7nm miner thread

[promojo]

Bit of food for thought for those that think 7nm will show up next year.

https://arstechnica.com/gadgets/2017/08/intels-next-generation-chip-plans-ice-lake-and-a-slow-10nm-transition/

Given the issues with 14/16nm, and the reported issues getting 10nm up to speed, I'm starting to think that *2019* might be excessively optimistic for actual 7nm production to arrive.

Thanks for sharing!

[QuintLeo]

Going by commentary out of IBM, *THEIR* 7nm process will be using a "hybrid" Silicon/Germanium wafer.
No solid word out of anyone else except an Intel comment that "10nm will be the end of the line for pure silicon".


 Economic reality is that no miner maker will be making new gear on the new node(s) for a while, but that they WILL have to make the move for competative reasons probably 2 years more-or-less from now when the new node(s) start getting affordable yield figures and the relative efficiency has become known between them.

If the new nodes take longer to deploy than the current projections (like the 14/16nm generation did), it could be 3-5 years before miner makers will be ABLE to move to them.

Interesting... Thanks for pointing that out...  Off to learn what Germanium wafer is I go.... lol

 Germanium, like Silicon, is a semiconductor material.
 Germanium was actually more common than Silicon in the early days of semiconductor manufacturing, but it turned out Silicon had a higher tolerance for heat AND was less expensive to make wafers for (silicon is literally dirt-cheap - a large proportion of "dirt" is Silicon, along with Aluminum and Iron).

 Germanium has the advantage of a lower "band gap" than Silicon, which makes it easier to push electrons around in the material - this might buy 1 or 2 more generations of semiconductor manufacturing before MAJOR structural changes have to happen due to the increasing issues with Quantum-level effects on small node sizes, but I'd not bet on Germanium going past about 5 nm without running into the SAME issues Silicon is hitting at 10 - and the relatively poor heat capacity is going to cause it's OWN issues.


 Moore's Law isn't dead yet - but it's been limping a bit for the past decade, and it's definitely not looking good past the NEXT decade.

[NotFuzzyWarm]

I'll say one thing - it will be interesting to see how thermally stable using Ge will be...
And I bet it rules out high-power apps such mining chips

The first transistors were made using Ge and even today it excels at having about 1/2 the band gap voltage that Si does, a Ge P-N junction voltage is around 0.3V and Si is around 0.7V (yes doping lowers it but still applies) and is more radiation tolerant (Aerospace industry loves it) but using Ge brings with it 3 things: High intrinsic leakage, thermal drift, and lower max allowable junction temps than silicon-based devices (also lower max Vc but that does not apply here). Those 3 things are what allowed Si-based devices to take over semiconductors so many decades ago.

Gotta assume that using different techniques the leakage factor has been solved, that just leaves the junction voltage drift vs temperature.. As for max temperatures -- the temps that current mining chips run at will destroy them in a heartbeat.

edit: damnit - ya beat me to the main points....

[promojo]

 Germanium, like Silicon, is a semiconductor material.
 Germanium was actually more common than Silicon in the early days of semiconductor manufacturing, but it turned out Silicon had a higher tolerance for heat AND was less expensive to make wafers for (silicon is literally dirt-cheap - a large proportion of "dirt" is Silicon, along with Aluminum and Iron).

 Germanium has the advantage of a lower "band gap" than Silicon, which makes it easier to push electrons around in the material - this might buy 1 or 2 more generations of semiconductor manufacturing before MAJOR structural changes have to happen due to the increasing issues with Quantum-level effects on small node sizes, but I'd not bet on Germanium going past about 5 nm without running into the SAME issues Silicon is hitting at 10 - and the relatively poor heat capacity is going to cause it's OWN issues.


 Moore's Law isn't dead yet - but it's been limping a bit for the past decade, and it's definitely not looking good past the NEXT decade.

Interesting points you mention here.   Thanks for the input!

[QuintLeo]

To be picky, the original transistor (and many original diodes) were point-contact devices.

 Reliability of those wasn't particularly impressive though, folks LOVED junction-type transistors when those started showing up as commercial production parts.

 Germanium still gets used even today in some signal processing and radio usage for detection diodes, but much less so then in the past as AM and SSB modulation has largely been replaced by phase and frequency modulation techniques.

[Philopolymath]

We made Erbium doped Fibre Grad Bragg Filters at JDSU R&D...and used several other exotic elements.
Pushing Light...

Re: Original topic..there is no big threat coming down the pipe for a LONG time..

[QuintLeo]

More issues to consider.

 Even when 7nm reaches actual production - what are the YIELDS going to be like?
 How long will they take to improve and by how much?

 What is the design cost going to end up being (14/16nm design is EXPEN$IVE compared to anything before it).

 What is the cost per chip going to end up being like?

 What is the CAPACITY going to be like the first couple years?



 Keep in mind that the S7 is still viable *NOW* if you have fairly cheap electric despite being a generation outdated.
 The S9 isn't going to immediately become unprofitable even when 7nm reaches actual production in a miner design.

[philipma1957]

More issues to consider.

 Even when 7nm reaches actual production - what are the YIELDS going to be like?
 How long will they take to improve and by how much?

 What is the design cost going to end up being (14/16nm design is EXPEN$IVE compared to anything before it).

 What is the cost per chip going to end up being like?

 What is the CAPACITY going to be like the first couple years?



 Keep in mind that the S7 is still viable *NOW* if you have fairly cheap electric despite being a generation outdated.
 The S9 isn't going to immediately become unprofitable even when 7nm reaches actual production in a miner design.

This  fact makes the rush to 7nm  become more of a walk then a run.

[promojo]

Unlike most people on this list, I am in the chip biz.  I cannot say much.  I can say this:  A US$300M investment in a 7nm-based chip design will get you pretty much that.  Then you'll need more money to get your chips fabbed and you systems built and installed and everything else.  A few $millions will go a long way towards physical plant for installation, but it takes a lot of money to play in the 7nm game. 

There were some comments in this thread about the move from 22/20nm to 16nm.  The 20nm node was largely skipped by the industry because 16nm node added so much value with the introduction of finFETs, that it wasn't worth it to continue investing in 20nm.  The 16nm node still has a ways to go, and has had a number of rounds of refinements.  I don't know which actual 16nm process is being used by BMT, but they could probably get some mileage just by re-spinning into a newer 16nm process.  If they cared about improving the product.  However they're in a situation where they can sell all they can make, and price is hardly an object right now, so why waste the engineering resources? 

7nm is a huge wall to climb for anyone, and I don't think anyone's going to be building a 7nm miner for at least another year, maybe 2. 

If someone does come along and proves me wrong, I'll buy it.

Thanks for the insight and comments.....

[Morguk]

Whether or not they are ready for 2018, It's cool they are producing PCI-E cards. Will be fun to have a decent miner in my PC.

[Sr.Urbanist]

Germanium, like Silicon, is a semiconductor material.
Germanium was actually more common than Silicon in the early days of semiconductor manufacturing, but it turned out Silicon had a higher tolerance for heat AND was less expensive to make wafers for (silicon is literally dirt-cheap - a large proportion of "dirt" is Silicon, along with Aluminum and Iron).
... the relatively poor heat capacity is going to cause it's OWN issues.

BitFury reports to use immersion cooling on their website to nearly double their efficiency. This would probably give them an advantage to understand the tech well if the next generation is going to run hotter.  Seeing those numbers makes me think about researching into it more.

Even when 7nm reaches actual production - what are the YIELDS going to be like?
 

They are reporting 10 Th/s with 500W on the GMO 7nm chipset, so that's about double the efficiency of an S9 - not a large jump like 28 to 16/14nm - but still significant.

 

 

[philipma1957]

Germanium, like Silicon, is a semiconductor material.
Germanium was actually more common than Silicon in the early days of semiconductor manufacturing, but it turned out Silicon had a higher tolerance for heat AND was less expensive to make wafers for (silicon is literally dirt-cheap - a large proportion of "dirt" is Silicon, along with Aluminum and Iron).
... the relatively poor heat capacity is going to cause it's OWN issues.

BitFury reports to use immersion cooling on their website to nearly double their efficiency. This would probably give them an advantage to understand the tech well if the next generation is going to run hotter.  Seeing those numbers makes me think about researching into it more.

Even when 7nm reaches actual production - what are the YIELDS going to be like?
 

