Was reading about scalability on wiki lately.

[fghj]

They estimate 12msec in total per transaction for ECDSA verification, at 80tps it leaves 0,5ms for rest.
Could ECDSA verification be outsourced to FPGA or GPU?

[DeathAndTaxes]

Sure an FPGA could be used and so could an OpenCL/CUDA capable GPU (either discrete or integrated into a modern CPU).  However it is very likely that neither will be needed.  The 80tps assumes existing codebase with costly double verification on a relatively old CPU and it is PER CORE.  Newer CPU are integrating GPU cores on the die. While they can't match their discrete brothers they have 500% to 1000% of the IntOp performance of even the fastest CPU.  This trend will only continue.  Just like today finding a non-64bit or single core CPU is almost impossible in 10 years it is unlikely you will be able to find a CPU which doesn't have some kind of GPU "like" cores on die.

CPU isn't a bottleneck.  

We are at ~0.35 tps today.  Say we double transaction volume every year for next decade (pretty impressive).  That would be ~350 tps or 11 billion transactions annually in 2012.  At 0.5 USD cent per transaction the network would be protected by 50 million in miner revenue.  If miners were willing to accept a 10% ROI that would mean nearly 500 million in hardware.    To put it into perspective PayPal is on the order of 50 tps.  

Oh noes a CPU maxes out at 80 tps.  Well once codebase is improved to remove double verification it is more like 160tps PER CORE.  Today a quad core would be capable of ~640 tps.  A decade of Moore's law says average CPU will be ~32x as powerful in 2022.  Lets be conservative and assume 1/3rd of that.  That puts a CPU at around 5000 tps in 2022 (more than VISA).

If that isn't enough GPU, FPGA, or ASIC could be used.   We can use SHA-256 hashing performance as a rule of thumb.  A total guestimate of ECDSA throughput of a GPU would be maybe 5x to 10x that of a TOP of the line CPU lets me conservative again and say only 4x the performance.   So we are looking at maybe 20K tps or more.  FPGA likely would be even more (and at lower power costs).  While a dedicated ECDSA ASIC might not makes sense, maybe the BFL devotes 2% of the die space on the SuperComputer 3.0 released in 2022 to enable hardware accelerated transaction verification while mining.

TL/DR:
The 80 tps shouldn't be seen as a hard limit.   Performance can be increased by
* removing double verification
* other code enhancements/optimizations
* the continual increase in computing power
* OpenCL acceleration
* dedicated processors (FPGA or even ASIC).

[fghj]

Great! Running full node will not be that expensive.
I could imagine that most nodes will have pruned blockchain on fast drives and maybe full version on tape drives for analysis and/or seeding, so network is last one left.
It will probably be no problem in Japan, South Korea and some other places.
In some more remote places I could imagine that there could be "half-full" nodes that do SPV and get rest via "IP over Avian Carriers" or "IP over Dog Sled" once a week/month.
Now it looks like even with current tech we could do full node doing 6000tps for less than 3000USD.

[Mike Hearn]

Double verification has already been fixed.

I don't think we fully understand the current bottlenecks. Technically we're not in the dangerzone yet, but in practice miners have started load shedding due to IO shortages. LevelDB is the next big win.