Not sure on power, someone bored enough to synthesize a design and run power estimations?
Don't get fooled by xilinx S6 perf/power, those are the worst of all 45/40nm FPGAs by a long shot; we pretty much only use them because they're cheap-ish and readily available in small qty.
Topic: 1GH/s, 20w, $500 — Butterflylabs, is it a scam? - page 18. (Read 123109 times)
b) only SIII based prototypes exist, once they get enough pre-orders HCIIIs get ordered (about 8wk if you rush assembly... possible.)
or
c) they somehow got a whole bunch of large SIIIs really cheap.
or
d) something entirely different.
or
c) they somehow got a whole bunch of large SIIIs really cheap.
or
d) something entirely different.
You'd agree what we could probably identify (b),(c) from the power usage though, no?
Those packages look like 1mm grid thermally enhanced flipchip BGA ... guesstimating by the 0.1" headers ... 29x29mm FF780
so if these are SIII, pretty much "could be anything from a L50 to a E260"
One thing kinda throwing me off... if those are HCs, 1.1V core doesn't make sense. HCII is 1.2V and HCIII and IV are 0.9V.
As for "what SIII would you need for that"... depends on how much time you spend optimizing, first guess would be 2 unrolled engines with 1 pipeline stage per round, each running @ 250MHz... gut feeling says a L150.
So, random guesses:
a) these are HCIIIs, they prototyped with SIII and forgot to update the label for vcore.
or
b) only SIII based prototypes exist, once they get enough pre-orders HCIIIs get ordered (about 8wk if you rush assembly... possible.)
or
c) they somehow got a whole bunch of large SIIIs really cheap.
or
d) something entirely different.
so if these are SIII, pretty much "could be anything from a L50 to a E260"
One thing kinda throwing me off... if those are HCs, 1.1V core doesn't make sense. HCII is 1.2V and HCIII and IV are 0.9V.
As for "what SIII would you need for that"... depends on how much time you spend optimizing, first guess would be 2 unrolled engines with 1 pipeline stage per round, each running @ 250MHz... gut feeling says a L150.
So, random guesses:
a) these are HCIIIs, they prototyped with SIII and forgot to update the label for vcore.
or
b) only SIII based prototypes exist, once they get enough pre-orders HCIIIs get ordered (about 8wk if you rush assembly... possible.)
or
c) they somehow got a whole bunch of large SIIIs really cheap.
or
d) something entirely different.
i highly agree with you with a combination of C and B.
they got some S3 for cheap.
than sell the product in cost price or a few percents of loss.
when pre-order number goes enough, start the hard-copy progress.
if not, kill this project.
Those packages look like 1mm grid thermally enhanced flipchip BGA ... guesstimating by the 0.1" headers ... 29x29mm FF780
so if these are SIII, pretty much "could be anything from a L50 to a E260"
One thing kinda throwing me off... if those are HCs, 1.1V core doesn't make sense. HCII is 1.2V and HCIII and IV are 0.9V.
As for "what SIII would you need for that"... depends on how much time you spend optimizing, first guess would be 2 unrolled engines with 1 pipeline stage per round, each running @ 250MHz... gut feeling says a L150.
So, random guesses:
a) these are HCIIIs, they prototyped with SIII and forgot to update the label for vcore.
or
b) only SIII based prototypes exist, once they get enough pre-orders HCIIIs get ordered (about 8wk if you rush assembly... possible.)
or
c) they somehow got a whole bunch of large SIIIs really cheap.
or
d) something entirely different.
so if these are SIII, pretty much "could be anything from a L50 to a E260"
One thing kinda throwing me off... if those are HCs, 1.1V core doesn't make sense. HCII is 1.2V and HCIII and IV are 0.9V.
As for "what SIII would you need for that"... depends on how much time you spend optimizing, first guess would be 2 unrolled engines with 1 pipeline stage per round, each running @ 250MHz... gut feeling says a L150.
So, random guesses:
a) these are HCIIIs, they prototyped with SIII and forgot to update the label for vcore.
or
b) only SIII based prototypes exist, once they get enough pre-orders HCIIIs get ordered (about 8wk if you rush assembly... possible.)
or
c) they somehow got a whole bunch of large SIIIs really cheap.
or
d) something entirely different.
ngzhang: Altera SIII is 1.1V core.
oh, i forgot to check altera S3.
so, is it every thing goes clear?
ngzhang: Altera SIII is 1.1V core.
Aye, looked it up since I am not familiar. To ask, do you think that's what they are using? Seems the ones on their boards have a silver housing and the Alter SIII are black. No clue if that's static or if a custom chip from them could be different though. Also, what would be the smallest cell count that could be capable of the ~500MH they are claiming? I notice those chips are not cheap.. Sorry to ask such useless questions but you're just about the most qualified brain about to answer them. ;p
Cheers
ngzhang: Altera SIII is 1.1V core.
on the basis of the pics, this is a very high quality product, with professional design.
it looks like 2 FPGAs with a MCU on a 4 layers PCB. and i see that 3 voltage test points, 1.1V/2.5V/3.3V.
which high-end FPGA uses 1.1V for the core?
not altera S4 (0.9V),
possible xilinx V6 is 1.05V (1.1V is OK),
possible xilinx V5 is 1.05V (1.1V is OK)
it looks like 2 FPGAs with a MCU on a 4 layers PCB. and i see that 3 voltage test points, 1.1V/2.5V/3.3V.
which high-end FPGA uses 1.1V for the core?
not altera S4 (0.9V),
possible xilinx V6 is 1.05V (1.1V is OK),
possible xilinx V5 is 1.05V (1.1V is OK)
V6 isn't going to do 1GH with 20w.
yeah, and i also consider the power on this board is VERY over engineered, 2top and 2botton high-end mosfets for just approx. 9A I-core? (19.8W, 9.9W/each,1.1V, so : 9Amax)
but 2 V5 or V6 can easily get 1.05G/s, that means 500M/each, just implemented 3 cores run @ 170M or 2 cors run @ 250M is ok. very, very, easy on V6.
and, Inaba, no need to remove the heatsinks, just let us know all the other chips on this board is enough for evaluation this board on a technology side.
on the basis of the pics, this is a very high quality product, with professional design.
it looks like 2 FPGAs with a MCU on a 4 layers PCB. and i see that 3 voltage test points, 1.1V/2.5V/3.3V.
which high-end FPGA uses 1.1V for the core?
not altera S4 (0.9V),
possible xilinx V6 is 1.05V (1.1V is OK),
possible xilinx V5 is 1.05V (1.1V is OK)
it looks like 2 FPGAs with a MCU on a 4 layers PCB. and i see that 3 voltage test points, 1.1V/2.5V/3.3V.
which high-end FPGA uses 1.1V for the core?
not altera S4 (0.9V),
possible xilinx V6 is 1.05V (1.1V is OK),
possible xilinx V5 is 1.05V (1.1V is OK)
V6 isn't going to do 1GH with 20w.
on the basis of the pics, this is a very high quality product, with professional design.
it looks like 2 FPGAs with a MCU on a 4 layers PCB. and i see that 3 voltage test points, 1.1V/2.5V/3.3V.
which high-end FPGA uses 1.1V for the core?
not altera S4 (0.9V),
possible xilinx V6 is 1.05V (1.1V is OK),
possible xilinx V5 is 1.05V (1.1V is OK)
it looks like 2 FPGAs with a MCU on a 4 layers PCB. and i see that 3 voltage test points, 1.1V/2.5V/3.3V.
which high-end FPGA uses 1.1V for the core?
not altera S4 (0.9V),
possible xilinx V6 is 1.05V (1.1V is OK),
possible xilinx V5 is 1.05V (1.1V is OK)
^ Cool, sounds like it's easily portable then. There would be no reason to develop a kernel driver anyway, as far as I know.
I guess I might as well send them and e-mail and ask them for the source code.
I guess I might as well send them and e-mail and ask them for the source code.
The demo plan is that I will:
A) connect it to a non-routable development side of the pool, so that the box is unable to communicate with the internet. I will then let it submit shares to the pool and I will have one of the getwork servers in debug mode and I will see what is sent out and what's sent back. As I found no evidence of any wireless communications on the board, and since the computer it's connected to will not be on the internet, it won't have any way of falsifying the shares submitted.
B) I will take a unit home that evening, disassemble and take more robust pictures. I will NOT be removing any heat sinks, however.
C) I will do further testing that evening on my own with a packet analyzer to see what packets are being transmitted, when and where they are going and coming from.
That's the plan as of right now, at any rate.
A) connect it to a non-routable development side of the pool, so that the box is unable to communicate with the internet. I will then let it submit shares to the pool and I will have one of the getwork servers in debug mode and I will see what is sent out and what's sent back. As I found no evidence of any wireless communications on the board, and since the computer it's connected to will not be on the internet, it won't have any way of falsifying the shares submitted.
B) I will take a unit home that evening, disassemble and take more robust pictures. I will NOT be removing any heat sinks, however.
C) I will do further testing that evening on my own with a packet analyzer to see what packets are being transmitted, when and where they are going and coming from.
That's the plan as of right now, at any rate.
Please make sure at some point during the actual mining you run it on a power meter so you can get a measurement under actual mining load.
There are various ways to fake this sort of thing and match one of {rate,power} but not both. (e.g. you could use a big-fast-expensive FPGA to get 1GH/s, but if they did that they couldn't deliver on their advertised price)
Yeah, I forgot to mention that specifically... but in the initial pictures, I had some pics of the Kill-A-Watt meter, but the flash blanked out the LCD screen on it. I will have the Kill-A-Watt and take a better picture while it's in operation.
Quote
Inaba, regarding the driver used to control this box, is it a simple user-space USB driver packaged into the program we can see running on your laptop in the last two picture you posted?
I'd very much like to take a look at the source code. I don't expect to see anything revolutionary, as anything they don't want out can be hidden in the firmware on the device. I'm just really curious to see how its built, and especially so if it requires a kernel driver, which would be more of a pain to get working on ARM.
I'd very much like to take a look at the source code. I don't expect to see anything revolutionary, as anything they don't want out can be hidden in the firmware on the device. I'm just really curious to see how its built, and especially so if it requires a kernel driver, which would be more of a pain to get working on ARM.
I actually mis-spoke a bit, I think, when I called it a driver. It's actually the mining software used to run it... it uses the standard USB driver already present for basic IO. To my knowledge, they will be releasing the source to the whole thing (it's GPL'd I believe, so they'd have to release it if anyone asked anyway).
Inaba, regarding the driver used to control this box, is it a simple user-space USB driver packaged into the program we can see running on your laptop in the last two picture you posted?
I'd very much like to take a look at the source code. I don't expect to see anything revolutionary, as anything they don't want out can be hidden in the firmware on the device. I'm just really curious to see how its built, and especially so if it requires a kernel driver, which would be more of a pain to get working on ARM.
I'd very much like to take a look at the source code. I don't expect to see anything revolutionary, as anything they don't want out can be hidden in the firmware on the device. I'm just really curious to see how its built, and especially so if it requires a kernel driver, which would be more of a pain to get working on ARM.
incredibly interesting thread, subbing for future.
looking forward to seeing test results.
looking forward to seeing test results.
I think lets stop the guessing and let the real results show. The conspiracy theories are getting more and more colorful.
[...] This thing looks to be seriously overdesigned! If they'd just added a cheap Ethernet PHY and port, it would have more than enough processing grunt to connect directly to a pool by itself and mine standalone with no computer.
As far as I can tell that would require a lot more software work. Like building a program for the Atmel chip that fetches work. I think this would, somewhat, be a waste of time, as this is what ckolivas has spent so much time perfecting with cgminer. Also, is there a TCP/IP implementation for this Atmel chip even? Or how'd they actually send packets with the PHY, let alone run a mining application on the board?Not to mention network equipment is a lot more expensive than USB hubs/cables if you want to run a cluster of them.
For future reference, what is the best method to use for removing heatsinks that have been attached with thermal glue (instead of using thermal grease and holding them on with screws)?
If they used a good epoxy adhesive, you might never get them off. At least not without breaking the chips. 
it seems a little silly since that knowledge will become public, assuming they hit their target ship dates
Nah, they've probably removed any silkscreening on the tops of the chips. Bet they're blank underneath. Don't count on learning anything from removing the heatsinks. If they don't want you to know what used to be silkscreened there, it's only going to cost them a few pennies worth of solvent and about five seconds of additional assembly time.
IMO it's a bit pointless to try to prove, with the demonstration, that BFL didn't pack the board with some expensive chips to make it look like they achieved the $/MH figure they did. Only with time will we get more and more certain whether this is the case or not.
Conspiracy theorists will never run out of ideas.
Conspiracy theorists will never run out of ideas.
Just to make sure the conspiracy theorists run out of ideas.. assuming they have several units, will they let you choose a unit at random?
Jump to: