Horsepower for ASIC

[AmDD]

I read a comment https://bitcointalk.org/index.php?topic=109297.msg1188865#msg1188865 that made me think...

How much computer power is an ASIC going to require? My thought is your internet will be the bottleneck, as fast as it can download your computer should be able to feed to to the miner... I would think. Is the data from pools compressed any and therefor require the cpu to uncompress before sending to a miner (GPU, FPGA, ASIC)

Thoughts?

[live627]

The ASIC does all the work. The computer is just a relay.

[crazyates]

As far as CPU usage is concerned, CGMiner is very lean and fast, and hardly uses any CPU time, even with multiple BFL Mini Rigs. It also preps new work for your miners while they're hashing so there's very little down time. As far as network usage is concerned, rollNtime already does a lot to reduce traffic, but it's not enough as we move into ASIC territory. There are two new proposals for condensing network usage, drastically reducing both network usage, and enabling ASICs to reach 100s of GH/s with zero downtime.

I made that comment about the C2D laptop because when you're mining on thousands of $ worth of ASICs, why introduce a weak point in the form of a 10 year old desktop CPU? I'm confident that it can run the mining software, but why take the risk?

[AmDD]

I understand that, what I mean is since it is so much faster than anything else we have used, how much computer power is it going to take to feed the ASIC work fast enough? Or will it not be any kind of issue? Right now we can run a FPGA off a router... will that router be able to feed an ASIC? Will a raspbery Pi? an Atom? etc...

[AmDD]

I made that comment about the C2D laptop because when you're mining on thousands of $ worth of ASICs, why introduce a weak point in the form of a 10 year old desktop CPU? I'm confident that it can run the mining software, but why take the risk?

Very true, however a rackmount server would be a bit overkill I would think. It would be nice and neat but for the pricetag Im not so sure. Maybe one second hand thats in good shape...

But either way, just thinking out loud...

[crazyates]

I understand that, what I mean is since it is so much faster than anything else we have used, how much computer power is it going to take to feed the ASIC work fast enough? Or will it not be any kind of issue? Right now we can run a FPGA off a router... will that router be able to feed an ASIC? Will a raspbery Pi? an Atom? etc...

IIRC, almost every bottleneck has been addressed. Mining programs use very little CPU/RAM and zero HD activity. The data sent to/from the miner through USB is nowhere near the max throughput of the USB spec. Network activity can be iffy, but new protocols and higher-diff pools are being built to address that.

Now noone's every actually mined on a 1TH rig before, or at least to my knowledge.  So there may be barriers once we do get there, but who knows? /shrug

I made that comment about the C2D laptop because when you're mining on thousands of $ worth of ASICs, why introduce a weak point in the form of a 10 year old desktop CPU? I'm confident that it can run the mining software, but why take the risk?

Very true, however a rackmount server would be a bit overkill I would think. It would be nice and neat but for the pricetag Im not so sure. Maybe one second hand thats in good shape...

But either way, just thinking out loud...

You can pick up used servers on ebay for under $100. Almost as cheap as a laptop, and fits right it in in a datacenter.

I'm totally with you there, tho. I'm always thinking outloud. I process best by explaining to others. 

[vitruvio]

I think all this is solved with higher dificulty shares.

There is already pools that serve work with a higher difficulty than 1, so indeed if you have a gpu mining at 350 MH/s in cgminer you will see 7-8 shares every 5 seconds ( so it supose 7-8 communications with server every 5s) you are right thinking that a jalapeno that would do 3.50 GH/s will need to communicate with pool server 70-80 times every 5 s, this is not a problem for a normal pc and actual ISP speeds, but if you think it may be for your CPU / network, use a pool with a variable difficulty as eclipsemc or a pool with a fix difficulty i.e. 10x, so the network transfer will keep the same, you ask for a work with 10x diff to pool and will get shares 1/10 times but will count as 10x for rewarding. Less net traffic, less work for miner soft, same result.


Regards

[AmDD]

Thats something else I dont understand, pool difficulty... 1) why do they use a difficulty of 1 when the actual difficulty is over 2 million? 2)how can they change this difficulty? If its easily changed doesnt that defeat its purpose?

I can see how adjusting it will reduce network traffic and thus less work to feed the ASIC but resulting in the same payouts in the end. Makes sense.

[vitruvio]

I don't know technically how is done, but to my knowledge, that is what a pool does, divide the total work necessary to find a block in portions, ("share" the work between miners). With high difficulty in solo mining you probably won't find a block at low hash rates, maybe before you finish to make a calculation with difficulty 2 millions the block will be found by another and you must restart the work with other data and so on... Somebody correct me o explain it better "in a human way"  please.

About how to change this difficulty, the same way as the pool can divide the total work in 2 millions of parts the can do it ins 2 millions/10, giving you a work to calc more complicated that by average you will take 10 times more.  This is a global view of thing, technically there a parameter called target, but I can´t explain you how it wokrs, there is some technical post that I don't even readed it

Regards

[eleuthria]

This problem is being solved by the Stratum protocol (and possibly by getblocktemplate).

Standard getwork protocol will not scale with ASIC chips.  Higher difficulty will only reduce the upstream traffic, not the downstream.

With Stratum, pools provide miner with a packet of work that is equivalent to 4.2 billion getworks (or more), while taking less bandwidth to transfer than a single getwork.