Topic: CoinTerra announces its first ASIC - Hash-Rate greater than 500 GH/s - page 85. (Read 231002 times)

You should have told that to hashfast a month ago

Nobody has yet commented on this:

"Warranties. CoinTerra warrants that its Product(s) will, at the time of shipment and for a period of thirty (30) days thereafter, be free from defects in material and workmanship, be free of all liens and encumbrances, and will conform to CoinTerra’s approved specifications. (…)"

This text is an excerpt from: http://cointerra.com/consumer-sales-agreement/

Besides thermal issues, probably the main reason for the huge package is the requirement to reduce the substrate resistance as much as possible.
Just imagine the overall current flowing to the die (or 4 dies if you like) is easily above 120A (@ 0.7V).
If you want to realize a substrate voltage drop below 100 mV, the substrate resistance must be below 0.8 mOhm, which requires a lot of parallel copper connections (and copper vias) resulting in big substrate area.
 

I fully agree, I would realize it with 4 dies too instead of one. Should also be possible, because the 4 KnC quads are fully self contained and could be diced to 4 dies easily.
But above I only analyzed published sources and KnC news is showing only one die containing 4 quads.
Maybe they did that to mislead their competitors. The arrow to the PLL is also pointing to nowhere instead to the real PLL macro on the left side.

Nice summary however I believe KNC is using a 4 die concept.  Their error reports refers to die 0,1,2,3.  Google "KNC die zero error".  The large size of the package (45mm x 45mm) also would indicate that.  It is unlikely they are using a giant package for a tiny die, and equally unlikely they have a massive single die underneath that package lid.  It will be interesting the first time a KNC module dies out of warranty.  Dissolving the package would give some good insight.

So what 28nm ASIC concepts do we have?

KnCMiner (running)
-   TSMC 28HPM (probably) @ 0.7V
-   die area ?
-   1 die per package (see KnC news from 7/19/2013)
-   192 double hash engines per die (see KnC news from 7/19/2013)
-   reported max. hash rate per chip package : 144 GH/s
-   resulting operating frequency : 750 MHz

Hashfast (spec)
-   TSMC 28HPM @ ?V (announced)
-   die area : 81 mm² (18x18 mm² announced by Amy divided by 4)
-   4 dies per package (see open source protocol spec)
-   192 double hash engines per die (see open source protocol spec)
-   specified hash rate per multi chip package (4 dies) : 400 GH/s
-   resulting operating frequency : 520 MHz

Cointerra (spec)
-   GLOBALFOUNDRIES 28HPP @ 0.765V (see FPGA video on Cointerra homepage)
-   die area 111 mm² (resulting from press release 1,5 mm² per GH/s & 3 dies per package for 500 GH/s, 100 mm² also mentioned in their FPGA video)
-   3 dies per package (see FPGA video)
-   121 double hash engines per die (see die layout picture on Cointerra homepage)
-   specified hash rate per multi chip package (3 dies) : 500 GH/s
-   resulting operating frequency : 1377 MHz (1.4 GHz also mentioned in FPGA video)

So, from my point of view the KnC and Hashfast concepts are quite comparable, besides the fact that Hashfast tries to integrate 3x the hash power in one package. Okay they will use more efficient water cooling, but they will still have a 3x time higher current density in their package substrate than KnC has. Maybe that explains their substrate concerns/problems.

Cointerra is much more ambitious. With their power density they will be much closer to the power density hot spots Intel or AMD has in their CPUs, because of the high operating frequency of almost 1.4 GHz. I’m curious how this will end.
Oh, I forgot, they have the same high package substrate current densities as Hashfast has.

Silicon isnt the best conductor of heat (nor the worst); so yes heat spreads across the die, but its thermal conductivity is about half that of a cheap aluminum heatspreader and closer to 1% of that of a copper heatpipe. 


Sure, the total cooling requirements for the entire bitcoin machine will generally be higher. But the lower your thermal density of the chips, the easier it is to cope with that.  Just look in the KnC thread; people are disconnecting the case fans and blocking half the chip cooler because for some weird reason, they perform better when running slightly warmer. But even without case fans, and half blocked coolers (which by default are running low speed fans), these 100-150W chips are running pretty darn cool (50-60C). If you'd try the same with an Intel CPU you would get a very different result.


Im going to pass on this question, though I would be interested in hearing the answer. Fact is intel (amd, whatever) cpu's achieve a far far higher thermal density in their cores (especially ALU's and FP units) than any bitcoin asic. Part of this might be due to higher clockspeeds (and in some cases, double pumped alu's). Core voltage of CPU's is also higher, but combined Im not sure thats enough to explain the difference. So I do wonder if it is not a sign of better optimization.

