Topic: Can anyone tell me what chip is used in BFL single? - page 3. (Read 12750 times)

These are totally unrelated.

The JTAG interface lets you ask the chip "what chip are you" -- IDCODE.  All Xilinx chips have this capability, and it cannot be disabled.  I don't know about altera, but I suspect it's the same deal.  This has nothing to do with the bitstream/firmware.

^ Are we talking about the same thing here?

As far as I can tell, the DMCA reference refers to its prohibition of breaking a copy-prevention mechanism in a piece of electronic equipment that you own.

It's actually sensible, very little known, single paragraph in a single directive.
It puts the effort of proof illegality to the claimee, and emphasizes that it may not cause extra costs for ISPs, or make them responsible for proactive monitoring, or to break laws (privacy usually) to cooperate with those.

In a case of movie for example, MediaSentry has to send in not only what was copied, but they have to also proof it was illegal for the user, which is almost impossible to know as certain countries allow private copies, other countries allow copying for backup purposes (afaik from any sources). etc.

You can check the EU ecommerce directive article 14 about this. It's actually quite hard to understand due to the wording, and some small companies opt to translate it to the worst possible, even worse than SOPA (blanket full censored rights to any party for any reason, without any kind of prejudice. Including your own logos etc.) Obviously those outlets are simply morons.

I think SOC is a good possibility, could also be some kind of older powerhouse FPGA that they got a really good deal on from a closed out design shop or something

I thought more about it and I think BFL uses some kind of SOC device, you know those FPGAs with integrated uC among other things. At first it seems counter intuitive, but if you really look at what these devices are capable of you start to re-think it.
Some of those have SIMD and/or MIMD instruction sets, in some way similar to GPUs.

Mining can only partially benefit from discrete logic and many cases can probably be done more efficient with the right ALU. Those present in the Spartan-6 we ruled all pretty much out but there could very well be a device which could be very well suited.
The market is large and there are lots of players involved, but I am almost certain that there lurks a device, lets call it a hybrid between a FPGA and a GPGPU (shift instructions!) which would match the performance of the BFL devices.

Look at those Altera APEX devices... aren't those like the thing I talked about?

In case I have been right, BFL,  hate me  

As I recall there is a EU equivalent to DCMA too, probably not as restrictive, but at least it enabled the researchers (they are in Germany) to do what they are doing.


Right, we may not even need to go as far.


Is there some chips identifiable as a flashrom (SPI, I2C, etcc..) on the boards?
As I recall most high performance FPGAs don't include flash rom. So it could be possible to find out which chip it is just by knowing from which pins the bitstream is loaded.
There also is a AVR32 on there which I can make out from the pictures available, that could be also used to store the bitstream.

I think once one of these singles gets into the right hands someone will find some meaningful data from the bitstream that will indicate what the chip is. The bitstream itself isn't really that important to most of us (I don't particularly care about the programming of the chip, but perhaps some of the hxc programmers around here might).

I'm curious if this project was sired by someone among us that spun off a project from an idea(s) on this forum. I remember seeing some pretty cool FPGA implementation ideas and the manual logic placement optimizations.

Well sadly in the US it would be illegal.  DMCA makes circumventing encryption for ANY REASON even a legit one a crime.  Granted the law is an abomination and should be repealed but as of today it is on the books.

As discussed, they are almost certainly using a cheap source of old-gen 65nm FPGAs: https://bitcointalksearch.org/topic/m.769355

He is probably right with JTAG Probing, modern FPGAs include pretty good bitstream encryption.

But it is as with any DRM, futile for the long run.
just one example: http://it.slashdot.org/story/11/07/21/1753217/fpga-bitstream-security-broken

The nice thing about it: It's perfectly legal to do it, basically the same thing as jailbreaking a phone. Now sharing the bitstream would be another story.
So BFL is safe as long as nobody with access to a decent lab comes along.

I thought that your bitcoin products were just incidental overlap with your real business of government-contracted supercomputers and medical imaging devices...

BTW, when will those imaging drivers be available?

A custom chip wouldn't need to be a "trade-secret", because no one would be able to get the same chip anywhere else. So clearly the BFL Single is using an off-the-shelf chip. That's the kind of "secret" that will not stay secret.

I didn't say you necessarily need the catalyst driver to utilize the GPUs, but I said it was necessary because writing your own OpenCL implementation for a GPU would make this project go from very, very difficult to insanely difficult. AMD has a team of about 80 developers working full-time on their Catalyst driver (I am not kidding you).
The open source Radeon HD driver (which is about four years old) just recently was able to run a proof-of-concept OpenCL application: http://phoronix.com/forums/showthread.php?69064-Using-OpenCL-with-the-radeon-driver-to-print-Bitcoins-is-this-posible#post252304
Probably at 1% the speed of the Catalyst driver.

Developing an OpenCL implementation isn't that hard, developing a well-performing OpenCL implementation is! Just look at the radeon driver (open source Linux driver for the Radeon HD GPUs). AMD has a man dedicated to work on this full-time, and there are plenty of volunteer developers, and it doesn't support a fraction of the features Catalyst does, and it performs at a fraction of what the Catalyst driver does.

Building a custom GPU motherboard and writing a driver that only achieves 50% the performance of the Catalyst driver is hard and a waste of time (and it wouldn't be able to churn out 800 MH/s in this case).
Necessary, no. Necessary if you want a device performing as well as the Catalyst driver without spending 100,000 man hours developing a custom Radeon HD driver with OpenCL capabilites, yes.

I'm telling you man, a GPU driver is a lot more complex than most people think. It's basically a small operating system running on the GPU: memory manager and process scheduler are both a necessary part of a GPU driver. It's basically a massive serial-to-parallel compiler (translating a series of Direct3D or OpenGL instructions into tens of thousands of threads with instructions the GPU chip can understand and execute in parallel). It's actually quite fascinating. This covers the very basics of a GPU driver: http://fgiesen.wordpress.com/2011/07/01/a-trip-through-the-graphics-pipeline-2011-part-1/

Starting in 2008, AMD released all documentation for some of their GPUs (1) (yes, that's about 2500 pages of hardware, firmware and software specification), after the open source Linux community had encouraged them to do so because "if we have the specifications some hacker will come around and develop a much better driver than the proprietary one". Today reads 2012, and the open source driver still isn't really usable for the average gamer (who buys a $200-$800 graphics card and then proceeds to use a driver that only performs at 20% the speed of the Catalyst driver?)

Tosh. You would need an OS to use the catalyst drivers -- but you don't necessarily need catalyst to use the chip, and you wouldn't need an OS to directly interface to the chips if you were given their specification (or reverse engineered it) [on this note: do the GPU chips directly implement PCIE, or is there an additional interface chip for this -- I suspect the former, meaning you would simply need a USB to PCIE interfacing chip...] -- which is however highly unlikely since I don't think ATI would publish this sort of data / reverse engineering would be tedious. You would be wasting a lot of time reimplementing what the catalyst drivers already do (although you would only need a small part of the functionality for bitcoin mining) once you had this data as well. In other words it is pointless to do the above, but it is not actually necessary to use an OS.

^ I think we can rule this out even disregarding the power consumption. It would simply not be possible to do this without too much time and effort (orders of magnitude more).

I actually doubt anyone but AMD employees would be able to do this, in which case it'd probably be illegal because of the NDA agreement they've signed (unless, of course, AMD gave them permission).

They'd still need an x86 CPU (the Catalyst driver isn't available compiled for anything else but x86), and either Windows or a Linux distribution running on that CPU (graphics cards don't work without an operating system). They'd need system RAM (for the CPU) and video RAM (for the GPU). They would basically be designing their own x86 motherboard from the ground-up, with embedded graphics and a custom OS running on this motherboard. I think I can safely say that this would simply be too expensive, too time-consuming and too difficult to make.

Yeah, when I bother to actually do the math, it becomes clear that you are right. I can take 400MHs out of each of my 5970 cores @ 840MHz, give or take. The full board, with 2 cores, sucks ~250W (RAM downclocked) and I'm sure a lot could be shaved off by removing everything but the bare essencial to mine. That 'not essencial' category will fit a lot of the GPU components, and thus the power draw has a good margin to be reduced, but to 80W from 250W is probably just impossible.

Pretty low, the wattage is much lower than a GPU. Looks like an outdated FCPGA chip that was bought in bulk for a dirt cheap price.

What are the chances of this being some ATI GPU or other stripped of everything else, just doing opencl? Kind of makes sense from both a price and performance standpoint

My feeling is that Bitcoin mining is just a small stepping stone on their way to greatness. So they probably will not want to make everything open-source at the outset.

That's how business work, remember? Even though all of Facebook's datacenter plans are open-source (http://opencompute.org/), they certainly weren't at the initial outset of the project. And that is only one example.

I bet ArtForz would disagree.

There are ways to make the chips talk just as there are ways to make people talk