Rybka Rental program is a nonsense.

[Don]

I understand, and I would prefer e.g. a "time to depth x"(*) comparison. But I didn't find any other performance figures from the cluster yet, than these nps which are from a Rybka 4 benchmark page. I think that page was initially intended to compare "normal" comps running retail Rybka 4.

*) I guess time to depth x is not an optimal comparison either, because if I get things right the search tree on a bigger number of cpus will be bigger, for the same depths? - Maybe solving times for some good test positions are more relevant, in terms of "effective performance" so to speak.

Anyway we don't have such figures and I doubt that we will get them. Although, anyone willing to pay for it could rent Rybe cluster time just for that, too.

If many people are curious, you can set up a collection to rent some Rybka-cluster time for those tests. Then make a page with the results (which is probably what Vasik should do as well, if the cluster actually performs well).

It would be useful to play a thousand game match to measure it's actual performance. It would be best to be done blind, without telling him about it so that he doesn't do anything to make it play optimally for the match. The question to answer is, "what would I be getting?"

Don

[Suj]

When rybka cluster ran with 248 cores and then 260 cores the nps count was still around the 18000 kns mark and I think one shouldn't be taking this for doing any speed up calculations.

Does this actually mean running 60 additional cores yielded no extra nps?

[Werewolf]

I understand, and I would prefer e.g. a "time to depth x"(*) comparison. But I didn't find any other performance figures from the cluster yet, than these nps which are from a Rybka 4 benchmark page. I think that page was initially intended to compare "normal" comps running retail Rybka 4.

*) I guess time to depth x is not an optimal comparison either, because if I get things right the search tree on a bigger number of cpus will be bigger, for the same depths? - Maybe solving times for some good test positions are more relevant, in terms of "effective performance" so to speak.

Anyway we don't have such figures and I doubt that we will get them. Although, anyone willing to pay for it could rent Rybe cluster time just for that, too.

If many people are curious, you can set up a collection to rent some Rybka-cluster time for those tests. Then make a page with the results (which is probably what Vasik should do as well, if the cluster actually performs well).

It would be useful to play a thousand game match to measure it's actual performance. It would be best to be done blind, without telling him about it so that he doesn't do anything to make it play optimally for the match. The question to answer is, "what would I be getting?"

Don

I did try to do this. But unfortunately when you pay for a machine you want it to do something other than satisfy curiosity. So most of the time I used the Cluster for was opening work.

But my impression was, in March, from limited testing, Cluster Rybka 40 Cores is 160 elo > Houdini 1.5a on a fast Quad.

Now it is April & the Cluster has made progress (strongly hinted at on the Rybka forum) and no doubt Robert has made progress too....

[tiger]

Any wealthy chess enthusiast would rather invest $30000 in a Cray Cx1
(Windows HPC Server 2008) with Houdini, rather than renting R4 cluster
for an equivalent amount. By the way, CX1 is only the first step in Cray' s
plan to take advantage from Windows hegemony.

What will the R4 cluster be worth when Cray 256-cores Windows power computers will be available ? Can amateur clustering compare with supercomputing experience ? Harsh times ahead for Rybka connection.

Qapla, Robert !

If you want to know if the business model is sound, just tell them you want to make a reservation for the full monthy cluster (100 cores) for 2 days (it's the minimum you must reserve, that makes around 1200€ or $1700 for these 2 days).

The information you want to get is WHEN the cluster is available.

They can't ask you to pay BEFORE telling you when you can get cluster time. So it won't cost you anything to get this information. They can't ask you who you are either. If you are a top chess player you want to be able to access this thing anonymously. It's a condition they will have to accept, or they will put down potential customers.

If you CAN get a reservation for today or tomorrow then you know they can't get 100% usage time.

Just repeat the experiment several times and deduce how much money they get from their customers.

Then we could, just for fun, evaluate the cost of running this thing. I'm pretty sure there are IT professionals reading this forum who could do it with a good accuracy.

And finally we could deduce how much money they are making.


But one thing I notice is that their business model is based on TRUST:

- You have to TRUST that you have access to the hardware you have reserved: in order to fully book the cluster they will be tempted to allow several users at the same time without telling them. It's just a matter of faking the NPS (if that sounds familiar).

- You have to TRUST that the results of your analysis are not going to be transmitted to someone else and that nobody else will be able to snoop on your connexion. What if I offer to pay them $3000 per day to snoop? No? Then I offer $5000. It's much more valuable to know your opponent's preparation than doing yours.

- You have to TRUST that if you discover an opening novelty the team is not going to use your discovery. What if I pay them $10000 in advance to get 3 opening novelties? I just have to sit and wait until 3 novelties are discovered: it could be monitored continuously by a separate piece of software, they don't even need to do it manually.


Chess players tend to be secretive and even slightly paranoid, and this Rybka cluster project is based on trust. Smart move guys.


// Christophe

[wolfv]

Any wealthy chess enthusiast would

If you want to know if the business model is sound, just tell them you want to make a reservation for the full monthy cluster (100 cores) for 2 days (it's the minimum you must reserve, that makes around 1200€ or $1700 for these 2 days).

The information you want to get is WHEN the cluster is available.

They can't ask you to pay BEFORE telling you when you can get cluster time. So it won't cost you anything to get this information. They can't ask you who you are either. If you are a top chess player you want to be able to access this thing anonymously. It's a condition they will have to accept, or they will put down potential customers.

If you CAN get a reservation for today or tomorrow then you know they can't get 100% usage time.

Just repeat the experiment several times and deduce how much money they get from their customers.

Then we could, just for fun, evaluate the cost of running this thing. I'm pretty sure there are IT professionals reading this forum who could do it with a good accuracy.

