Glad to hear I didn't screw up completely. Going to spend tomorrow morning placing/routing the PCB with the goal of submitting it to OSHPark for a run next week. Will report back once I have the gerber files. This is my first attempt at a single-chip schematic. No joke, I was more or less copying the c-scape design for their initial test PCB. Electrically I think this is all correct? I'm going with the same decoupling capacitor values* that were mentioned in a post by intron.
Looks fine. Toch? intron What is Toch? Didn't see anything like that mentioned in the chip pinout, not sure if I'm missing something for the SPI connections? Maybe add a solder jumper from INCLK to GND.
As far as calculating bypass caps, to do a proper job you need to know the layout of the board and the exact specifications of the capacitor. As they say: "ain't nobody got time for that!"
Excellent point for the INCLK pin. I don't see myself driving the chip with an external clock source so I'll pull it to ground. since we're on the topic - I'm assume switching to 0603 size caps doesn't pose a problem, correct? After hand soldering 150 0402 caps, I'm ready to not ever do that again  I used all 0603, and they're fine. My supply planes are nice and quiet. I didn't fancy placing those tiny 0402 capacitors. Since I'm just doing a cheap OSHPark run of 3 boards, I hopefully won't lose too much if I run into issues. |
|
|
|
This is my first attempt at a single-chip schematic. No joke, I was more or less copying the c-scape design for their initial test PCB. Electrically I think this is all correct? I'm going with the same decoupling capacitor values* that were mentioned in a post by intron.  I couldn't figure out how to add a ground pad for the 5GHash chip but in the PCB design phase I'm going to have a copper net underneath the chip with vias to the ground plane for heatsinking and ground purposes. I plan on using a 4-pin female molex connector to provide ground/VDD/IOVDD and a 8-pin header for the various I/O pins. After I layout this I'll have a second PCB with dc-dc converters to provide IOVDD and VDD. That same PCB will have the level-shifter to shift the spi signal voltages between the Raspberry Pi and 5GHash voltages. My goal is to dip my toes into this to see if I enjoy it. So far I do but I haven't had to debug anything yet  I'd like to chain 2 of these together and then build a 2 or 4 chip PCB using the lessons learned from this project. *- I found a couple very interesting PDFs on calculating the decoupling capacitance and the resulting inductance. Still educating myself on this so I decided to go with values that have been proven to work. |
|
|
|
No, 0.25A is not right.
I just measured it on the S-HASH board by putting a 1Ω resistor on the output of the 1.8V regulator. Voltage drop was 95mV, so current is 95mA for 16 chips, or about 6mA per chip, which sounds a lot better.
That is a relief, thanks! |
|
|
|
Speaking of over-clocking, has anyone got a chip hashing with an external oscillator yet? If so, could you make it go faster than the internal oscillator would have been running at whatever core-voltage you're running?
I think bitfury had some results from externally driving the chips when he was first testing them for correctness? Would be interesting what fmax we could get out of the chips if we ignored power efficiency. Power consumption goes up exponentially... I don't think you're going to get much beyond 3GH. Exponentially with voltage, but linearly with frequency(at identical voltages), correct? Yeah, that's correct. I'm not sure using an external clock alone is going to help that much. I'm definitely interested to see what folks come up with though! I imagine, having no professional education on this, that the internal oscillator scales in frequency in a predictable manner. Might be linear, might be exponential, or some other function. I know bitfury showed results where at a certain voltage/frequency, the chip was returning enough errors that a lower frequency was resulting in more overall work. Perhaps the cores required more voltage at that point to be stable? Would have been really neat if the internal oscillator took in a separate Vcc pin so you could tweak each chip's oscillator and core voltage for ideal performance. Bitfury's chips interest me greatly and make me sad I chose computer science instead of computer engineering while in school.. now I'm stuck reading up on CE/EE college websites at night while pretending I understand what I'm reading. |
|
|
|
Speaking of over-clocking, has anyone got a chip hashing with an external oscillator yet? If so, could you make it go faster than the internal oscillator would have been running at whatever core-voltage you're running?
