IBM takes a leap to 7nm

[NotFuzzyWarm]

<snip> meet the first guy that introduce the first IBM pc it was joke i was told and a bet the first day they sold out with in hours then the PC time started were we could buy affordable ones.

Ja. IBM - and Apple for the most part - had no idea what they had given birth to. For the IBM-PC it was the accountants buried in dimly lit back rooms or small store fronts that gave praise to it when the first spreadsheet program came out. One story has several breaking out in tears when they first saw how they could now tally up the books and produce records vs filling in each entry on paper and using a comptometer (adding machine) or doing the math in their heads and scribbling down the tallies...

[1714850886]

1714850886

[1714850886]

1714850886

[1714850886]

1714850886

[1714850886]

1714850886

[NotFuzzyWarm]

As for mining ASIC's at the 10/7nm nodes: Not for a long long time. Read the articles again: The focus is on putting massive #'s of gates in the dies eg, more cpu cores, more and larger L1/2/3 caches, etc. Not fitting more chips on a wafer or dropping power needs. That implies that doing scads of chips with fewer than several billion gates just ain't worth it.

We've seen the results of ASIC companies trying to fit entire miners on just a couple huge chips and the results were not good. Even then they were talking about probably only a few million gates at most in BFL's Monarch chips for one example..

[sidehack]

...and then requiring about a 300A multiphase buck at probably 85% efficiency, which is pretty generally a bad idea.

[NotFuzzyWarm]

Ja. Silly question perhaps but how many hashing cores/engines do Bitmains chips have (or Avalon's, BitFury's)? In the 1385 data sheet they always refer to things as 'core'. Singular, not plural or more.

I believe the Monarch had 256 cores/chip or something like that. I know Inno/Bitmine.ch's A1 had "27 highly optimized hashing engines based on custom ASIC cells up to 1.6M" (quote from the A1 spec sheet). IF what they refer to as cells are actually gates (or if each cell is actually 2 gates) then that raises the count but even w/256 nearly 10x more gates it's a long way from the near 2-billion in the latest (affordable) CPU/GPU/s.

2112 wanna pop in on this? Should be right up your alley.

[sidehack]

BM1380 had 8, 1382 had 64 I think, BM1384 55, BM1385 50.

[NotFuzzyWarm]

BM1380 had 8, 1382 had 64 I think, BM1384 55, BM1385 50.

Hmm. just looked at the 1385 data sheet and dinna see it spelled out. All references are singular, as in 'core'.

[sidehack]

I just did the math. Frequency times chip count times core count equals hashrate, for the numbers given for an S7.

[2112]

Ja. Silly question perhaps but how many hashing cores/engines do Bitmains chips have (or Avalon's, BitFury's)? In the 1385 data sheet they always refer to things as 'core'. Singular, not plural or more.

I believe the Monarch had 256 cores/chip or something like that. I know Inno/Bitmine.ch's A1 had "27 highly optimized hashing engines based on custom ASIC cells up to 1.6M" (quote from the A1 spec sheet). IF what they refer to as cells are actually gates (or if each cell is actually 2 gates) then that raises the count but even w/256 nearly 10x more gates it's a long way from the near 2-billion in the latest (affordable) CPU/GPU/s.

2112 wanna pop in on this? Should be right up your alley.

I really don't have anything important to add to this thread.

1) As far as core counts: I wouldn't pay much attention to the marketing blurbs. Technical accuracy isn't an objective there. The words and sentences are mostly meaningless, whatever sounds most impressive wins. Only trust sidehack's calculations or equivalent methods. "Custom cell" may mean as little as "standard cell" with  JTAG/testing connections/circuitry removed/defeated.

2) As far as IBM sharing the access to the 7nm process with fabless boutique designers: I see no problem from the IBM side, IBM typically is quite good with sharing their technology with their long term partners. I see a problem from the coin mining entrepreneurs side: this niche doesn't attract trustworthy/stable people. IBM does thorough vetting of their prospective partners / developers / resellers. Even people much less skeezy that in the Bitcoin milieu fail to make an entry grade in the IBM Developer's Program. Even Spondoolies, the least weird company in the niche, found hiring full time hardware engineering staff impossible.

[QuintLeo]

The part that caught my eye was the "die made of a Silicon/Germanium alloy"

 One of the reasons Silicon took over from early Germanium usage was that it could handle HEAT a lot better - gotta wonder what the heat limits on this new alloy are going to be like, and the power handling capability.

