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!
I looked through the most recent Nvidia transcript today. A few points jumped out...
1. An analyst (Ian Ing) stated that TSMC's FinFET was not considered the best out there. This would imply that Samsung's version is currently viewed as the superior process technology. (DISCLAIMER: I don't know if this is true or not...but it was a very interesting point at the end of the conference call).
2. Jen-Hsun didn't deny looking at other foundries but stated that TSMC is still considered to be the "most strategic" & would be a "very, very important partner in the foreseeable future". This would appear to mean that TSMC would continue to be Nvidia's foundry of choice but does open the door for Samsung, GF, UMC, and others.
3. Nvidia is open to licensing their IP. They are in discussions (and litigation) with Samsung & Qualcomm.
4. The laptop GPU market appears to be growing fast & they are in prime position to take advantage of this growth
5. Nvidia's entry into the autos appear to be successful & growing. They highlighted their partnership with Tesla.
6. Jen-Hsun spoke of the importance of mobile technology (not just for mobile phones but IoT) and how a "mobile cloud" can extend the benefits of computing to billions of people.
NOTE: Apologies guys...I got sidetracked. This entry should probably be posted to the main thread instead of the one on process technology. I originally wanted to just highlight what was said about TSMC but just kept adding more & more points
I looked through the most recent Nvidia transcript today. A few points jumped out...
1. An analyst (Ian Ing) stated that TSMC's FinFET was not considered the best out there. This would imply that Samsung's version is currently viewed as the superior process technology. (DISCLAIMER: I don't know if this is true or not...but it was a very interesting point at the end of the conference call).
Was not considered the best out there? Nice wording. Ian is a sharp guy, one of the few analysts that actuall know the semiconductor industry but on this one he is wrong. TSMC 16FF+ is the superior process. Maybe he was talking about TSMC 16nm then I would agree with him.
Any mention of 20nm GPUs? I find it VERY hard to believe that NVIDIA and AMD are skipping 20nm.
Well, at the moment, the "customers" can only compare 16FF vs 14LPE and it appears that 14LPE has an edge (I'm not talking about yield and cost, but only about performance, let's be clear).
16FF+ will perform better than 16FF (and probably also than 14LPE), but then, the fair comparison would be against 14LPP (and in principle, I expect the same gap there).
The question was brought up. Jen-Hsun mentioned that the Tegra X1 would be a 20 nm. He stated that Nvidia was almost moving forward in regards to process technology but he didn't specifically say there would be 20nm GPUs.
[T]he equation is much more complicated than that, and the reason for that is because we have the ability because of the fact that we architect and innovate from architecture all the way to software, we have the ability to improve our performance and efficiency and features on all of those dimensions. We can improve at the architectural level using the same process. We can innovate at the design level using the same process and same architecture. We can innovate at the process level. We can innovate at the software level. We can innovate at the algorithm level. And because we control all of that, we have the ability to move the needle forward in multiple dimensions.
Daniel, it looks like the discrete GPUs are too power hungry for the current 20nm node. Sure, they have tried, but it appears more convenient to move straight to 14nm rather than put additional efforts in developing at 20nm.
Tegra X1 is a mobile SoC, mainly designed for automotive applications: Tegra X1 Super Chip | NVIDIA Tegra | NVIDIA
Daniel, it looks like the discrete GPUs are too power hungry for the current 20nm node. Sure, they have tried, but it appears more convenient to move straight to 14nm rather than put additional efforts in developing at 20nm.
Tegra X1 is a mobile SoC, mainly designed for automotive applications: Tegra X1 Super Chip | NVIDIA Tegra | NVIDIA
I had breakfast with a GPU friend this morning and was told that 20nm was in fact developed for Apple. There is however a higher performance variant for CPU/GPU but clearly it is an SoC process. There was a big change at 20nm with Apple stepping up with a big check while the FPGA and GPU guys were loving 28nm and writing little checks for 20nm.
We live in an SoC world now so at 20nm and below SoCs will be used for process development and ramp instead of FPGAs and GPUs. Look at the recent tone of the comments from Xilinx and NVIDIA when talking about TSMC:
Xilinx: Okay. So there are no issues with TSMC, they have had numerous tape outs already, they are giving us full support. The design whenever you encounter a new generation of product tends to unearth problems that you did not anticipate and as a result the closing of all of these issues is taking a little longer plus the challenges related to design for FinFET transistors are more significant. So it's not a TSMC challenge or issue at all, it's just our ability to finish the design with their support.
NVIDIA:
Well, first of all, TSMC is a fabulous supplier. And number two, their FinFET technology is excellent, and we've been evaluating it. And we -- of course, we do test chips, and our test chip technology is incredibly rigorous, and we take it very, very seriously. This has come from the years of learning that and working with advanced process nodes and fast process ramps. We take it incredibly seriously. So we're working with TSMC on FinFET now for a couple of years, and so we have quite a bit of confidence in their ability to deliver amazing FinFET transistors. I guess with respect to that, we always look to look at all foundries, and TSMC remains our most strategic, of course, and they're going to continue to be a very, very important partner for us for the foreseeable future.
Quite a bit of difference compared to:
Nvidia Blames TSMC's 28nm Process Technology for Slow Sales
Nvidia: TSMC 20nm Essentially Worthless
Bottom line: FPGA and GPU wafer volumes pale in comparison to SoC so they are no longer in the driver's seat. It's an SoC world and we are now SoC girls.... That is a Madonna "material girl" reference by the way for you younger folks.
Hi,
why nvidia is so binded to TSMC? why they do not try other companies do make their chips?
by the way, what are also the conditions that nvidia require to licence their IP?
Hi,
why nvidia is so binded to TSMC? why they do not try other companies do make their chips?
by the way, what are also the conditions that nvidia require to licence their IP?
Nvidia was one of TSMC's first customers and the CEOs became best friends. Nvidia owes their success to TSMC and visa versa. I have seem them appear in public together several times and know this to be true. But business is business and I know Nvidia has done test chips @ Samsung for 14nm but according to the latest comments they will stay at TSMC for 16nm:
