News TSMC Might Cut 3nm Prices to Lure AMD, Nvidia

[Admin]

TSMC is reportedly considering lower quotes for 3nm production to stimulate interest from chip designers such as AMD and Nvidia.

TSMC Might Cut 3nm Prices to Lure AMD, Nvidia : Read more

 

[cyrusfox]

Sounds like TSMC was overly aggressive on price increases as they are now not able to fill new top node by apple alone. Priced out everyone else.

As N3 is the first new node where you don't get all the same scaling benefits. We are approaching diminishing returns of scaling, coupled with competition from Intel Foundry, as well as Samsung, Foundry market looks to be heating up! Too bad Global bowed out of this competition.

 

[InvalidError]

With AMD having to drop prices on Zen 4 due to low sales, Nvidia's 4000-series not selling particularly well due to exorbitant prices and AMD presumably not faring a whole lot better based on how the 7900XT has consistently been in stock since launch, I can't imagine anyone being in a rush to sign major commitments for N3/N3E.

Interesting times ahead as the industry adjusts to grossly overshooting the consumers' spending limits.

 

[Deleted member 14196]

they obviously have no idea how economics works. Skyrocketing, inflation and peoples wages, stagnating or going down. Etc

 

[DavidLejdar]

The disposable income varies among countries though. E.g. in Germany inflation exists as well, but last year the minimum wage was increased by more than 10% (without the economy crashing, and with jobs still available), and rents are way lower than e.g. in London. So here I sit in Berlin, with an income below the local median, and I am like: "Wait, what? 3nm-based GPUs are not in production already? Lame! First these companies get me all hyped about the newest technology and stuff, and now all they want to sell me is like some last-gen pen and paper? What a let-down!".

I mean, e.g. a 4090 would be quite an expense for me. But because it is just an expense, I can not afford to purchase such a GPU every year, and why then go for some technologically last-gen stuff?

Joking aside though, I think what the GPU makers have been somewhat overestimating is the interest for 4K gaming. Many gamers do not want that, or do not care about it, or simply upscale (as consoles do), possibly thinking that they have full 4K already. And in such a context, offering a 4-nm based GPU primarily designed for 1080p gaming, and one for 1440p, and reasonably priced of course, that may arguably drive sales a bit more than just trying to sell what is way over the top for many a gamer.

 

[PlaneInTheSky]

This TSMC/Samsung foundry duopoly really has to stop.

I am not buying any GPU until prices are halved.

 

[Co BIY]

This TSMC/Samsung foundry duopoly really has to stop.

I am not buying any GPU until prices are halved.

The major "duopolist" having to drop prices just to secure sufficient utilization of its enormously expensive fab investments does not sound like market conditions that would entice others to join!

But Intel is jumping in anyway ...

 

[Co BIY]

they obviously have no idea how economics works. Skyrocketing, inflation and peoples wages, stagnating or going down. Etc

Pretty sure they understand it - They are lowering prices in response to reduced demand in order to most efficiently use the fabs and processes they build though years of investment.

That they are overbuilt for this particular moment on this node is not some foolish mistake. It takes years to develop these kinds of plans. They don't stop or turn on a dime. I don't think their major competitors are in better shape

TSMC may be in for rough times - 45% profit instead of 60%. They are running a Sale when others are performing mass layoffs.

 

[sweepyjoelschl]

So Question: who exactly bears the brunt of cost for the low yields? Does Apple pay full price for each wafer even if only 60% is good? or does TSMC have to issue "make good refunds". If it is the former then what company in their right mind want to sign up until they get the process snafus worked out and yields in very high percentages? At $20,000 per wafer I would not bewilling to fork that out if half of the dies end up bad and unusable

 

[digitalgriffin]

So Question: who exactly bears the brunt of cost for the low yields? Does Apple pay full price for each wafer even if only 60% is good? or does TSMC have to issue "make good refunds". If it is the former then what company in their right mind want to sign up until they get the process snafus worked out and yields in very high percentages? At $20,000 per wafer I would not bewilling to fork that out if half of the dies end up bad and unusable

From what I heard, if TSMC makes a claim "for a given XXX sq mm chip running our yield rate is Y% +- Z%" that has to be written into the contract. The client would have to prove it's a true defect vs a design problem.

It is a lot more complicated complicated than that because of binning saying "v% will have 1 defect. Or u% to will have 2 defects." But you get the idea. If you design a chip where you allow two from national units to be defective you might have close to 100% yield if it's a small enough chip. (Think 3600x vs 3800x)

Calculating yield stats with Poisons distribution is something they teach you in second year comp sci.

 

[digitalgriffin]

TSMC did and did not get greedy.

TSMC was running full capacity during the beer virus. They were also facing supply constraints. So TSMC invested heavily into new fabs for future demand.

But cost of new fabs is exponentially growing. And the number of steps it takes to make a finished wafer is obscene now. So yes it is costing them more.

But TSMC wanted considerably more than organic growth due to all the demand. So they charged more with higher margins during high demand to pay for that growth. Now it is biting them in the tail.

 

[PlaneInTheSky]

It will not be that we can't technically get smaller nodes, but that consumers will no longer be willing to pay the cost to do so.

There is a point where the cost/benefit analysis for consumers no longer makes any sense.

TSMC builds fabs with money Apple, AMD and Nvidia gives them. And consumers are supposed to pay Apple, AMD and Nvidia in return.

But the moment consumers have had enough of paying for $800-$2,000 GPU and $1,000 iPhones, TSMC goes down.

Imo we are at an inflection point where consumers are no longer willing to foot TSMC's expensive bill.

 

[-Fran-]

Given how many concessions/incentives Taiwan gave them, this was expected to happen, I guess?

I just hope Samsung can compete and, later down the line, Intel. Although I'm not sure AMD or nViia would be too comfortable using Intel foundries. Plus, I doubt Intel would open their latest nodes (i.e: most advanced) to their direct competitors.

I doubt this will materialize in lower prices on consumer devices using their 3nm node anyway. At least, not in any significant manner. I hope I'm wrong.

Regards.

 

[InvalidError]

I doubt this will materialize in lower prices on consumer devices using their 3nm node anyway. At least, not in any significant manner. I hope I'm wrong.

The whole point of Intel opening up its fabs for hire is to immunize itself from having under-utilized 20+G$ investments due to lulls in demand for its own chips. I doubt Intel can afford to be choosy about which credible clients it will lend its fabs' excess capacity to. I doubt Intel would get many foundry clients if clients know they may get arbitrarily denied access for encroaching on Intel's in-house product turf.

 

[Endymio]

It will not be that we can't technically get smaller nodes, but that consumers will no longer be willing to pay the cost to do so.

The problem is that N3E isn't a full node jump. It's not even a half-node: 10%-50% increase in logic density, essentially no increase in SRAM density, and a per-transistor cost that, depending on the specific process used, may actually be higher than N5. Amortize in the design costs of switching to that node, and it's just not that attractive ... not without a substantial price cut.

 

[InvalidError]

It will not be that we can't technically get smaller nodes, but that consumers will no longer be willing to pay the cost to do so.

When transistors, traces and other structures are mere atoms thick or wide, there isn't much room to go any smaller where any single atom to be misaligned, wrong isotope, trace contaminant, etc. At 3nm actual physical feature size, a single out-of-place atom is ~11% of your total silicon atoms in that dimension.

 

[GasLighterHavoc]

they obviously have no idea how economics works. Skyrocketing, inflation and peoples wages, stagnating or going down. Etc

They obviously do understand how economics work since they are raising prices when everyone is desperate to buy from them and cutting prices when demand slackens.

No business cuts prices proactively until the situation forces them to do so.

 

[GasLighterHavoc]

When transistors, traces and other structures are mere atoms thick or wide, there isn't much room to go any smaller where any single atom to be misaligned, wrong isotope, trace contaminant, etc. At 3nm actual physical feature size, a single out-of-place atom is ~11% of your total silicon atoms in that dimension.

But consider that all the fab companies no longer accurately report transistor size in their node names. Is any transistor actually 3 nm physically at the 3 nm node? I don't think so.

For example, from the 60s to the 90s, the gate length is what defined a new node. The last time the gate length matched the node name was the 350 nm node which had a gate length of 350 nm all the way back in 1995! The half-pitch is another measure of a node. The last time the half-pitch matched the node name was in 2004 with the 90 nm node (coincidentally around the time that Dennard scaling died).

Your overall idea is right in that sub-elements of transistors are small enough for single atoms to make a large difference. But nothing in the 3 nm node is actually 3 nm in size. It is a purely commercial name at this point.

 

[digitalgriffin]

The whole point of Intel opening up its fabs for hire is to immunize itself from having under-utilized 20+G$ investments due to lulls in demand for its own chips. I doubt Intel can afford to be choosy about which credible clients it will lend its fabs' excess capacity to. I doubt Intel would get many foundry clients if clients know they may get arbitrarily denied access for encroaching on Intel's in-house product turf.

Intel won't get choosy, but you can bet the bank they won't let them use the latest node. I think Fran is correctly inferring that to keep dominance Intel will keep the fastest nodes for themselves. It would be like Tesla allowing Ford to build a Mustang Mach E at their factory when they are capacity as it is.

This is to bankroll older fabs that lost relevance. GloFo still has significant business for their 12nm node.

 

[hannibal]

If we want to have decent priced GPU we need to go back to 12nm... Of course those GOU would be slower also...
But in reality it is hart to see GPU prices to come down unless they start to make much smaller GPU chips aka less powerful chips. If TSMC reduce the price of 3nm, it still is much more expensive than 6nm or above... Or even 5nm that Nvidia is using now. That will mean than GPU made in 3nm will be more expensive than Nvidias current 40000 series...

 

[InvalidError]

If we want to have decent priced GPU we need to go back to 12nm... Of course those GOU would be slower also...

7nm EUV will probably be the sweet spot for long-term bang-per-buck chips: although EUV equipment may be more expensive, EUV eliminates many of the multi-patterning steps required to make sub-22nm DUV chips, which saves time, labor, consumables and improves yields.

 

[hannibal]

7nm EUV will probably be the sweet spot for long-term bang-per-buck chips: although EUV equipment may be more expensive, EUV eliminates many of the multi-patterning steps required to make sub-22nm DUV chips, which saves time, labor, consumables and improves yields.

True 7nm is rather good node or 6nm variant that AMD use in their Zen4 memory controller chips. Both have excellent yields and are not the most pricey options in the market.