They are reporting 10 Th/s with 500W on the GMO 7nm chipset, so that's about double the efficiency of an S9 - not a large jump like 28 to 16/14nm - but still significant.

 

 

That article makes me think less of them as of today no one has good 10 nm

And they are talking 7nm 5nm and 3.5 nm

My thoughts are this. 7nm in 2019 at best.

5nm in ?

3.5nm. I think comes in over 10 years.

[Sr.Urbanist]

That article makes me think less of them as of today no one has good 10 nm

I, too, am beginning to think more and more that they are dreaming as I look into their project.   

[AtraxPool]

I think we are still to see the advantage of immersed chips and other forms of liquid cooling that would be the next step I see being developed.

Liquid cooling is still serious option available and the performance improvements are also something that could be looked into.

[QuintLeo]

Germanium, like Silicon, is a semiconductor material.
Germanium was actually more common than Silicon in the early days of semiconductor manufacturing, but it turned out Silicon had a higher tolerance for heat AND was less expensive to make wafers for (silicon is literally dirt-cheap - a large proportion of "dirt" is Silicon, along with Aluminum and Iron).
... the relatively poor heat capacity is going to cause it's OWN issues.

BitFury reports to use immersion cooling on their website to nearly double their efficiency. This would probably give them an advantage to understand the tech well if the next generation is going to run hotter.  Seeing those numbers makes me think about researching into it more.

Even when 7nm reaches actual production - what are the YIELDS going to be like?
 

They are reporting 10 Th/s with 500W on the GMO 7nm chipset, so that's about double the efficiency of an S9 - not a large jump like 28 to 16/14nm - but still significant.

 You're addressing a question there.
 Yield has to do with how many usable die you get from a wafer as a percentage of the POSSIBLE die you should get if there are no flaws.

 Efficiency - I'm starting to question their figures a LOT, 28nm to 14/16nm in Mining ASIC so far has seen about half the power usage per hash in most usage on comparably optimised products (S7 to S9 for example).

 Given that quantum effects have been limiting the gains in both possible clock rate AND in efficiency at an increasingly worse level per generation for GENERATIONS now, I doubt we're going to see an improvement of 2.5 TIMES the efficiency from 7nm - even if it really IS 7nm and not just CALLED that while having real feature size closer to 9-10 nm range.

[philipma1957]

I think we are still to see the advantage of immersed chips and other forms of liquid cooling that would be the next step I see being developed.

Liquid cooling is still serious option available and the performance improvements are also something that could be looked into.

Liquid cooling  is not  bad  it works well for cpu's and gpus.

if GMO  ends up  seling a gpu style miner with liquid cooling I would buy it.

[AtraxPool]

I think we are still to see the advantage of immersed chips and other forms of liquid cooling that would be the next step I see being developed.

Liquid cooling is still serious option available and the performance improvements are also something that could be looked into.

Liquid cooling  is not  bad  it works well for cpu's and gpus.

if GMO  ends up  seling a gpu style miner with liquid cooling I would buy it.

I have looked at gmo I think what we are talking about could be a possiblity.

Not to forget this guy..

https://bitcointalk.org/index.php?topic=346134.0

[Sr.Urbanist]

Not to forget this guy..

https://bitcointalk.org/index.php?topic=346134.0

That is rad.

[AtraxPool]

Not to forget this guy..

https://bitcointalk.org/index.php?topic=346134.0

That is rad.

I think we have yet to see full potental from what we have already let alone 10nm and 7nm chips.

This is from the 16nm series chips from bitfury a while back.

https://youtu.be/DSbmuBvDWo0

There should be more development with the chips that are currently available and I think chips around 7nm will have no choice but to be cooled in this way.

Esoteric methods of manufacturing like extreme ultraviolet lasers will aid in the process of such chips.

[Sr.Urbanist]

There should be more development with the chips that are currently available and I think chips around 7nm will have no choice but to be cooled in this way. 

It'd be nice to save some energy, right now, but I'm worried about playing with the S9s.  One could probably increase efficiencies with liquid cool, but I'm afraid I'd ruin them.  It took me eight months before I was comfortable taking them apart to clean.  I'd probably try it on an S7 first if I owned one.