Community Miner Design Discussion

[Searing]

Just throwing this out there, but I won't use that part unless I can't find something that was actually designed for board mounting instead of spliced inline on a wiring harness. Heck of a big hassle, not worth talking about.

side note:


sure does not help us titan users with the 2 plug psu y adapters and the overclocking of those machines (2x) by knc to 'ahem' stop the returns via RMA's imho

by the time a fuse went with one of these babies it is too late indeed


back to regularly scheduled community miner design discussion

[navigatrix]

less than 500w per tera is really a great result, i cna even begin to mine again bitcoin directly with that ratio

i hope this time they do not make huge miners with 10 tera and kill the casual mining with 3kw at home

Wha'?
The Antminer S7 is <280W per Terahash. @Amph, 500w/TH is S5 specs.
Did you mean 50?

I'd be happy if the finished product (with Bitfury chips) was between 90 - 110W per Tera. 

[philipma1957]

less than 500w per tera is really a great result, i cna even begin to mine again bitcoin directly with that ratio

i hope this time they do not make huge miners with 10 tera and kill the casual mining with 3kw at home

Wha'?
The Antminer S7 is <280W per Terahash. @Amph, 500w/TH is S5 specs.
Did you mean 50?

I'd be happy if the finished product (with Bitfury chips) was between 90 - 110W per Tera.  

goal is 70 to 120  as stated by sidehack .  that wide spread is very helpful for adding life to the miner.

[sidehack]

Say, got a question. Who's the best domestic quick-turn prototype PCB fab? We've gotten prototype PCBs made by SEEED, but they're in China - which is grating for obvious patriotic reasons and because it makes shipping suck for any kind of speed.

[NotFuzzyWarm]

Say, got a question. Who's the best domestic quick-turn prototype PCB fab? We've gotten prototype PCBs made by SEEED, but they're in China - which is grating for obvious patriotic reasons and because it makes shipping suck for any kind of speed.

I've always used Sunstone Circuits  http://www.sunstone.com/pcb-products along with their PCB123 software.
Home for them is in Oregon and with turnaround of 24hrs (if ya pay for it) or a week production has to be in the USA.
Great company.

[sidehack]

We've used Sunstone for all our PCB batches that weren't the DPS2K boards (we opted for someone who specialized in heavy copper) but I've not looked into their quick-turnaround service. I did recently read something of theirs on 2-layer prototyping, but not 4-layer.

[jstefanop]

Say, got a question. Who's the best domestic quick-turn prototype PCB fab? We've gotten prototype PCBs made by SEEED, but they're in China - which is grating for obvious patriotic reasons and because it makes shipping suck for any kind of speed.

I use http://www.pcbunlimited.com

Their Panel share is ridiculously cheap for prototyping if you can wait 5 days, but their quick turn is not too bad either.

[NotFuzzyWarm]

Say, got a question. Who's the best domestic quick-turn prototype PCB fab? We've gotten prototype PCBs made by SEEED, but they're in China - which is grating for obvious patriotic reasons and because it makes shipping suck for any kind of speed.

I use http://www.pcbunlimited.com

Their Panel share is ridiculously cheap for prototyping if you can wait 5 days, but their quick turn is not too bad either.

Panel Share would be good for controller boards but the restriction of 1oz copper makes it a no-go on hashboards.

[sidehack]

And since I'm not making any controller boards, at least for this project...

Just doing a quick check, Sunstone quickturn looks like 1oz copper and I don't see a way to change that on the online form. And for what I want we're looking at around $900 for half a dozen boards. $900 from Full-feature will get me a heavier copper, ten boards, and only a couple days later. SEEED will get me heavier copper still, and ten boards, and a couple days later, for $300. That's gonna be hard to argue with.

[philipma1957]

And since I'm not making any controller boards, at least for this project...

Just doing a quick check, Sunstone quickturn looks like 1oz copper and I don't see a way to change that on the online form. And for what I want we're looking at around $900 for half a dozen boards. $900 from Full-feature will get me a heavier copper, ten boards, and only a couple days later. SEEED will get me heavier copper still, and ten boards, and a couple days later, for $300. That's gonna be hard to argue with.

So seed is the most copper at the best price?

Yeah sounds good.

Let us know when you order.

That way we can all ask are we there yet?

[sidehack]

Well this is a sub-scale Avalon-based proof of concept commissioned by Kilo, though I would like to show it off once it works. Very unlikely to become an actual product for anyone but him.

[Mitak]

There are no news from Bitfury , avalon is the second choice did you both with kilo managed to get chip docs or you  just do reverse engineering ?

[2112]

I moved sidehack's reply to this thread, because the original thread is going on back-burner.

Novak chose an LPC11u23 ARM, which is USB-capable and has a variety of ports and ADCs. It's in the same family as the ARM on the Avalon Nano and Ava6 control board (which I believe is LPC11u14? Going from memory on all these part numbers). I've worked with 8051 programmed from Windows, but not ARM programmed from Linux. He had a Linux toolchain set up, and before he left I believe he shifted it over to one of the general shop machines rather than his own box. I figure on utilizing the USB bootloader supported by the chip, so I can write code to it without requiring extra hardware - that also makes firmware updates in the future more possible.

So, I reckon if someone is pretty good with ARM, and maybe also c for some cgminer drivers, and has time to work on it in the next couple weeks, let me know.

I would guess that Avalon designers had chosen LPC11U14 because of the popularity of Embedded Artists' cheap LPCXpresso development kits. http://www.embeddedartists.com/products/lpcxpresso

They are about EUR20 in Europe, about USD18 in the United States.

LPC11U14 is now obsolete, replaced by LPC11U24, which is also soon getting obsoleted. They have a cute break-away debugging sub-circuit. It can be used to debug the included micro-controller or after breaking of can be used to debug an external board with similar compatible NXP micro-controller.



It is also supported by the free version of the LPCXpresso toolchain, no need to pay $500 for the full version. I took a quick look at the Avalon's firmware source posted on Github. They seem to be using Keil's toolchain which is very expensive. Edit: about USD5000, depending on the license details. Edit2: although Keil recently started supplying some free/subset version of their toolchain for very small chips, which I think this one will qualify.

It is going to be a tight on this MCU: 50MHz max clock, 24-32kB FLASH, 2kB RAM, 1 UART, 1 I2C, 2 SPI.

[sidehack]

Not sure what you mean by tight. 50MHz with 2k RAM and 32KB code space is huge for what I need it to do.

[2112]

Not sure what you mean by tight. 50MHz with 2k RAM and 32KB code space is huge for what I need it to do.

It looks tight to me. But my experience with embedded systems is mostly related to medical-grade devices or with high reliability demands.

Couple of weeks ago we've discussed daisy-chain versus star designs. With only 2 SPIs true star is doable for at most 2 mining chips.

I also look at it less from the production system angle and more from testing and qualification system angle. Personally, I would grossly overspec the MCU in first system to make sure that I can do serious reliability testing and essentially reverse-engineering the chip I/O protocol. Then in production I would use some pin-compatible MCU with much lower specs.

Or maybe design the board that the MCU could be tri-stated out of the real chip control and provide some pins to hook up external MCU with much better specs. Those pins could be left unpopulated in the shipping version to save costs.
 

[sidehack]

I don't think any miner has ever used more than one port out of the micro. Most chips are daisy-chained; ASICMiner used address-decoded chip selects on a common SPI bus (which I really like). It looks like this and the previous generation of Bitfury chips use a comm multiplexer that probably only needs one bus connection to the micro and then breaks out a couple dozen data pairs that would each go to a single chip.

My experience with embedded was writing 8-bit assembler, where I did in about 65 bytes what the C compiler wanted about 1k to accomplish. Most of what this chip will have to do are interrupt-based timer routines, and relaying data packets back and forth. The only limitation might be RAM buffering USB data packets, but the clock is plenty fast and there's probably an order of magnitude more code space than should be needed.

[2112]

I don't think any miner has ever used more than one port out of the micro. Most chips are daisy-chained; ASICMiner used address-decoded chip selects on a common SPI bus (which I really like). It looks like this and the previous generation of Bitfury chips use a comm multiplexer that probably only needs one bus connection to the micro and then breaks out a couple dozen data pairs that would each go to a single chip.

We had this discussion about a half-year ago:

2) Star topology vs daisy-chain topology. On this I prefer star because the ICs need to be running at the edge of failure (thermal or noise), otherwise the project is not competitive.

2) I also prefer star over chain for comms.

I don't know exact circumstances of your decision-making and MCU-selection process. I'm not changing my opinion. To me daisy-chaining mining chips is still "tight" and "skating the edge of failure". Personally, I would never compete with Chinese by out-lame-ing their lame designs.

My experience with embedded was writing 8-bit assembler, where I did in about 65 bytes what the C compiler wanted about 1k to accomplish. Most of what this chip will have to do are interrupt-based timer routines, and relaying data packets back and forth. The only limitation might be RAM buffering USB data packets, but the clock is plenty fast and there's probably an order of magnitude more code space than should be needed.

IIRC your experience was 8051. This is not representative with respect to modern architectures like AVR or ARM. With those the difference between hand-coded assembly and C/C++ is about 10%-20% at most. And not always hand-coded assembly wins, especially on ARM.

In my experience the C/C++ firmware gets abnormally big only when the programmer doesn't understand the toolchain: pulls needless library routines like malloc() or printf(), does unnecessary exception handling, etc.

[sobe-it]

I have used SEEED for my nanofury boards..... it was a little wait but was by far the cheapest and their metal solder paste stencil was dirt cheap. They also give coupons that you can use on the next order, or at least I know I got one but it expired a few months ago.

[sidehack]

At least according to those quotes, it looks like we agree on comms topology.

And like I said, there should still be plenty of code space for the not-a-whole-lot I intend to do on this micro. It's not gonna be like an AM Blade or Cube where I have an ethernet stack and stratum implementation built in. If there's enough RAM to buffer USB packets that get relayed out to the bus, it should be fine.

[chiguireitor]

...

IIRC your experience was 8051. This is not representative with respect to modern architectures like AVR or ARM. With those the difference between hand-coded assembly and C/C++ is about 10%-20% at most. And not always hand-coded assembly wins, especially on ARM.

...

This is because ARM has thumb mode, which does mad compression on code on a static manner, and can rearrange code blocks so you get large swathes of code compressed on thumb code.