You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
According to a new report today, Google is indeed working to replace Samsung with TSMC for Tensor G5 development...
9to5google.com
According to a new report today, Google is indeed working to replace Samsung with TSMC for Tensor G5 development.
The Information details how the plan was originally to unveil its “first fully customized chip” for Pixel phones in 2024. Instead, “Redondo” missed the trial production deadline last year after already cutting features, with a handoff to Taiwan Semiconductor Manufacturing Company only occurring “earlier this year.” Today’s report says “it won’t be ready in time for mass production by 2024,” with Redondo now being used as a test chip ahead of the next generation.
Laguna, with these chips having a beach motif, is targeted for 2025 and likely be branded as the Tensor G5. According to today’s sources, Tensor G5 will be based on TSMC’s 3-nanometer manufacturing process and Integrated Fan-Out for reduced thickness and increased power efficiency.
Google is sticking with Samsung for next year’s Tensor chip design and manufacturing, but has been replacing more and more Samsung components (“everything from communications and audio to image and graphics processing”) with its own IP every generation.
Meanwhile, “multiple Tensor chips” have been canceled over the past two years. That comes as a Tensor-powered Pixelbook was rumored but canceled last year. Notably:
Google’s delay in getting fully custom Tensors to market partly stems from challenges dividing and coordinating work between the U.S. and India, where the majority of its Tensor silicon engineers are based, along with high turnover inside the group, according to a former Google chip executive with direct knowledge of the effort.
Meanwhile, that former executive also told The Information that they were bearish on how much Google is spending on custom chips given that Pixel is not yet selling in volume.
A former Google executive making negative remarks to the press about secret Google development programs. I'd be wary about the objectiveness of the information.
With all of the little guys leaving Samsung for TSMC (Google Tensor, Tesla FSD), Samsung should get more aggressive in pursuing customers. Like a special discount program. Stay with us for 2 more years, and everything will be half off. Masks, wafers, packaging, etc...
With all of the little guys leaving Samsung for TSMC (Google Tensor, Tesla FSD), Samsung should get more aggressive in pursuing customers. Like a special discount program. Stay with us for 2 more years, and everything will be half off. Masks, wafers, packaging, etc...
Samsung is already hemorrhaging money compared to last year. They are barely profitable as is. Also this is not feasible as a strategy because those at the leading edge and ai arms race are not going to go to a clearly inferior process in every way just to save a couple bucks while their competitors are on TSMC. You cut elsewhere not cheaping out on your chip suppliers
Samsung is already hemorrhaging money compared to last year. They are barely profitable as is. Also this is not feasible as a strategy because those at the leading edge and ai arms race are not going to go to a clearly inferior process in every way just to save a couple bucks while their competitors are on TSMC. You cut elsewhere not cheaping out on your chip suppliers
Barely profitable yes. Hemorrhaging money, no. They have yet to lose money. (they have in the past but that was a long time ago). This is just a temporary situation given the memory glut. It will pass. They have been making money hand over fist for so long that they probably have a huge warchest saved up.
Their process may be inferior, but it's all about yields. Throw out the bad chips and pick out the high performing ones. At a low yield, this is very expensive, and would not be cost effective normally. But they have money to burn.
Throw in some V-NAND. You don't want V-NAND? Okay, take some DRAM instead. Everyone wants HBM. We'll drown you in it if needed. Boom. Customer locked in. Press release. Let's go, next...
With all of the little guys leaving Samsung for TSMC (Google Tensor, Tesla FSD), Samsung should get more aggressive in pursuing customers. Like a special discount program. Stay with us for 2 more years, and everything will be half off. Masks, wafers, packaging, etc...
QCOM and NVDA left Samsung as well. TSMC is getting more aggressive in regards to customer retention (packaging etc...) so the foundry business will be much more competitive. Intel Foundry is starting from the ground up and so is Samsung.
2nm will be a much more complex node, unlike FinFETs which are an open standard. It will be much more difficult for foundries to make PDKs and for customers to straddle foundries.
3nm will be a long node and it is very good timing for TSMC to have a dominant process, absolutely.
This is a very reductive and inaccurate assessment of how the foundry business works. I’m not even sure where to start. Just offering random memory to fabless designers has no relevancy to picking a process. It’s just not a thing. So many other things come first. PDK, collaborative ecosystem, reliability, scale. Price is of course a factor but it’s not enough. I’ll let Daniel and others handle this one lol
Barely profitable yes. Hemorrhaging money, no. They have yet to lose money. (they have in the past but that was a long time ago). This is just a temporary situation given the memory glut. It will pass. They have been making money hand over fist for so long that they probably have a huge warchest saved up.
Their process may be inferior, but it's all about yields. Throw out the bad chips and pick out the high performing ones. At a low yield, this is very expensive, and would not be cost effective normally. But they have money to burn.
Throw in some V-NAND. You don't want V-NAND? Okay, take some DRAM instead. Everyone wants HBM. We'll drown you in it if needed. Boom. Customer locked in. Press release. Let's go, next...
Barely profitable yes. Hemorrhaging money, no. They have yet to lose money. (they have in the past but that was a long time ago). This is just a temporary situation given the memory glut. It will pass. They have been making money hand over fist for so long that they probably have a huge warchest saved up.
Their process may be inferior, but it's all about yields. Throw out the bad chips and pick out the high performing ones. At a low yield, this is very expensive, and would not be cost effective normally. But they have money to burn.
Throw in some V-NAND. You don't want V-NAND? Okay, take some DRAM instead. Everyone wants HBM. We'll drown you in it if needed. Boom. Customer locked in. Press release. Let's go, next...
I believe Samsung already did several things you suggested but achieved limited success. One problem is companies, like Qualcomm or Nvidia, cost is important but without getting enough quantity of chips with good quality at the right time can kill them.
Samsung is already hemorrhaging money compared to last year. They are barely profitable as is. Also this is not feasible as a strategy because those at the leading edge and ai arms race are not going to go to a clearly inferior process in every way just to save a couple bucks while their competitors are on TSMC. You cut elsewhere not cheaping out on your chip suppliers
Barely profitable yes. Hemorrhaging money, no. They have yet to lose money. (they have in the past but that was a long time ago). This is just a temporary situation given the memory glut. It will pass. They have been making money hand over fist for so long that they probably have a huge warchest saved up.
Their process may be inferior, but it's all about yields. Throw out the bad chips and pick out the high performing ones. At a low yield, this is very expensive, and would not be cost effective normally. But they have money to burn.
Throw in some V-NAND. You don't want V-NAND? Okay, take some DRAM instead. Everyone wants HBM. We'll drown you in it if needed. Boom. Customer locked in. Press release. Let's go, next...
Just sharing data. Samsung LSI division reported loss of 0.51tn for Jun 23 qtr compared to loss of 0.27tn Won for Mar qtr. EBIT margin at June qtr stood at negative 9%. At Samsung corporate level it does make good money on handsets, OLED. Memory has been the dominant cashflow which currently is suffering losses.
9to5google.com
