the 4060 is already announced, and there's nothing to stop a boardmaker from selling a 12GB version as well. However, I'll bet you dollars to doughnuts their $299 8GB version (not $399 as you claimed) will be their best seller, and will fly off store shelves like hotcakes.

This is incorrect. Nvidia does not allow AIBs to add or remove memory from their initial board designs. If different memory layouts are released it is at Nvidia's allowance, not by reason of creativity on the AIBs behalf.

 

[Endymio]

Not really: in November 2022, N3 was quoted at around 20k$ per wafer.

That's the figure I used for my estimate of $100 per 200 mm^2 chip. If we use your estimate of $50, however, it's even worse for your argument, as it reduces NVidia's gross markup to $32/chip -- less, even, since their margins on consumer gpus are substantially lower than data center or visualization.

This is incorrect. Nvidia does not allow AIBs to add or remove memory from their initial board designs.

It's my understanding that an AIB maker is free to alter the reference design in any manner they choose, subject to the following limitations:

-- they cannot reduce the base specifications for any given model
-- they may not create their own model names or trademarks, i.e they can't sell an "RTX 4075", no matter how much they beef up a 4070.

If you have a reference that says otherwise, I'd enjoy seeing it.

 

[helper800]

It's my understanding that an AIB maker is free to alter the reference design in any manner they choose, subject to the following limitations:

-- they cannot reduce the base specifications for any given model
-- they may not create their own model names or trademarks, i.e they can't sell an "RTX 4075", no matter how much they beef up a 4070.

If you have a reference that says otherwise, I'd enjoy seeing it.

Of course I cannot find the source I read, however, paraphrasing, "AIBs need Nvidia approval to change performance aspects of the graphics cards such as, adding a second GPU to a board, increasing vRAM capacity above reference board specifications, and ..." This means they cannot sell a 48gb version of the 4090, for instance, or a 20gb version of the 3080 without Nvidia's express approval. This is for a couple reasons I can think of. For example, Nvidia’s RTX 4060 Ti card has two variants: one with 8GB of GDDR6 memory and one with 16GB of GDDR6 memory. Both cards have a 128-bit memory bus, which means they can support either 4 or 8 memory chips. AIBs can choose which variant to produce and sell, but they cannot go beyond 16GB of memory without changing the memory bus width or getting denser memory chiplets, which do not exist. Another thing to note is the the 8GB and 16GB versions of the 4060 Ti cards have different GPUs technically. The 8GB version is called the AD106-300 and the 16GB version is the AD106-400. Meaning they are technically not just the same GPU with more or less vRAM.