Not sure of your point. Here, we are talking about a "high-performance application". The claim is that they can offer access to a cluster running Rybka, and that this hardware will be far more powerful than a single user can afford. Remember that part? A large cluster is lousy with any sort of ethernet connection. Latency is through the roof. But even if they choose to do that, this cluster would be far more powerful than what a single user could realistically afford, until you reach N users, where N is the number of nodes on the cluster. Now it no longer offers such high performance, because it is a single resource shared among N users. Only way to use it is in a sort of "simul mode" where it searches a second or so on one position, and then moves on to the next user's position, until it reaches the end and wraps around.Gian-Carlo Pascutto wrote:Already factored in, as pointed out in the original post. No need for silly expensive things like infiniband.bob wrote: What about connections? We use infiniband. 128 nodes requires a bunch of switches. Plus the bandwidth coming in to make this usable. More expense.
Nobody says you have to start with 128 quad-cores. That's just a number you came up with. But anyway, let me take you up on it:If you believe that usage will be very light, it might work, although you are still stuck with that initial half-million dollar investment to buy the hardware and environmental controls.
There are hosters offering Core i7 920 hosting in Germany for 50 Euro per month, or 600 Euro per year.
Running cost of 128 of those boxes: 77k Euro per year.
Investment cost: 20k Euro setup fees.
That doesn't exactly sound like earth-shattering performance, it sounds like it will end up being _weaker_ than what a user can buy off the shelf, because it won't be dedicated to that user.
As far as operating 128 nodes for 77K Euro a year, that sounds like someone is ignoring A/C costs. But in any case, I can only quote the numbers for the clusters we are running here. And cooling is a _huge_ problem, we've had to upgrade twice over the last 4 years as we added more hardware.
I'm not convinced this model for computer chess will work. It might solve the piracy issue, but it will almost certainly fail due to user dissatisfaction. Doesn't concern me either way however.