Topic: Tabula & FPGA (Read 2199 times)
That's true. TSMC is just ramping 28nm and it's occupied with AMD, Nvidia, etc. fighting for its chips. And from what I can gather, ngzhang and ztex are basically supply constrained on their LX150-based boards. 2013 would be a good year. Get some ROI on the LX150 and upgrade in 2013!
I would imagine the FPGA guys like ngzhang and ztex are already planning if not already in development.
I wonder if they are. Sourcing 28nm parts (even single units for testing & design) is still tough. In 6-9 months it should be much easier. By 2013 28nm will start to displace 45nm parts as the FPGA companies want to move customers to the faster, cooler, smaller, cheaper chips.
Usually each generation has a low end, medium, and high end chip. The high end chips tend to be loaded with lots of gee-wiz bells and whistles which take up lots of silicon real estate but don't do crap for mining. This is why Spartan-6 despite being a low end chip is economical for mining.
Artix-7 is likely going to be the "next gen" chip for mining.
http://www.xilinx.com/products/silicon-devices/fpga/artix-7/index.htm
It is low end part like Spartan-6. Eventually (supply, demand, yields, all that good stuff) it should be available at sub Spartan-6 prices. So we are talking <$150 maybe <$120 per chip.
Its largest model has 350K LUTs vs 150K on Spartan which should make 3 hashes per clock possible (or at least 2). If routing w/ 4ns critical path is possible that means 250 Mhz * 3 = 750 MH/s. Granted that is very optimized so first versions won't be that good. It will take some fine tunning of bitstreams.
To put that into perspective if
* Artix-7 350K LUT is available for <$150
* It can be clocked @ 250 MHz with fully unrolled hashing bitstream.
* You can fit 3 fully unrolled SHA-256 double loops in 350K LUTs
It would be possible to build a dual chip board getting ~1.5GH/s @ ~40W for ~$500 .... someday.
Artix-7 is likely going to be the "next gen" chip for mining.
http://www.xilinx.com/products/silicon-devices/fpga/artix-7/index.htm
It is low end part like Spartan-6. Eventually (supply, demand, yields, all that good stuff) it should be available at sub Spartan-6 prices. So we are talking <$150 maybe <$120 per chip.
Its largest model has 350K LUTs vs 150K on Spartan which should make 3 hashes per clock possible (or at least 2). If routing w/ 4ns critical path is possible that means 250 Mhz * 3 = 750 MH/s. Granted that is very optimized so first versions won't be that good. It will take some fine tunning of bitstreams.
To put that into perspective if
* Artix-7 350K LUT is available for <$150
* It can be clocked @ 250 MHz with fully unrolled hashing bitstream.
* You can fit 3 fully unrolled SHA-256 double loops in 350K LUTs
It would be possible to build a dual chip board getting ~1.5GH/s @ ~40W for ~$500 .... someday.
Thanks for the nice analysis! Let's see how fast the 28nm ramps. I would imagine the FPGA guys like ngzhang and ztex are already planning if not already in development.
Usually each generation has a low end, medium, and high end chip. The high end chips tend to be loaded with lots of gee-wiz bells and whistles which take up lots of silicon real estate but don't do crap for mining. This is why Spartan-6 despite being a low end chip is economical for mining.
Artix-7 is likely going to be the "next gen" chip for mining.
http://www.xilinx.com/products/silicon-devices/fpga/artix-7/index.htm
It is low end part like Spartan-6. Eventually (supply, demand, yields, all that good stuff) it should be available at sub Spartan-6 prices. So we are talking <$150 maybe <$120 per chip.
Its largest model has 350K LUTs vs 150K on Spartan which should make 3 hashes per clock possible (or at least 2). If routing w/ 4ns critical path is possible that means 250 Mhz * 3 = 750 MH/s. Granted that is very optimized so first versions won't be that good. It will take some fine tunning of bitstreams.
To put that into perspective if
* Artix-7 350K LUT is available for <$150
* It can be clocked @ 250 MHz with fully unrolled hashing bitstream.
* You can fit 3 fully unrolled SHA-256 double loops in 350K LUTs
It would be possible to build a dual chip board getting ~1.5GH/s @ ~40W for ~$500 .... someday.
Note that is just back of napkin estimate based on the the relative performance of the chips. Until someone gets one and starts trying routing no way to know for sure but it should be > 2 MH / $ and > 30 MH / W.
Artix-7 is likely going to be the "next gen" chip for mining.
http://www.xilinx.com/products/silicon-devices/fpga/artix-7/index.htm
It is low end part like Spartan-6. Eventually (supply, demand, yields, all that good stuff) it should be available at sub Spartan-6 prices. So we are talking <$150 maybe <$120 per chip.
Its largest model has 350K LUTs vs 150K on Spartan which should make 3 hashes per clock possible (or at least 2). If routing w/ 4ns critical path is possible that means 250 Mhz * 3 = 750 MH/s. Granted that is very optimized so first versions won't be that good. It will take some fine tunning of bitstreams.
To put that into perspective if
* Artix-7 350K LUT is available for <$150
* It can be clocked @ 250 MHz with fully unrolled hashing bitstream.
* You can fit 3 fully unrolled SHA-256 double loops in 350K LUTs
It would be possible to build a dual chip board getting ~1.5GH/s @ ~40W for ~$500 .... someday.
Note that is just back of napkin estimate based on the the relative performance of the chips. Until someone gets one and starts trying routing no way to know for sure but it should be > 2 MH / $ and > 30 MH / W.
I see 
Fancy technology, lots of I/O, not so many LUTs. That can't be cheap = poor MH/$. Move along...
I saw the news as well. 22nm is so cutting edge I can't imagine that FPGAs built on that node would be cost effective for mining atm. Even 28nm aren't cost-effective yet.
I would love to have Intel build me some custom ASICs on the 22nm node. Too bad my last name is not Otellini...
I would love to have Intel build me some custom ASICs on the 22nm node. Too bad my last name is not Otellini...
Just as i thought, No one knows S**T
o well
o well
After the announcement that this company is using Intel's 22nm tech that we knew for a year i was wondering if anyones looked further into making H/W with anything they produce.
With FPGA's looking the logical choice for the future, Teaming up with intel to work on the 22nm versions is only going provide some serious mega hasing, Maybe even giga hashing???
ANy info/advice on this company or what you know is good
With FPGA's looking the logical choice for the future, Teaming up with intel to work on the 22nm versions is only going provide some serious mega hasing, Maybe even giga hashing???
ANy info/advice on this company or what you know is good
Jump to: