Nvidia Turing Speculation thread [2018]

TSMC accelerating 7nm process production plan to meet demand
June 8, 2018

https://www.digitimes.com/news/a20180608PD211.html

 

They started volume production early Q2 already, how is this "mass production" different from that?

 

Takes about 3 months to produce a single wafer, so Mass production follows Volume production and is the point where the process is complete and silicon wafers are available to customers.

https://www.fool.com/investing/2018/04/25/apple-incs-a12-chip-is-likely-in-mass-production-r.aspx

 

September release likely:

More in link.

https://translate.google.com/translate?sl=de&tl=en&js=y&prev=_t&hl=de&ie=UTF-8&u=https://www.tomshw.de/2018/06/12/nvidia-geforce-gtx-2080-1180-oder-was-wir-ueber-turing-nicht-wissen-update/&edit-text=&act=url

 

Lines up with the rumors of a Gamescom announcement. I just hope it delivers.

2+ years since Pascal has certainly raised my expectations.

 

I still don't get it. Either they're on 12nm launching in September or so, and the [don't expect it for a long time] thing was a really weird dodge, because what they'd be dodging I don't know.

Or they're on 7nm and launching next year for mass consumer. Or, well ok. If AMD can put out a pro only 7nm chip from TSMC this year then so can Nvidia. Charging mid 5 figures or more offsets the low volume and high chip production costs you'd find making large dies on a new process. Consumer chips of the architecture could come next year. Wouldn't explain the comment, but hey even modern tech CEOs, coached as they are, can make minor flubs.

 

I don't believe Nvidia will release 7nm chips this year, but as far as that sentence goes, based on recent history, I'd say that "if AMD can put out a pro only 7nm chip from TSMC this year then Nvidia can do better with almost any chip."

What die size would a potential 7nm "Turing" chip need be to slightly beat a 1080 Ti? TSMC claims:

and

EDIT: ^^This is about the 10nm process

I mean, I don't know how accurate those claims really are, but the gap is immense with 16FF+. Even being conservative it looks like a 180mm^2 would suffice. taking those claims at face value, it would be less than 140mm^2!!

And as I said above, I'd say that if AMD can put out a 300+ mm^2 in enough quantities for any market, Nvidia could certainly handle a 170-200 mm^2 chip in enough quantities for a GTX 1180 (paper) release, if that's what they really wanted to do.

 

From a trustworthy source in a forum but I will not link to it.

 

EETimes did an article back in May regarding the 7nm multi-patterning and 7+ EUV process from TSMC compared to their 16FF+ (rather than the "12nm" mature 16FF process).
https://www.eetimes.com/document.asp?doc_id=1333244

The non-EUV DUV 7nm multi-patterning process cost is higher relative to the TSMC 7+ EUV, so would influence its use.

For contrast Global Foundries mention for 7LP 40% more speed or uses 60% less power, density a bit less than TSMC; GF product brief spec goes back to late 2017 but still dated 2018 so would seem accurate.
https://www.globalfoundries.com/sites/default/files/product-briefs/pb-7lp.pdf

 

https://be.hardware.info/nieuws/57763/lnvidiars-nieuwe-lturingr-geforce-gpu-komt-in-septemberr
Turing for September apparently

 

Even though that the German Tom's guy (the original source) is apparently pretty reliable, he did also say earlier that it would be July.

 

Maybe NVIDIA is planning on launching a founders version a month or two before the partner boards ship. All the recent rumors are coming from board partners it seems and not NVIDIA.

 

Re: the leaks about a "new boost clock algorithm" for Turing:

It seems to me as though Nvidia's boost clock algorithm is already quite mature, and as such there is minimal clock headroom left for overclocking beyond what the boost algorithm can already achieve. If Nvidia wants to "pull an AMD" and push Turing toward the outer edge of the thermal/power envelope they will only gain perhaps 5% over what is already achievable. That leaves process improvements to make up the rest of the expected performance gains over Pascal, a notion I find dubious given the nature of the "12nm" (i.e. 16nm+) process involved.

All this makes me wonder, perhaps even hope (naively) if perhaps the leaks as they have been interpreted thus far don't quite tell the whole story. If Turing is to gain any sizable amount of performance over Pascal on the average (gaming workloads), given the same number of ALUs and only minor architectural changes, there will need to be some major clock speeds improvements to achieve this. Given what I've already mentioned about the likelihood of this prospect, to me this only leaves one possible solution:

The return of clock domains.

Discuss.

 

Where did you see that they will maintain the same number of ALUs? I doubt we will see the return of double pumped ALUs, given NVidia's focus on efficiency since Kepler. Fermi was the end of the road for NVidia's Netburst.

 

Every rumor/leak that we have seen up until this point has indicated "GT104" will contain 3584 ALUs - i.e. the same amount of enabled ALUs found in *most* GP102 SKUs. It is upon this assumption that I base my premise. If this turns out to be incorrect, then my premise will of course be untrue.

This is not to say that there won't be more ALUs in larger SKUs e.g. "GT102/100", but that could be a year or more away.

 

Oh I see what you mean. Well, GP102 clocks around 1.5Ghz, some 200 to 300 Mhz lower than GP104. With the 12nm shrink they probably can clock even a bit higher (2Ghz?) and together with some architecture changes, plus GDDR6, that might be enough to have a 30% increase over GP102? I really do not think we will see double pumped ALUs.

 

Specified "OEM" boost clocks do not match reality, in my experience. I've owned both GP104 and GP102-based graphics cards and they both hit the same boost clocks ~1900MHz without overclocking. Now, that is with power maxed out and under waterblocks but there was no manual overclocking involved. When overclocking is involved, the cap seems to be around 2.1GHz for Pascal GPUs regardless of SKU.

 

This is to be expected.

• Recently, the 100/102-type parts have been roughly 1.5x of the 104-type parts (ALUs, memory bus, etc).
• The 100-type parts have had as many ALUs as the 102-type parts (at least for Pascal).

And then we can apply that knowledge:

• GV100 has 40% more ALUs than GP100/GP102.
• If the 1.5x ratio holds, then we'd expect "GT104" to have 40% more ALUs than GP104.

Sure enough, a 40% bump to GP104's 2560 ALUs is exactly 3584 ALUs.

So I think it's very reasonable to expect 3584 ALUs in "GT104".

 

I would be shocked if anything other than this were to occur when GT104-based SKUs launch in the next 1-3 months.

Just for the sake of clarification, my theory pertains to the method by which GT104 may surpass GP102 in raw arithmetic performance, which would require something more exotic than minor process and boost algorithm optimizations would allow, in my opinion.

 

With boost it's average is 1800Mhz, oc'ed easily to 2000Mhz.