I think bitfury had some results from externally driving the chips when he was first testing them for correctness? Would be interesting what fmax we could get out of the chips if we ignored power efficiency. Power consumption goes up exponentially... I don't think you're going to get much beyond 3GH. Exponentially with voltage, but linearly with frequency(at identical voltages), correct? |
|
|
|
Speaking of over-clocking, has anyone got a chip hashing with an external oscillator yet? If so, could you make it go faster than the internal oscillator would have been running at whatever core-voltage you're running?
I think bitfury had some results from externally driving the chips when he was first testing them for correctness? Would be interesting what fmax we could get out of the chips if we ignored power efficiency. |
|
|
|
Don't forget Bitfury's private pool(at least I believe it is his pool) -> https://ghash.io/. I haven't been keeping up with it but I know it has grown to 100 TH/s very quickly. This has nothing to do with the 100TH(now 200TH) mine. Wouldn't be surprised to see Bitfury just maintain 20% of the network for months to come given how efficient his chips are. edit: Just realized this post may not be useful for the purpose of your thread. Just something that should be considered though given the opportunity that Bitfury has to replace ASICMINER as the 800 pound gorilla on the network. If ASICMINER's 2nd generation chips can compete with Bitfury you could end up with a large chunk of the network held by those two. |
|
|
|
Price for 2.2 gh miner?
I believe 1499 RMB or ~250 USD. About equal to what ASICMINER is charging for their devices with similar power consumption numbers. A little late to the market but interesting nonetheless. |
|
|
|
For your gas to water condenser look into brazed steel plate heat exchangers. I've used the ones sold by dudadiesel as heat exchangers in my cascades/autocascades and other refrigeration systems with no ill effects. Just be careful if you are condensing liquid such that you don't end up with a liquid trap. This is a bigger issue with oil for compressors in refrigeration systems though and luckily I used oil carrying refrigerants to minimize this problem. I think there may be some confusion. I looked up the consdenders made by dudadiesel and they are liquid to liquid.  In open batch immersion cooling the primary working liquid (Novec or Flourinert) boils and the gas rises. It isn't piped anywhere it just rises until it contacts the "cold" (cooler than boiling point) condenser which causes it to condense, and "rain" back into the batch as a liquid. I guess I am missing something or you misunderstood. Ah, for that a large coil would be sufficient and the flourinert will just drip off the coil's fins. My mistake, I was thinking of something else. |
|
|
|
I am dumb and understand nothing of this technical talk, with hashrates and whatever. But I have 20 BTC invested in AM and am wondering why is value dropping so hard. I am now at half a BTC profit and will probably go in the negative soon. And this shit came at a time I REALLY needed to cash my profit out.  Someone care to ELI5 ?  The value is dropping because ASICMINER controls less of the hash share then it did a month (2 months?) ago. It went from having 25-30% network share to having 8-12% share, leading to dividends from mining decreasing quite substantially. There are the Hardware sales, but Friedcat mentioned they are being held back for the next-gen hardware implementation. TLDR; Dividends decreased? SELL SELL SELL! Lots of weak hands getting out of AM due to the historically low dividends along with a lot of competition springing up. The future of AM is fuzzy. If the 2nd generation chips are competitive with the network then we might see AM spring back up to holding a significant portion of the network AND turn over a significant amount of hardware sales. That'd all be good for dividends and potentially secure AM's future until a '3rd' generation of chips hits the network. |
|
|
|
It is weird that the company gives all the specs except the important one. If this reseller is accurate it looks like it can "remove" (transfer from water to air) up to a 3300W heat load. http://www.orionairsales.co.uk/hailea-water-chiller-hc2200bh-1800-watt--2200-litre-cooling-capacity-1796-p.aspThat would make sense 3300W heat transferred / 1800 W power = COP 1.83 Larger (think whole house) heat pumps have a COP of 3 to 4 and cheaper ones close to 2.5. So 1.83 at peak capacity for a smaller unit makes sense. Your experiment has got me designing a immersion cooling system. To lower operating costs I am considering a two stage cooling loop. "Hot" water from the condenser is pumped to an aircooled heat exchanger (radiator and fan assembly) which lowers the temp to within 5C to 10C (depending on how large/expensive of an assembly you want. Preferably this would be outside so the waste heat is simply dumped. Now in cooler temps this alone likely would be sufficient but on hot summer days it might not be enough so the water then flows to a chiller and back to the condenser. The only issue is picking the right boiling temp. Air cooled heat exchange is more efficient the larger the temp difference between water inlet and air inlet. Water is never going to be hotter than the immersion boiling point. Have you found a suitable condenser. That has actually been harder then I thought it would be. Really one just needs a bundle or tubes, nothing more complex. I am sure a radiator would work but that doesn't seem optimal. For your gas to water condenser look into brazed steel plate heat exchangers. I've used the ones sold by dudadiesel as heat exchangers in my cascades/autocascades and other refrigeration systems with no ill effects. Just be careful if you are condensing liquid such that you don't end up with a liquid trap. This is a bigger issue with oil for compressors in refrigeration systems though and luckily I used oil carrying refrigerants to minimize this problem. All the water loop stuff should be a piece of cake. |
|
|
|
Is the level-shifter required to transform the voltage coming out of the GPIO pins of the raspberry pi(5v?) down to levels that the bitfury chips operate on(3.3v?)?