And finally we could deduce how much money they are making.


But one thing I notice is that their business model is based on TRUST:

- You have to TRUST that you have access to the hardware you have reserved: in order to fully book the cluster they will be tempted to allow several users at the same time without telling them. It's just a matter of faking the NPS (if that sounds familiar).

- You have to TRUST that the results of your analysis are not going to be transmitted to someone else and that nobody else will be able to snoop on your connexion. What if I offer to pay them $3000 per day to snoop? No? Then I offer $5000. It's much more valuable to know your opponent's preparation than doing yours.

- You have to TRUST that if you discover an opening novelty the team is not going to use your discovery. What if I pay them $10000 in advance to get 3 opening novelties? I just have to sit and wait until 3 novelties are discovered: it could be monitored continuously by a separate piece of software, they don't even need to do it manually.


Chess players tend to be secretive and even slightly paranoid, and this Rybka cluster project is based on trust. Smart move guys.


// Christophe

Hi Christophe,

yes, the key word may well be 'trust'. However, the Rybka cluster team have explicitly stated the limits of their legal (and probably moral) liability, responsibility, commitment, etc. If you care, you can infer and 'visualize' the amount of trust you mention reading the Disclaimer closely. Actually, being ignorant about the possible legal issues, I wonder what is the extent of the guarantees the Rybka cluster team offer to a customer. Here goes:

"Disclaimer

Time on the Rybka Cluster is offered as is. The Rybka Cluster has been extensively tested in a wide range of scenarios, but we make no warranties of any kind, either express or implied, including but not limited to the implied warranties of performance, merchantability, satisfaction, or fitness for a particular purpose."

[ozziejoe]

That is an intersting idea for the cluster use, but it would seem to apply to only a small consumer base. As an 1900 player, I won't benefit from having the extra power, even if my opponent has just won in my pet line. I suspect nobody under 2400 would benefit with the cluster...They just need to go out a buy a good quad core and they will have an advisor that is over 3200 playing strength.

[Don]

That is an intersting idea for the cluster use, but it would seem to apply to only a small consumer base. As an 1900 player, I won't benefit from having the extra power, even if my opponent has just won in my pet line. I suspect nobody under 2400 would benefit with the cluster...They just need to go out a buy a good quad core and they will have an advisor that is over 3200 playing strength.

I'm curious to see how well this type of business model works. I think the killer clones and Houdini must have surely taken some business away from them, but there was always a risk of that, clones or no clones.

If they can sell it, it works for them. It's hard for me to believe that anyone would pay for time on this machine as you can run a strong program on your own quad and get the same strength by just letting it think longer. If I were a strong grandmaster I would personally only consider this if it was so strong that I needed a factor of 10 to beat it. I would rather let a program like Houdini think for an hour than to pay for 15 minutes to get the equivalent strength (and I don't know what is equivalent, this is just an example scenario.) I seriously doubt this setup is so far ahead of Houdini to make it worth it to a strong grandmaster. Unless it's so cheap or they are so rich ....

I also would feel a lot of anxiety over the fact that I would not really know how much effective CPU time I'm really getting. They don't give any guarantees do they?

Don

[Gian-Carlo Pascutto]

It seems they have not even done some basic checking of their pricing level either. They are claiming their cluster is much cheaper than Amazon EC2, but some simple calculation shows that they are 30% more expensive.

And this is in the best case calculation for them, i.e. assuming that you stay up the full 5 days in sequence to fully utilize your time. The Amazon EC2 instances can be rented per hour, whenever you want, and you can put stronger programs than Rybka on them.

It does not make sense to rent the Rybka cluster.

[Dann Corbit]

It seems they have not even done some basic checking of their pricing level either. They are claiming their cluster is much cheaper than Amazon EC2, but some simple calculation shows that they are 30% more expensive.

And this is in the best case calculation for them, i.e. assuming that you stay up the full 5 days in sequence to fully utilize your time. The Amazon EC2 instances can be rented per hour, whenever you want, and you can put stronger programs than Rybka on them.

It does not make sense to rent the Rybka cluster.

There could be some esoteric move that is pivotal for a world championship correspondece event or some such thing.

If the compute power of the cluster is the only way to get the right answer, then it could make sense to rent the cluster, since no other programs that I am aware of which are of similar strength are cluster capable (though the skeleton of Daniel Shawul's program shows that it is not incredibly difficult to achieve this).

On the other hand, I suspect that the scenario I have described affects 3-4 players in the world and then only once or twice in a lifetime.

It could also make sense for a book author who is verifying assertions made in a chess book and who has a limited time schedule to complete the verifications and wants to be very sure about them and still meet the deadline.

On the other hand, I don't see how you can make money with that model.
On the other, other hand -- often marketing hype is more important than factual analysis of cost/benefit so it might make money anyway.

[Don]

It seems they have not even done some basic checking of their pricing level either. They are claiming their cluster is much cheaper than Amazon EC2, but some simple calculation shows that they are 30% more expensive.

And this is in the best case calculation for them, i.e. assuming that you stay up the full 5 days in sequence to fully utilize your time. The Amazon EC2 instances can be rented per hour, whenever you want, and you can put stronger programs than Rybka on them.

It does not make sense to rent the Rybka cluster.

On the other hand, it does not always have to make sense to sell it. There are many things that are sold today based on ignorance, much better values available for much lower prices or even free. For example how many people buy a chess program at Wallmart or somewhere off the shelf that is much weaker than Stockfish 2.0 which can be obtained for free?

If everyone made really intelligent buying decisions, a LOT of companies would immediately have to go out of business.