 On the other hand, the band gap on Germanium is a bit less than half that of Silicon (.3v vs .7v or something like that IIRC?), which should help a little on lowering power usage.


 I'm not sure about the A1, but the Innosilicon A2 has 432 hash cores / chip (going by the highest core counts on my 88MH unit, the software doesn't show that on the 110s).
 I'd guess the A1 should have had MORE as SHA256 is a simpler algorythm than Scrypt is.



 I suspect IBM will share the tech with their partners - eventually - but they have to get it working reliably first. I think the claim about 7nm being in full production by 2017 is crazy and 2018 wildly optimistic, especially since they need new wafer manufacturing of their NEW wafer material to happen to SUPPORT the new process - that's not an overnight thing to achieve. I just don't think the INFRASTRUCTURE to support the new tech is in place enough to manage 2018, much less 2017 - and even if it IS managed by 2018, IBM will be using it internally for a while first before they start farming it out at all.


 I doubt we'll see this technology used in cryptocoin mining before 2020 - and wouldn't BET on it by then.

[2112]

I suspect IBM will share the tech with their partners - eventually - but they have to get it working reliably first.

The last part of the above sentence is plainly wrong. IBM first shares the technology with the partners to debug it, then once it works reliably starts selling for money to the regular customers. This is why IBM is so picky when forming partnerships: they have to be reasonably assured that the partner relationship will provide them with truthful feedback required to achieve IBM's customers' expectations of quality.

You seem to have this partnership relationship backwards.

[Mr. Kashif]

I suspect IBM will share the tech with their partners - eventually - but they have to get it working reliably first.

The last part of the above sentence is plainly wrong. IBM first shares the technology with the partners to debug it, then once it works reliably starts selling for money to the regular customers. This is why IBM is so picky when forming partnerships: they have to be reasonably assured that the partner relationship will provide them with truthful feedback required to achieve IBM's customers' expectations of quality.

You seem to have this partnership relationship backwards.

So, Mr. 2112, Do you think we'll be seeing 7nm miners soon in the industry?

[RoadStress]

So, Mr. 2112, Do you think we'll be seeing 7nm miners soon in the industry?

He should be the Lead Designer/Project Manager for this chip!

[brontosaurus]

No one is going to be seeing a 7nm miner in the next 3-4 years, that's almost certain (assuming Bitcoin is still around then) because it's still at the prototyping stage, the process has to be fully engineered and characterised and that will be a majot challenge in itself. Even when IBM or thr others start taking orders it'll be for chips that really, really need the density and/or power characteristics. The NRE's wll initially be truly horrific.

As for some of the comments about devices with lots of 'core' (or more likely pipelines) it's no more difficult to make a big chip with lots of them than a small chip with few, there is such a thing as Designing for Manufacturability which a lot of chip designers seem to ignore just as they ignore the importance of proper product engineering at the back end. Chip design is, or rather should be a team effort.

Just my views, please feel free to express yours.

[RoadStress]

No one is going to be seeing a 7nm miner in the next 3-4 years, that's almost certain (assuming Bitcoin is still around then) because it's still at the prototyping stage, the process has to be fully engineered and characterised and that will be a majot challenge in itself.

I thought a bitcoin mining ASIC chip is pretty basic and straight forwards. Wouldn't it be wise to prototype and to calibrate the machines using mining chips since they are so simple? Just asking.

[NotFuzzyWarm]

No one is going to be seeing a 7nm miner in the next 3-4 years, that's almost certain (assuming Bitcoin is still around then) because it's still at the prototyping stage, the process has to be fully engineered and characterised and that will be a majot challenge in itself.

I thought a bitcoin mining ASIC chip is pretty basic and straight forwards. Wouldn't it be wise to prototype and to calibrate the machines using mining chips since they are so simple? Just asking.

To put it in short form, ja miner ASICS could make very simple handy process targets to tinker with. The companies researching 10/7 are not about to make the process targets as for-saleable items.

The fact that IBM and friends have already done a functional mammoth count 7nm test chip says that they are way beyond needing simple targets and have to concentrate on the process and materials basics to make the process usable on a commercially viable scale.

For 1, the EUV laser based light source process is last I heard from Trumpf is still falling short of the expected target of 250w of EUV light on-target (the photo masks). Current output of the system is around 125-150w. Usable but not a profit maker. To say the least, zapping 50k per-sec 10um droplets of tin with several MW of peak power to turn into the EUV light emitting plasma and then collecting/focusing the light is no small feat. Did I mention that it is also al done under high-vacuum? Air is a very strong absorber of the wavelength produced.

[2112]

To put it in short form, ja miner ASICS could make very simple handy process targets to tinker with. The companies researching 10/7 are not about to make the process targets as for-saleable items.

The fact that IBM and friends have already done a functional mammoth count 7nm test chip says that they are way beyond needing simple targets and have to concentrate on the process and materials basics to make the process usable on a commercially viable scale.

The "for-sale" problem is usually solved by leasing the new chips with mandatory return to the manufacturer after the end of the useful life of the product. I've learned about this trick a while back: hardware RAM disk (mainframe device) manufacturers would lease large quantities of experimental and not-up-to-spec DRAM chips to assemble their "disk drives". IBM (and others) were happy to just lease those "RAM disks".

The general idea of leasing would actually mesh quite well with the Bitcoin mining business model.

[NotFuzzyWarm]

For industrial scale btc farms and other large blockchain processing concerns ja it might if IBM & friends could be talked into it. BW's dealling with Samsung may or may not help there. The peta farms are the only ones with large enough equipment turnover and certainty of where the gear is at (ownership) to make it worthwhile. For us lowly 'small' miners best we can do is the recycle yard or pass-it-down resale.

[2112]

For us lowly 'small' miners best we can do is the recycle yard or pass-it-down resale.

I was under impression that many of the 'small' miners are actually hosting/co-locating their mining equipment at the 'moderately large' mining farms. That would be an equivalent of the olden mainframe days where many 'small' mainframe owners were actually co-locating them at reasonably large data centers and frequently renting/leasing peripheral devices (like large hardware RAM disks) only when needed.

Then the only remaining issue is the physical security: that the miners using experimental chips did not go missing in the night or the owner of premises for the mining farm repossesses/places a lien on the equipment for nonpayment of the rent or electricity bills.

Edit: I remember seeing the photographs of the DRAM banks for the aforementioned hardware RAM disks. The DRAM chips were all in DIP packages and all socketed. To avoid the theft the DRAM banks were then cemented to steel plates on top to prevent individual removal. One of the "thefts" to prevent was actually not theft (chips goes missing) but unauthorized replacement (experimental chip replaced with an off-the-shelf equivalent). Apparently in the DRAM business seeing and measuring un-binned chips would allow detailed reverse-engineering of the manufacturing process. This includes not only taking the actual possession of the experimental chip and de-capping it but also a temporary removal from the original socket to run a battery of post-manufacturing electrical tests and then putting the chip back in the original circuit.

[ZedZedNova]

Then the only remaining issue is the physical security: that the miners using experimental chips did not go missing in the night or the owner of premises for the mining farm repossesses/places a lien on the equipment for nonpayment of the rent or electricity bills.

If it's a "legit" co-lo facility, security can be part of the package. How secure depends on the facility and how much you want to pay...

Edit: I remember seeing the photographs of the DRAM banks for the aforementioned hardware RAM disks. The DRAM chips were all in DIP packages and all socketed. To avoid the theft the DRAM banks were then cemented to steel plates on top to prevent individual removal. One of the "thefts" to prevent was actually not theft (chips goes missing) but unauthorized replacement (experimental chip replaced with an off-the-shelf equivalent). Apparently in the DRAM business seeing and measuring un-binned chips would allow detailed reverse-engineering of the manufacturing process. This includes not only taking the actual possession of the experimental chip and de-capping it but also a temporary removal from the original socket to run a battery of post-manufacturing electrical tests and then putting the chip back in the original circuit.

When I was working for a prominent US DSP chip maker on one of their engineering test floors, there was a Schlumberger scanning electron microscope (I believe) that was used to test silicon. What it displayed on the screen was pretty interesting and what it could do was pretty sweet. You could see all sorts of details, and based on the colors tell what elements were used and whether a particular trace was energized or not, and if I recall you could tell the difference between voltage levels. Virtual probes allowed you to measure signals and display them in a "scope" window. Granted this was in 1989-1990 time-frame so I know there is way more capable and cooler sch!@t available now.

- zed

EDIT: fixed spelling of Schlumberger...

[sidehack]

...it gets cooler? Dang, I'm in the wrong business. That's freakin' sweet.