Nvidia Turing Speculation thread [2018]

Several people stated that this "long time" phrase was actually an attempt to shut down the hordes of journalists asking the same question repeatably, as no company is stupid enough to announce their next gen product is a long time away! especially not NVIDIA.

 

lying hurts bussines. So i, otoh, say they aren't that stupid to do it.

 

Can you define a long time?

Its a vague term that different people take to mean different things.

 

Yup, for me, a long time is 1 year or more.

However, my nephew last night let me know that 2 hours is a VERY long time.

Regards,
SB

 

Obviously vague. I already stated my personal definition in the context. >3 months.
The context being that of a corporate entity

 

You also have the semantics between actual launch (still vague to have several interpretations), the review window before that, and potentially a soft launch and announcement, looking at each individually it changes the context of when.

 

Or, entirely plausibly, they've given up 12nm as a bad node to put a year+ worth of products into when AMD seems adamant in getting out products on the much, much better "7nm" node. Little sense, perhaps, in investing in an entire product lineup that could be outdated within less than a year. Especially when Nvidia is already on top.

Why not wait till sometime next year, when TSMC's (already ahead of everyone else) 7nm should be mature enough, and have enough production capacity, to meet large scale large die manufacturing. AMD already announced its (no doubt very limited and expensive) TSMC 7nm card would be shipping by the end of the year. What's good news for them is good news for Nvidia too, as far as scheduling is concerned.

 

A primary driver will be whether they feel they can sustain GP102 and GP104 for Inferencing rather than going to Tensor Gx102 and Gx104 version, along separately with what solutions are required for datacenters/clouds.
There is a lot of competition out there that is beyond Geforce gaming, and like I mentioned earlier it is very difficult to truly split Geforce/Quadro and Tesla (apart from possibly node size and one could expect Tesla at 7nm), as can be seen that no Tesla models are yet announced apart from V100 (Tesla can be announced-presented months in advance before launch) and V100 does not fit all the Tesla segment requirements/deployments; I can see though you could argue this point fits your post but would be more Geforce, also need to consider the HotChips schedule recent change due to high profile visibility.
But yeah I agree Nvidia with the lateness of the cycle are now entering closer to the phase of 7nm for all models and not just certain Tesla, but they also have their obligated contracts for GDDR6 for this year (probably not the only supply chain commitment) so from a logistics/line process it can become a headache; there is a logistical formula used for the ideal logistics-manufacturing line efficiency by quite a few of the larger international tech companies.
I think multiple factors have caused this delay from technical-maturity to BOMs (GDDR6 fitting in both as one example and maybe GDDR5X was a misstep) to situations such as mining craze that will cause a business drop (no point selling new model into the mining craze and better yet when sales start to dry up from a business growth narrative), then from a Tesla side it is getting the clear narrative on their recent announcements and steps for their high margin even by Nvidia standards all-in-one solutions.

But as you say they are in the situation where they may had to change their strategy very late in the day.
Beyond your point/context that are all valid I would say it would be very difficult to conclude/predict either way with what has been said publicly so far.

 

I think it's wrong. GF was never mentioned as far as i know and gf themself said they are a bit behind. This year no other factory than TSMC can deliver 7nm products. Epyc is also 7nm TSMC because TSMC is ready first.

 

I think that might be underplaying the 1st stage CLN7FF from TSMC, remember it is taped out by quite a number of tech manufacturers, and it is being compared as a 1st stage replacement to 16FF+ by TSMC and others.
The issue may be more to do with manufacturing/scaling costs-performance due to DUV/multipatterning but still viable as stage 1, albeit a move to EUVL next year still has some performance/current/scaling improvements over the CLN7FF.
Separately a recentish article at EEtimes.

May 2nd article: https://www.eetimes.com/document.asp?doc_id=1333244

 

In their last quarterly investor call they worded it pretty straight that first 7nm chips would be TSMC. They will be using both GloFo and TSMC 7nm regardless.
edit: also, wasn't GloFo supposed to start risk production early H2, which would rule out working 7nm silicon from them

 

The last slide of the deck posted at anandtech says, AMD has working silicon of Zen2 in the lab:
https://www.anandtech.com/Gallery/Album/6399#24
Zen2 was supposed to be fabbed at GloFo, no?

 

I think it was on their last conference call, when lisa mentioned zen2 will come from tsmc and GF and it seems Epyc will be produced at tsmc.

 

March 18 2019 is the date for the Volta successor announcement. No reason to announce earlier.

 

Their roadmaps seem to be very … let's say flexible. In one earlier iteration, Maxwell was supposed to be the big leap from Kepler in terms of DGEMM/W. Pascal appeared only lateron and originally, Volta was scheduled to introduce stacked DRAM.

 

Volta was also to introduce Tensor Cores according to one of the R&D CUDA engineers even if it was not in the public presentations, or at least became part of it in the early stages of design function scope.
Makes sense as it is not something they could develop and integrate as cleanly as they have unless it has been part of the R&D design-scope feed for the Volta project for awhile, although fair to say the push was mixed-precision narrative for a long time with regards to Pascal-to-Volta.
Possibly part of the reason to introduce Pascal 1st without that, while also ticking some technical risk milestones with Pascal for their high profile commitments.

 

Perhaps Nvidia have done what AMD done, focus on AI(or something else) at the expense of gaming.
Nvidia do appear to have their fingers in many pies lately.