Yes, but the voltages are 3.3 and 1.8. Thanks for the confirmation. Maybe one day I'll get around reading up on PCB design and tinker with this. |
|
|
|
|
Is the level-shifter required to transform the voltage coming out of the GPIO pins of the raspberry pi(5v?) down to levels that the bitfury chips operate on(3.3v?)?
|
|
|
|
Highest is 68deg in a 76deg room. Not too bad for making it much easier to sit next to.
Huh? How are you cooling below ambient with just a fan? Sorry that was misleading. Two things to add. 1) running caseless and 2) unit sitting next to a house AC vent. (not always on but helps) Also Celsius vs Fahrenheit? |
|
|
|
I'm not so sure there will be that many incremental changes made to chips once they start to hit the process limit. I say this as I expect the margins on chips to become so slim that it would be financial ruin to invest in the NRE for another 28nm chip once you already have one. I guess it all depends on how "good" existing designs are. Taking HF just because they provided die size, hashrate and power info. One 18x18mm die is able to do 400 GHash (nominal - more overclocked**)
Hashing per square mm:
18x18mm = 324mm^2
400 GHash / 324mm^2 = 1.23 GHash/mm^2 Reported power consumption is 400 GH / 250 W = 1.6 GH/W. So HF claims 1.23 GH/mm^2 and 1.6 GH/W @ 28nm. It remains to be seen how "good" those are for 28nm. Are they very good, or barely adequate? If their simulations show an improved version is only marginally better (say 1.4 GH/mm2 and 1.8 GH/W) I agree it wouldn't make sense to try and squeeze out more efficient chips. On the other hand say >2.0 GH/mm^2 and >3 GH/W) are possible. Once we get multiple vendors with stats based on real silicon we can start to get a better idea of relative efficiency. Just responding to say that I agree on all points. I imagine the first 28nm designs may be more safe than elegant to minimize risk, leaving performance on the table. 2014 will be exciting, 2015 maybe not so much. |
|
|
|
@Gomeler
Agreed on both points. Design matters and we won't be moving past 28nm for a while.
As for design there is something horribly broke in BFL design and I can't see them doing a significant redesign at this point. Bitfury and BFL aren't the only two with vastly different efficiencies at the same scale. Granted both of these are vaporware at this point but I find it interesting that Hashfast is reporting ~0.6W/GH (at the chip) and KNC is reporting >2W/GH (at this chip) and both of them are on a 28nm process. We will see when it gets to real silicon but if true that is significant.
As for next gen. We won't be seeing anything smaller than 28nm. 22nm is cost prohibitive and you are looking at huge NRE. By the time it becomes possible the market will be flooded. I do think there will be incremental improvements on the same process (the "tock" in Intel tick tock strategy). There is also a troubling trend which Nvidia highlighted a while back in that the cost per transistor isn't going down much even when processes become mature so the smart money may be on optimization.
