Merrick6 for Bitcoin

[yohan]

I wasn't suggesting Merrick1 for Bitcoin for those reasons. The cut down Merrick6 is a better board initially to talk about. Doing a derivate design for Bitcoin it could be possible to up the power supplies to support all devices by simply say removing DDR chips to make space. The 12A regulator circuit we use could fit into the space taken by 2 DDR chips and we could end up with 6/7 FPGAs each with 12A of core voltage.

That's all viable to do and that sort of design change could be done in a few days maybe a week. A prototype could be turned in about another week if we pushed although that fast gets expensive.

[1714823813]

1714823813

[1714823813]

1714823813

[1714823813]

1714823813

[Gomeler]

I wasn't suggesting Merrick1 for Bitcoin for those reasons. The cut down Merrick6 is a better board initially to talk about. Doing a derivate design for Bitcoin it could be possible to up the power supplies to support all devices by simply say removing DDR chips to make space. The 12A regulator circuit we use could fit into the space taken by 2 DDR chips and we could end up with 6/7 FPGAs each with 12A of core voltage.

That's all viable to do and that sort of design change could be done in a few days maybe a week. A prototype could be turned in about another week if we pushed although that fast gets expensive.

That would be an interesting option. If possible, a 4-pin Molex or SATA power connector might be more useful. I know a lot of my PSUs have 1 floppy power connector but a dozen SATA and 4-pin connectors.

[DeathAndTaxes]

Have you looked at other board offerings?

Various 2x Spartan 150 boards are available for ~$500
One 4x Spartan 150 board is available for ~$1000

Your price point is roughly double and the other Spartan boards face stiff competition from BFL which acheives roughly double the performance (MH/$).  Your board would be in the ballpark of ~450 MH/s for $1000 vs BFL Single ~820 MH/s for $600.

A dual Spartan board for $1000 isn't going to fly. 

[DeathAndTaxes]

I wasn't suggesting Merrick1 for Bitcoin for those reasons. The cut down Merrick6 is a better board initially to talk about. Doing a derivate design for Bitcoin it could be possible to up the power supplies to support all devices by simply say removing DDR chips to make space. The 12A regulator circuit we use could fit into the space taken by 2 DDR chips and we could end up with 6/7 FPGAs each with 12A of core voltage.

That's all viable to do and that sort of design change could be done in a few days maybe a week. A prototype could be turned in about another week if we pushed although that fast gets expensive.

That would be an interesting option. If possible, a 4-pin Molex or SATA power connector might be more useful. I know a lot of my PSUs have 1 floppy power connector but a dozen SATA and 4-pin connectors.

Or PCIe 6 pin connector.  Most ATX PSU group all the MOLEX, SATA, and floppy connectors to a single rail because they are generally low current.   Even single rail designs will use current limiting circuit for safety.  So connecting a large number of boards to MOLEX connectors may be problematic.  A 1000W PSU may have 980W of 12VDC but it may only have <200W on the rail powering MOLEX connectors.  Even a single rail design will have some over current circuit on each "virtual rail" limiting the usable power on MOLEX connectors.

[yohan]

When I say disk connector that's hard disk i.e. Molex. Not always manufactured by Molex they are usually rated at 5 to 7A but can deliver a lot more under stress. We have see wires melt before these connectors fail and that's 10's if not 100's of amps.

On Merrick6 we usually fit either a 7A or 10A fuse on the disk drive connector which is way more than needed for the Bitcoin version. Similarly the 12V from the PCIe has a 2.6A fuse which is a reasonable limit for PCIe operation. If customers want this latter fuse can simply not fitted when we build boards forcing all the current through the DD connector.

[DeathAndTaxes]

Similarly the 12V from the PCIe has a 2.6A fuse which is a reasonable limit for PCIe operation. If customers want this latter fuse can simply not fitted when we build boards forcing all the current through the DD connector.

That likely is optimal.  While 30W on PCIe board is within the spec someone building a rig with 8 board likely will overload the PCIe bus. We have seen this w/ GPUs.  The solution is to use powered PCIe extenders to reduce the load on motherboard VRMs.  Having a board draw no 12VDC from the PCIe connector bypasses that entire issue.

The optimal board would be
PCIe 1x (allows max # of boards per rig)
PCIe 6 pin power connector
at least 4 (preferrably 6) Spartan-6 150 FPGAs
"beefy" 12VDC power supply.

A 6 Spartan board using current bitstreams could acheive ~1.2 GH/s per board @ ~50 - 60W.
Using the hand placed 3 loop bitstream (will need to be purchased) a 6 spartan board could achieve ~1.5 GH/s (possibly up to 1.8 GH/s with improvements) @ ~60W to 70W.

[Transisto]

I think there may be a lot of people interested in Mini-rig like hardware but would not trust BFL with this much money.

The delay of supposedly 3 month is also a big drawback.

If you can solve the later and have a reasonable price I'm in.

[Gomeler]

Similarly the 12V from the PCIe has a 2.6A fuse which is a reasonable limit for PCIe operation. If customers want this latter fuse can simply not fitted when we build boards forcing all the current through the DD connector.

That likely is optimal.  While 30W on PCIe board is within the spec someone building a rig with 8 board likely will overload the PCIe bus. We have seen this w/ GPUs.  The solution is to use powered PCIe extenders to reduce the load on motherboard VRMs.  Having a board draw no 12VDC from the PCIe connector bypasses that entire issue.

The optimal board would be
PCIe 1x (allows max # of boards per rig)
PCIe 6 pin power connector
at least 4 (preferrably 6) Spartan-6 150 FPGAs
"beefy" 12VDC power supply.

A 6 Spartan board using current bitstreams could acheive ~1.2 GH/s per board @ ~50 - 60W.
Using the hand placed 3 loop bitstream (will need to be purchased) a 6 spartan board could achieve ~1.5 GH/s (possibly up to 1.8 GH/s with improvements) @ ~60W to 70W.

These would be great drop-in replacements/upgrades for GPUs. If a slim enough heatsink could be attached to keep the cards as a single slot design I'd happily dump all my GPUs for these. Hell, even a double slot design or bigger would work on PCIe extenders, which I already use to cram more GPUs per board.

[BR0KK]

Is defenetly would consider a board with 5 to 6 spartans for my gpu miner.

Slim aka single slot cooling would be great.

U could implement an USB interface or something else, so that we could use the board without that pice interface.

[yohan]

We can actually do a USB interface slightly indirectly on this board. There is a debug port on the board that we have FT232 adaptor for and that can plug into a PC as a serial port. We were going to do it directly on the board but we ran out of board area for the chip + connector.

[Glasswalker]

I don't see a problem with the PCIe interface.

If they can fit 6+ (or seeing some of their others, 24+) fpgas on a board, and find a way to power and cool them. A doublewide 24x FPGA board on a pcie which draws 250watt over a pair of 8pin PCIe power connectors. You can put 2 in a standard motherboard, allowing for 48FPGAs per rig. With current bitstreams that's just under 10GHash/s per rig.

Also consider, PCIe allows for MUCH more bandwidth to the card (which is not important at all for bitcoin). But it's the standard right now for FPGA based HPC accelerators. Meaning if you have a PCIe card which works both for bitcoin, and for doing crypto acceleration, or other HPC applications, you now have decent resale value. Current bitcoin mining solutions aren't really good for much else (maybe as a "play" development board, but not for any real serious applications) because of the low IO count, and poor throughput (and lack of available RAM)

I would love to see a 4chip, 8chip and 24chip card lineup, all PCIe, with decent cooling/power arrangements. and offering pricing under $1/MHash (preferably closer to $0.75/MHash or lower, at least on the higher density cards) Even more appealing would be a doublewide card, that has the comm/control chips on the PCIe card, and power supply connectors enough for 24 FPGAs, and can take up to 4 daughter cards, each holding 6 FPGAs. That would allow for high density, and a nice scalability path, to start with a single module on a card for cheap, and add modules over time. Much more approachable than jumping from smaller modules to a $15K purchase for a mini-rig.

[rjk]

I don't see a problem with the PCIe interface.

If they can fit 6+ (or seeing some of their others, 24+) fpgas on a board, and find a way to power and cool them. A doublewide 24x FPGA board on a pcie which draws 250watt over a pair of 8pin PCIe power connectors. You can put 2 in a standard motherboard, allowing for 48FPGAs per rig. With current bitstreams that's just under 10GHash/s per rig.

Also consider, PCIe allows for MUCH more bandwidth to the card (which is not important at all for bitcoin). But it's the standard right now for FPGA based HPC accelerators. Meaning if you have a PCIe card which works both for bitcoin, and for doing crypto acceleration, or other HPC applications, you now have decent resale value. Current bitcoin mining solutions aren't really good for much else (maybe as a "play" development board, but not for any real serious applications) because of the low IO count, and poor throughput (and lack of available RAM)

I would love to see a 4chip, 8chip and 24chip card lineup, all PCIe, with decent cooling/power arrangements. and offering pricing under $1/MHash (preferably closer to $0.75/MHash or lower, at least on the higher density cards) Even more appealing would be a doublewide card, that has the comm/control chips on the PCIe card, and power supply connectors enough for 24 FPGAs, and can take up to 4 daughter cards, each holding 6 FPGAs. That would allow for high density, and a nice scalability path, to start with a single module on a card for cheap, and add modules over time. Much more approachable than jumping from smaller modules to a $15K purchase for a mini-rig.

The problem is, if you design for bitcoin then the product instantly becomes pretty much worthless for other computations. Bitcoin is all about density, and it can achieve good density because it has no requirement for high speed (and therefore carefully routed) communications. Nor does it need memory. Other applications want lots of high speed, high bandwidth communications and lots of memory, which is more difficult to design for and tends to decrease density.

I think this single slot x1 formfactor is great, and I don't see any reason to use USB when you already have PCIe. Of course it would be nice to have more devices on a card, but very careful attention must be paid to power design.

[Glasswalker]

I agree, my point was that the guys from Enterpoint seem to already have the design under control with the high bandwidth. If they were to modify it slightly (no major overhaul) to get the price point down, they could hit a sweet spot where it's applicable (and cost effective) for bitcoin, while still carrying over much of the value for application in other spaces too.

The main thing I want to see is smooth scalability. Something that allows very high density, but without the "choppy" growth offered with solutions like the Mini-Rig. Something allowing more granular growth. But still offers close to their cost/performance ratio would trash the competition (especially if they can provide more consistant customer service, and meet their estimated deadlines reliably).

[Zubilica]

if you can sale this at 1k EUR , i'll take 2 : -> A 6 Spartan board using current bitstreams could acheive ~1.2 GH/s per board @ ~50 - 60W.

[yohan]

In addition to this product see my new post on quad XC6SLX150 board.

[Jaryu]

out of curiosity, how much in USD is that 24 x Spartan6 card?

[yohan]

Merrick3 sells for $10K. It's a complete different product level than needed for Bitcoin so there lot of work in this design. There 15K via holes in this board and roughly 7.5K connections that are all hand routed for quality of result. Just try imaging how you would arrange chips and make this number of wire connections. It's not easy.

[wogaut]

Similarly the 12V from the PCIe has a 2.6A fuse which is a reasonable limit for PCIe operation. If customers want this latter fuse can simply not fitted when we build boards forcing all the current through the DD connector.

That likely is optimal.  While 30W on PCIe board is within the spec someone building a rig with 8 board likely will overload the PCIe bus. We have seen this w/ GPUs.  The solution is to use powered PCIe extenders to reduce the load on motherboard VRMs.  Having a board draw no 12VDC from the PCIe connector bypasses that entire issue.

The optimal board would be
PCIe 1x (allows max # of boards per rig)
PCIe 6 pin power connector
at least 4 (preferrably 6) Spartan-6 150 FPGAs
"beefy" 12VDC power supply.

A 6 Spartan board using current bitstreams could acheive ~1.2 GH/s per board @ ~50 - 60W.
Using the hand placed 3 loop bitstream (will need to be purchased) a 6 spartan board could achieve ~1.5 GH/s (possibly up to 1.8 GH/s with improvements) @ ~60W to 70W.

I'd definitely consider buying the 6 Spartan version.