Dabs New ASIC, Grass-Cooled. Bamboo Heatsinks!

These aren't high end water cooling, they are about as cheap as water cooling gets.  Note this shouldn't be taken as a hit against either company high end water cooling is expensive, maintenance prone, and complex to install.  It really isn't suited for mass production.  These are the OEM (unbranded) version of popular sealed (self contained rad, pump, res, waterblock, & fluid) water cooling systems.

Kudos to both companies for ingeniously using off the shelf parts but I think you overstated the cooling challenges.  Cooling something 24/7 isn't any harder than cooling it for an hour.  You are going to reach equilibrium (especially with water) within minutes and if you don't have sufficient cooling the system isn't going to last an hour.   If it does lat an hour it isn't going to get harder to cool over time, baring a fan or pump failure the temp after an hour, day, week, year will be roughly the same.

my point is that a high end bitcoin miner is designed to run at a TDP of over 300W, ALL OF THE TIME - and not just one chip doing that.. all of them are doing that in one box (in ct.'s case thats 4 chips, each with a TDP over 300W, all next to each other)... whereas an Intel cpu, when over clocked, isn't going anywhere near this zone.  In normal usage, even when being maxed out, it runs at 125 watts (for the big ones) and 90w for the smaller ones.   It only ever runs at 250_ watts when its being tweaked by an over clocker, and even then, it doesn't run at that wattage all of the time... it is bursty.  its only during a task, like a video compression or math problem.. that it might do that, for hours maybe.. but not for days and weeks upon a time.  that full time load puts a lot of strain on the cooling system.

and yes, both hashfast and cointerra are using high end pc cooling technology, but even then, I'm sure they're using the extra strength version of it.. not the standard common or garden models.   cointerras radiators look custom (and big!)... they're not similar to anything I've seen in a store.. though hashfast's radiators do look similar to high end pc ones like the Corsair Hydro series (probably H80)

Raw wattage isn't the whole story in thermodynamics.  Your lawn on a hot summer day probably absorbs a couple KW of heat.  Do you have exotic water cooling or install heat sinks on your grass blades to keep it from bursting into flames?

Now that was a intentionally silly example but heat flux especially localized heat flux is relevant in the heat transfer requirements of the cooling system.  This is why both HashFast and Cointerra are able to use OEM CPU coolers despite as you point out no CPU has a TDP of 300W+.  They are designed to handle the much higher heat flux (W/cm2) present in modern CPU and cooling a large die with relatively even and low heat flux is less of a challenge.

just out of interest, i found some articles that discuss the watts that a heavily over clocked intel cpu might draw, when running a pretty maximum application... and here's the source... even with max over clocking (and over vaulting), and maxing out the cpu usage, it still doesn't get close to 300 watts!  and this is an extreme test, not an application that will be running 24/7/365 like a bitcoin miner.

http://www.bit-tech.net/hardware/2013/06/01/intel-core-i7-4770k-cpu-review/7

I'm still not sure how relevant that is, as the heat dissipates through the entire die, and via the package and board, into the cooling system... so its the total heat generated thats important and I'm not sure how relevant the spot power is... though maybe I'm missing something.

anyway, we talk in averages over the 'package' that needs to be 'cooled'... and an intel chip has a TDP of 125 Watts, which pretty much means that the entire package, at most, will generate 125 Watts of power (aka Heat) output, at its maximum usage...   whereas these bitcoin mining chips (both hashfasts and cointerras) will generate 300 watts of power, ALL OF THE TIME.   And if we overclock them, which I'm pretty sure we will all do, they will no doubt hit 350 or maybe even more watts per chip !   A bitcoin mining chip will use a LOT more power, and require much better cooling, than a regular intel chip.

even the heavily over clocked and over-volted intel chips that have been tweaked by enthusiasts, probably don't draw 350 watts of power, 24/7/365 !   thats my point.

i also think its questionable whether there's any parts of an intel chip that will toggle faster, in a denser layout than a hashing core... as almost all of the core, toggles almost every cycle.. thus how does it get any more power consumptive than that?

CT gives you the die size indirectly.  Claimed 1.5 GH/mm2.

Still I think Puppet's point was that in a CPU die cooling is more of a challenge.  While the die on average may only be 5W/cm2 in certain spots under certain conditions there can be localized heat flux of 10x that say 50W/cm2.  A SHA processor, any SHA processor is going to be more uniform in heat distribution.

so far, we know performance and power specs are a claimed >500 GH/s per chip at 0.6 W/GH...  ergo, each chip will consume at least 300 watts of power.  they haven't stated the die size publicly (though i believe its similar)

hashfast has announced 324mm2 running at 400 GH/s... which - as you say - would consume around 260 watts at 0.65W/GH... though its been designed to overclock til it hits max temp (90? 100?)...   if it didn't need any additional voltage, a linear scaling would be a similar wattage to the cointerra chip, i.e. 300 watts at 500 GH.. but since its nominal speed is 400 GH it might be safe to assume it might need to be over-volted to reach 500 GH... so lets 'guess' it might draw 350 watts at 500 GH assuming its possible.

both hashfast and cointerra utilise liquid cooling... so their cooling performance could be pretty much similar.   the cooling airflows on the two boxes are slightly different but not drastically (hf have 3 smaller radiators, ct have two larger radiators)

Thx aerobatic for nice summary. I'm pretty sure all competitors after first laugh form others mistake/fail go to table and double check  own project
About competition.
Yes, bitfury is admittedly a bit unstable, but still sells a lot of Gh at a price comparable to the KNC.
Avalon sellout all next gen chip in few days  How is it possible Payments in BTC only.
There are other player like Chinese BTCGarden who  still produce cheap equipment and deploy many Th http://www.btcgarden.com//figures.jsp
All together it consists of ascending difficulty. Starting in January will be hard.

Dont know CT provided specs, but IIRC, hashfast uses 4 ~80mm² dies per chip. ~65W per 80mm² die is pretty low thermal density compared to many cpu's and gpu's, and thats even ignoring the hotspot issue.

Now I do wonder if this is because these chips arent nearly as well optimized as, say x86 chips, or if its just because of the workload, but as it is, cooling those chips is "simple" compared to the highend x86 world.

Good points- but just a small correction- they made a bitfury design becasue I paid them, and They wanted to get extra if I was paying for some anyways

Status..   Cointerra announced they've taped out (a month later than planned).  They're still saying they're on time for arrival of chips in Dec (which is almost unbelievably fast).   they say they're on time for their promised delivery dates and that their schedule isn't impacted by the delay (which presumably means they've paid a heck of a lot of expedite fees to their fab, and that whatever slack they had in their schedule is probably now used up).

Expected date.  They said they sold out of december (small) and january (big) batches.  most customers are in the january batch.  presumably when they get their chips back from the fab, if yield goes better than expected they might have some extra from those batches to sell but they probably don't know yet til they arrive...   - yields for bitcoin chips should in theory be better than regular asics since there's a lot of repeated cores in a bitcoin mining chip and if one of the cores has a defect the others may still work so yields may be much better than expected...  - meanwhile they're now selling a feb batch.   since the dec batch was sold as a late dec batch, and the jan batch was sold as a mid jan batch... i think even before the late tape-out, many of their december customers had voluntarily switched into the january batch to take advantage of the significant lower price point in jan ($3/gh).

Threats, risks...   well... that depends on how well you think they will execute on their production, as the main risk has now been eliminated which is they've taped out, which means they've had enough funds to pay the fab and all their partners quite a few million bucks...  the remaining risks are production and schedule led.  Of the two competitors, one - KncMiner, executed flawlessly and got chips back when they thought they would, and stuffed the chips on boards and shipped within days of when they said they would...  and the other competitor - hashfast - was on an aggressive schedule... sold everyone october systems (at an october price) and has now delayed two months due to unforeseen production problems so is now planning to deliver in december.  says the chip were on time but the substrates caused them a delay.  quite why their substrates were delayed, no one really knows.   there are a few other competitors like BitFury with an excellent 55nm chip (but not professional grade systems)... and there's some newcomers like BitMine, AMT and Black Arrow.. who may also be in the running.  Black Arrow's delivery is a bit too far away.. they're promising feb but haven't taped out.. and are selling bitfury based systems to try and raise funds for their tape-out... which I've no idea if that strategy will work... they've got to sell a lot of bitfurys in a very competitive market...

you can be pretty damn sure that all the newer asic companies making chips will be calling their partners and telling them whatever happened to hashfast better not happen to them... so I'm quite sure they will be triple checking their substrates for flaws, for a start!   When one company has a major disaster like that, all the others can watch from the sidelines and if they're any good, they will learn and try not to make the same mistakes.

Topic is stretched so much and there are plenty of entries other than the main subject. Can anyone be tempted to do a short summary?

• What is the status of the chip production?
• Expected date of delivery based on your experience?
• Threats, risks at the time completion?