Question [Toms] TSMC hiking prices

[jpiniero]

TSMC Hikes Price of Chip Production: CPU & GPU Costs Set to Rise

As demand increases, chip production costs rise.

www.tomshardware.com

Says 10% for 7 nm and below and 20% for 16 nm and above (!) This for wafers delivering starting in January.

Toms source is a Digitimes article so FWIW.

 

[jpiniero]

Samsung and Key Foundry to raise foundry services price by 20%

Samsung and Key Foundry is planning to increase the prices of their foundry services during the second half of the year, TheElec has learned.The pair has notified their customers recently that they are planning to increase wafer prices by between 15% to 20%, people familiar with the matter said.Some

www.thelec.net

This article says Samsung is raising prices 15-20%.

 

[A///]

Except Intel has dumped billions into capex only to get nowhere. TSMC is effectively helping Intel out here when they know they don't have the tech nor the engineers to get there on their own. It'd be smarter for TSMC to let Intel keep floundering in the sand and become hooked on TSMC by funding their expansion not unlike Apple.

 

[DrMrLordX]

Except Intel has dumped billions into capex only to get nowhere.

If that's what Intel would be doing for the next 2-5 years without help from TSMC, it would overall be better for TSMC to just let Intel suffer. The only way TSMC truly benefits is if Intel becomes a permanent vassal. Intel has already shared plans publicly indicating that they won't be vassals starting with the 20a node.

 

[Mopetar]

Without access to TSMC's contract with AMD, it'd be odd if TSMC retroactively increased prices if they were locked in with one of their major partners. And realistically, TSMC would prefer to see Intel suffer more. Most of these TSMC deals you read about were signed a long time ago. I wouldn't be surprised if Apple, for example, signed up for 5nm and maybe 4nm 3-5 years ago.

TSMC doesn't mind Intel running the price up. They don't have to exclusively sell to Intel at any point or have to worry about their other customers all going out of business and then Intel abandoning them.

TSMC knows that those ~$5 Zen chiplets they manufacture are going into $5000 Epyc CPUs. They know AMD can afford to pay a little bit more and some competition from Intel for those wafers will let TSMC charge a little bit more.

TSMC isn't so stupid as to screw themselves over by letting Intel completely cripple their main competitors.

 

[A///]

TSMC doesn't mind Intel running the price up. They don't have to exclusively sell to Intel at any point or have to worry about their other customers all going out of business and then Intel abandoning them.

TSMC knows that those ~$5 Zen chiplets they manufacture are going into $5000 Epyc CPUs. They know AMD can afford to pay a little bit more and some competition from Intel for those wafers will let TSMC charge a little bit more.

TSMC isn't so stupid as to screw themselves over by letting Intel completely cripple their main competitors.

You've simply reiterated what I implied. TSMC isn't looking here to help Intel. If the prices are locked in for a node through a contract, and that's if they are, then they can only raise prices on additional wafers purchased, if possible, should another client fall through the cracks due to constraints. TSMC can also simply charge Intel double the wafer cost and they'd still likely pay because they need access to leading nodes, because their own progress is paltry in comparison to TSMC, and will be for a very long time. Your $5 suggestion is fairly low compared to the real cost involved. It's more around $9-30 depending on generation and progress of the node/economies of scale. As more advanced nodes come out, the cost sink will be slower.

I do not see Intel screwing over or pulling another Core era out of their hat. For all intents and purposes, the stars were aligned for Intel's decade of invincibility because AMD dropped the ball multiple times along with tying up their cash resources in overvalued acquisitions that wouldn't pay off for a long time. Intel's Big.Little may make Intel competitive, but I sincerely doubt they'll ever pull out another whopper like they did.

 

[Doug S]

I do not see Intel screwing over or pulling another Core era out of their hat. For all intents and purposes, the stars were aligned for Intel's decade of invincibility because AMD dropped the ball multiple times along with tying up their cash resources in overvalued acquisitions that wouldn't pay off for a long time. Intel's Big.Little may make Intel competitive, but I sincerely doubt they'll ever pull out another whopper like they did.

While I don't really disagree with your overall premise, you have to be cautious with blanket statements like "I don't see Intel pulling another Core era out of their hat". I see people making predictions all the time that what happened in the past will not happen in the future because reasons.