The 110/130nm tech is going to be obsolete very soon. I don't mean people won't be able to mine with them I just mean nobody is going to be able to sell new units. If AsicMiner and Avalon are going to stay in the game they will be forced to make a move to 55nm if not 28nm. When everyone is on the same process it is going to come down to design, system optimization, larger runs to reduce incremental cost, and having a solid supply chain. It will be interesting. I don't think every company that is around today will still be around a year from now.
I agree, it sounds like BFL there is something wrong with BFL's chip. Perhaps it has a chunk of the chip that is unusable but they can't disable. We don't know the finer details but it might not be out of the question that they are taking their 65nm design, correcting the issues and shrinking it to 28nm. I'm not completely familiar with the process but I suspect this would involve less engineering effort especially if they aren't hand-placing the design. I'm not so sure there will be that many incremental changes made to chips once they start to hit the process limit. I say this as I expect the margins on chips to become so slim that it would be financial ruin to invest in the NRE for another 28nm chip once you already have one. Unless of course we discover a major optimization to SHA256 which would yield sufficient sales to recover the NRE and turn a profit. I fully expect we'll hit a lower-bound on energy efficiency for a given process (let's say 0.25W/GH for sake of argument) and then it'll come down to the GH/wafer as everybody competes on maximizing profit/wafer and GH/USD. |
|
|
|
BFL is predicting their 28nm chips will use barely less W/GH than BitFury's 55nm chips, when they should be at least 3x better than BitFury, if their design was as good. BitFury's 28nm chips (if they plan to do that) ought to be awesome.
Without getting into BFL's 28nm plans, it's boggling how their 65nm design is eating 4w/GH even after their epic development time, whereas this 55nm design is well under 1w/GH and seems to have come out of no where. BitFury's design didn't come out of nowhere. His FPGA design(from what I've read) was superior to any other design on the market. Reading the little information that was fed to us, BitFury's 65nm shrunk 55nm chips were designed to be a low-voltage and efficient design. They missed their hashing/chip goals but the design has been proven to display fantastic efficiency and power-consumption/throughput scaling. A 28nm poorly designed chip could easily be worse than a well designed 55nm chip. I am excited for all of these upcoming 28nm chips but I'm really excited to see what comes out of BitFury(28nm die shrink perhaps?) and CoinTerra(due to the credentials of the team). I think this next generation of chips will be the generation that we mine on for a year+ as great designs will be limited by the manufacturing process, unlike the past where 110-180nm designs had a lot of room to scale with optical shrinks. |
|
|
|
|
What happened to the chip sales on megabigpower.com? Thought there were cheap packs of ~10 chips for October delivery? $98/chip compared to ~$31/chip in the H-Board seems like highway robbery.
edit: Just realized the $98/chip order ships NOW vs the October delivery of the H-Board. Will you be relisting bulk chip sales for October? I'd be interested in a few dozen chips to experiment with overclocking the chips(moar voltage!).
|
|
|
|
Why not just build a pressure chamber since you have to condense the working fluid and use R134a as its a fraction of the cost? Would only take 80-100PSI to hit a similar range.
I personally would be afraid of building a pressure vessel to contain the R134a. I suppose you could reuse something like a boiler or other already existing pressure vessels that would be sufficiently large. Immersive cooling with fluids that change phase at low temperatures brings all sorts of issues with safety. It would be very interesting to see if any chips could scale in performance with temperature. Eventually, when chips are cheaper, I'd like to do such testing but sadly now is the time for that to potentially make economic sense. Use 1-10x more power to achieve 1-2x higher hash rate = more profit given the high fixed costs of the chips and general low power requirements. I imagine in the future when I can afford to blow up dozens/hundreds of chips in testing, the fixed costs of the chips will be a fraction of their expected energy consumption over their lifetime and thereby make the process of chilling them a waste of money. |
|
|
|
Very clean Supra. A year or two from now I'd be interested solely to own one as a dyno queen. I'm sticking with AWD for my track rat GLWS, hopefully it goes to someone who will respect the car. Makes me sad when I see trashed Supras although I see fewer and fewer as they get wrecked by people. |
|
|
|
|
Show Posts