"There's no way mobile SoCs will ever compete with x86, they don't have the massive resources the Intel x86 monopoly gives them to throw at the problem." Witness M1 Macs proving that wrong (first by Apple helping make the mobile market larger than the PC/server market could dream of, then by launching ARM Macs that easily beat everything Intel offered in the same product classes)

"The era of big increases in IPC is over, the low hanging fruit has already been grabbed." Witness Apple's big IPC jumps in each generation over the past decade.

"We might see big IPC jumps for Apple because they were catching up from behind and designing for a lower clock rate, but we'll never see more than minor improvements for x86 because the design is mature and its complex instruction coding makes it impossible to increase width." Witness the AMD's jumps with Zen and Intel's claimed improvements for its upcoming CPUs.

"Apple needs to make a deal with Intel if they want access to the best process, the foundries are more than a generation behind and often late with their next generation thus falling further behind while Intel delivers a new process every two years like clockwork." Who could have predicted the situation we find ourselves in today when the "Intel will fab Apple SoCs" rumors first appeared a decade ago?

So I'll take any claims that just because Intel has had a rather underwhelming second half of the 2010s decade as far as IPC improvements that they can't get back on the train with "another Core era" if you will. Whether they can depends on the extent to which improved architectures were shelved because they were designed for 10nm and never saw the light of day due to process delays, and Intel had little incentive to push hard because they had no real competition the PC/server marketplace.

There are still plenty of ideas out there to try, from things their colleagues are doing (I'm sure Intel, AMD, Apple and ARM are all taking a close look at what IBM's architects did with Z16's cache architecture) to academic papers detailing new concepts. Now obviously sometimes when you try new things you end up with Core, but other times you end up with Bulldozer, so there are no guarantees.

 

[A///]

While I don't really disagree with your overall premise, you have to be cautious with blanket statements like "I don't see Intel pulling another Core era out of their hat". I see people making predictions all the time that what happened in the past will not happen in the future because reasons.

Because prior to the Core era, Intel was neck-in-neck with AMD. Intel had to constantly work at trying to beat AMD only for AMD to quickly turn around and release a better product. Recall how AMD began and how they evolved. Core was and still is a good uarch. But Core being good was only one part of it. Phenom was the beginning of AMD's problems. It was alright, borderline good, but it wasn't a match for Intel's faster offerings. That is the crux of everything. They simply weren't good enough until Ryzen, and even I'd argue 1st gen Ryzen (1000) was good, very good, but still fell short. When Zen+ came along, the tables began to turn. Though a lot of people would argue that Zen itself, when it launched, put a fire under Intel's bottom. And it did. 5 years later Intel is still struggling to catch up.

Core was a big deal for Intel in the mid 2000s. Core was great, but it was great for many reasons and not simply because it was perfect execution. You have to realize that AMD wasn't competitive by any stretch due to their own internal issues, and again, their overspending. The ATI purchase is almost a catch-22. They overpaid and caused them financial distress when their Core response didn't pan out as they hoped and floundered as they years went by. Of course, had they not, they wouldn't have gotten that console income that kept the lights on at AMD. Or their cards that were on and off again competitive with NVidia.

 

[A///]

"There's no way mobile SoCs will ever compete with x86, they don't have the massive resources the Intel x86 monopoly gives them to throw at the problem." Witness M1 Macs proving that wrong (first by Apple helping make the mobile market larger than the PC/server market could dream of, then by launching ARM Macs that easily beat everything Intel offered in the same product classes)

"The era of big increases in IPC is over, the low hanging fruit has already been grabbed." Witness Apple's big IPC jumps in each generation over the past decade.

"We might see big IPC jumps for Apple because they were catching up from behind and designing for a lower clock rate, but we'll never see more than minor improvements for x86 because the design is mature and its complex instruction coding makes it impossible to increase width." Witness the AMD's jumps with Zen and Intel's claimed improvements for its upcoming CPUs.

"Apple needs to make a deal with Intel if they want access to the best process, the foundries are more than a generation behind and often late with their next generation thus falling further behind while Intel delivers a new process every two years like clockwork." Who could have predicted the situation we find ourselves in today when the "Intel will fab Apple SoCs" rumors first appeared a decade ago?

So I'll take any claims that just because Intel has had a rather underwhelming second half of the 2010s decade as far as IPC improvements that they can't get back on the train with "another Core era" if you will. Whether they can depends on the extent to which improved architectures were shelved because they were designed for 10nm and never saw the light of day due to process delays, and Intel had little incentive to push hard because they had no real competition the PC/server marketplace.

There are still plenty of ideas out there to try, from things their colleagues are doing (I'm sure Intel, AMD, Apple and ARM are all taking a close look at what IBM's architects did with Z16's cache architecture) to academic papers detailing new concepts. Now obviously sometimes when you try new things you end up with Core, but other times you end up with Bulldozer, so there are no guarantees.

People who make claims like that are morons. But the fact is big.little at Intel is still x86, and Lakefield was their first attempt. It's not a bad processor, but it's also not great. Preliminary and dubious data aside, I expect the 12900K to be as good as the 5950X while running hotter and using more power. The rest of your post is littered with fallacy based on not understanding my point. Placing Core on a pedestal of greatness when the competition was in the corner sucking their foot and hitting themselves in the head with a mallet isn't a great look.

I worked at Broadcom for a handful of years. Anyone who thinks ARM isn't a threat to the x86-64 landscape needs a serious dose of reality over and over again until they get it through their skull. Both Intel and AMD need a hardcore reckoning to up the ante and work their little butts off. Frankly, I don't know if x86 is still gonna be around in 20 years or we'll have moved on or hybrid, then again, I doubt I'll be alive in 20 years... LOL

 

[Doug S]

Because prior to the Core era, Intel was neck-in-neck with AMD.

Huh??? They were neck and neck for only 2-3 years, with the Athlon and Opteron, and only because Intel chose the wrong direction with Pentium 4 and was stubborn about admitting their mistake since their marketing department loved the high clock rate to fool consumers.

Before that they were as far behind as they were in the Bulldozer years, Intel was destroying them in the 486/Pentium/PII/PIII years.

 

[Thibsie]

2-3 years ? Athlon/Pentium 3 era was what then ?
Your memory seems a bit lacking.

 

[DrMrLordX]

2-3 years ? Athlon/Pentium 3 era was what then ?
Your memory seems a bit lacking.

Kind of the wrong thread for this discussion, but you are correct. AMD took control of the desktop performance crown with the 1400 MHz Thunderbird-b and only gave it up briefly to the Northwood-b before retaking it with A64. AMD was a dominant force in raw CPU power from 2000/2001 until early 2006.

I am, however, having a little trouble figuring out how this meandering discussion relates to TSMC (and Samsung) raising their rates.

@A///

Your speculation that TSMC could charge double their normal per-wafer rates to Intel is interesting. Do you think that Intel is actually paying that much for N3?

 

[Mopetar]

Phenom was the beginning of AMD's problems. It was alright, borderline good, but it wasn't a match for Intel's faster offerings. That is the crux of everything. They simply weren't good enough until Ryzen, and even I'd argue 1st gen Ryzen (1000) was good, very good, but still fell short.

Until Bulldozer, AMD had good architectures, but the real issue they faced was that Intel had a massive process lead and AMD was basically a node behind Intel. Intel was bigger and could out-invest AMD so that they'd always have a lead.

Zen was a great design, but still on the GF 14nm process. When AMD moved their chiplet production on a 7nm TSMC process that had parity with (or even surpassed) Intel's node it became apparent that they had not just a great architecture but also a process that let it compete.

Intel's problem is that they've stuck with the same foundational design for too long. They can make modifications to it and get improvements, but at a certain point they need a fresh design from the ground up that isn't beholden to limitations of the current design. They just never bothered until recently because they didn't take AMD seriously.

 

[moinmoin]

Intel's problem is that they've stuck with the same foundational design for too long. They can make modifications to it and get improvements, but at a certain point they need a fresh design from the ground up that isn't beholden to limitations of the current design. They just never bothered until recently because they didn't take AMD seriously.

Nor did Intel take TSMC seriously. It apparently never crossed their minds in their market forecasting that a pretty competitive AMD may actually join force with a pretty competitive TSMC.

 

[Zucker2k]

When AMD moved their chiplet production on a 7nm TSMC process that had parity with (or even surpassed) Intel's node it became apparent that they had not just a great architecture but also a process that let it compete.

Nice strawman.

 

[A///]

Until Bulldozer, AMD had good architectures, but the real issue they faced was that Intel had a massive process lead and AMD was basically a node behind Intel. Intel was bigger and could out-invest AMD so that they'd always have a lead.

So basically what I said. Intel having a process lead would only account for so much. You can have a process lead or lead parity with garbage uarch, this has happened before. You can have an older node with an excellent uarch, and this has happened (Intel on 14nm). You can have a junk node with a good uarch, and still get good results, this has happened (NVidia on Samsung for RTX 3000).

Zen was a great design, but still on the GF 14nm process. When AMD moved their chiplet production on a 7nm TSMC process that had parity with (or even surpassed) Intel's node it became apparent that they had not just a great architecture but also a process that let it compete.

Again, one facet of the problem. Intel still being on an older, less robust lead was still capable of keeping up with Zen to a certain point. A fine honed node over several product generations. You forget it took a 2nd generation of 7nm chiplets under Zen 3 uarch to match or beat Intel's gaming performance and some single core activities. Zen 2 blew away Intel when it came to multicore. We all know this.

Intel's problem is that they've stuck with the same foundational design for too long. They can make modifications to it and get improvements, but at a certain point they need a fresh design from the ground up that isn't beholden to limitations of the current design. They just never bothered until recently because they didn't take AMD seriously.

Foundation design is but one of Intel's current problems. This foundation issue is not new. Intel has an incredible habit of being short sighted until they're pulled around by their ears like some Edwardian dark masochistic elf movie. Intel's problem has always been a leadership problem that isn't looking into the future but what to do to ensure their meal ticket bonuses come in. Gelsinger is a refreshing face in a company that has so many problems, but even for him, the time will come where he makes a simple mistake that won't rear its ugly head until years go by.

I'll leave it up to @scineram to downvote this post too because they're obsessed with me.

 

[A///]

Your speculation that TSMC could charge double their normal per-wafer rates to Intel is interesting. Do you think that Intel is actually paying that much for N3?

Pure hyperbole. Outside of TSMC's own announcements, speculation about the company always comes off as a method to pump up their in-country hype and get people to raise the roof, as the kids say. But, hell, if TSMC has the quality goods and Samsung doesn't, what's stopping them from charging what they want to Intel? I don't know of any laws that force TSMC to charge fair and uniform pricing across the board, especially to their competitors.

It's akin to that one person or persons on certain sites who enjoy using the pejorative, "Amdumbs." Intel bought companies' tech to infuse it with their expensive hardware lines. Fair enough, but those don't move enough units to justify the price spent even if the company makes record revenue. Only people who will see that kind of product performance will be research labs, major DCs and maybe some enterprise. Those people don't have product loyalty. Average Joe can't afford those products.

I can't tell you what Alderlake or Intel's DGPU is going to be like, but I'm not expecting amazing stuff like most people are. I couldn't give a dang about Intel or AMD. I neither invest into them nor do I care what I use as long as it's good and affordable. At the end of the day I don't owe any company my loyalty because all they've done is made a product I went out and bought. It only matters if both companies can compete and drive prices down. Even I can agree with you that AMD is getting a little too feisty with their prices, especially given Zen 3's been out for about a year and has sold countless SKUs.

 

[Mopetar]

Nice strawman.

Which part of that was a straw man? Calling something a fallacy doesn't make it one.

It's pretty obvious that AMD would not be able to compete with Intel trying to use a Bulldozer derivative even on TSMC 5nm and that even Zen 4 wouldn't be able to keep up if they were still stuck on GF 12nm.

I'll leave it up to @scineram to downvote this post too because they're obsessed with me.

He does that to me all the time as well so I doubt it's just you.

 

[Zucker2k]

Which part of that was a straw man? Calling something a fallacy doesn't make it one.

It's pretty obvious that AMD would not be able to compete with Intel trying to use a Bulldozer derivative even on TSMC 5nm and that even Zen 4 wouldn't be able to keep up if they were still stuck on GF 12nm.



He does that to me all the time as well so I doubt it's just you.

We all know how OG Zen fared against Skylake. And given given how wide both Zen 2 and 3 got what GF node could've done as excellent a job as TSMC 7? Yes, AMD deserves credit for going chiplet with their design but you can't downplay TSMC's part of the equation, which your post was at pains to do.

 

[DisEnchantment]

The Arm China Debacle and TSMC - Semiwiki

Having spent 40 years in the semiconductor industry, many years working with Arm and even publishing the definitive history book “Mobile Unleashed: The Origin and Evolution of ARM Processors in Our Devices” plus having spent more than 20 years working with China based companies, I found the...

semiwiki.com

Daniel Nenni vented his frustration with the fake news including the price hike.
Very hard to disagree

Pretty much the same as getting accurate political information from Facebook.

https://twitter.com/x/status/1431354133848731655

Price hike, Intel boxing AMD out of N3 and N3 delays. All made up stuff

 

[remsplease]

IMHO, this is valid and in response to hoarding to drive up market prices. TSMC is trying to stay out of the mess and push availability back to normal figures. They (tsmc) know how many cores they're delivering to vendors vs. market sales numbers (historical turn around time metrics from chip to finished product + quantity sold + available stock) and they don't line up.

Simply stated: If you're going to hoard chips to drive up price and throw us under the bus by claiming our delays are to blame, you pay more for that choice. We're staying out of this mess and delivering product.

 

[maddie]

IMHO, this is valid and in response to hoarding to drive up market prices. TSMC is trying to stay out of the mess and push availability back to normal figures. They (tsmc) know how many cores they're delivering to vendors vs. market sales numbers (historical turn around time metrics from chip to finished product + quantity sold + available stock) and they don't line up.

Simply stated: If you're going to hoard chips to drive up price and throw us under the bus by claiming our delays are to blame, you pay more for that choice. We're staying out of this mess and delivering product.

JIT systems breaking worldwide can easily be misinterpreted as hoarding.

 

[DrMrLordX]

IMHO, this is valid and in response to hoarding to drive up market prices. TSMC is trying to stay out of the mess and push availability back to normal figures. They (tsmc) know how many cores they're delivering to vendors vs. market sales numbers (historical turn around time metrics from chip to finished product + quantity sold + available stock) and they don't line up.

Cores or dice? Yeah I guess technically they know how many cores since they aren't ignorant of what they fab for their customers, but still.

Simply stated: If you're going to hoard chips to drive up price and throw us under the bus by claiming our delays are to blame, you pay more for that choice. We're staying out of this mess and delivering product.

Who is hoarding chips fabbed on the bleeding-edge processes we're discussing here? Is anyone hoarding N7, N7P, N7+, N5, or N5P? N6? Honest question.

 

[jpiniero]

TSMC plans 5-9% price hike starting in 2023

Taiwan Semiconductor Manufacturing (TSMC) has notified clients of a 5-9% price hike for both its advanced and mature processes. The new prices are set to go into effect in January 2023.

www.digitimes.com

Another TSMC price hike coming in Jan 2023. 5-9%. So instead of nodes getting cheaper they are getting more expensive.

 

[Doug S]

TSMC plans 5-9% price hike starting in 2023

Taiwan Semiconductor Manufacturing (TSMC) has notified clients of a 5-9% price hike for both its advanced and mature processes. The new prices are set to go into effect in January 2023.

www.digitimes.com

Another TSMC price hike coming in Jan 2023. 5-9%. So instead of nodes getting cheaper they are getting more expensive.

This isn't caused by processes getting more expensive, it is caused by TSMC having more customers than they can serve. They can either make you wait in a long line (except for VIPs like Tim Cook who get waved through) or they can charge more to balance supply and demand.

The real fix is to increase supply but there's a several year lead time to build new fab capacity and outfit it with all the equipment required like EUV scanners. They are doing that too, but in the meantime would you rather pay more for stuff containing TSMC chips and get them right away or be told you'll have to wait six months for delivery?

 

[jpiniero]

or they can charge more to balance supply and demand.

Which means the node is getting more expensive rather than cheaper.