Poker-AI.org

I've done a bit of Googling, and there's a fair amount of contradictory information, so I was hoping someone with knowledge could help me out.

What CPU should I get for decent number crunching performance? (At a reasonable price)

Wouldn't take too long before I may as well have bought a high end machine myself.

What's "a reasonable price"?

I'm looking at a complete build for under £1000 - the cheaper the better.

I was looking at this lot:
Gigabyte GA-990FXA-UD3
AMD FX-8350 4GHz
G-Skill 32GB (8x4GB) DDR3 1866MHz
WD 1TB 3.5" SATA-III Caviar Green
Sapphire HD 6450 512MB
CIT 650w Power supply
AvP Triton Mid Tower Black Blue

Total of £611.82 on Ebuyer

Obviously for a number cruncher, graphics and drive speed are not a concern.

The FX-8350 with 8 cores for £156.64 is very tempting compared to the Intel Core i7 2700K for £274.54 with 8 threads. But I was wondering if it'd be worth the extra money, or even the Intel Core i7 3930K with 12 threads for £466.12. I don't really know what the performance comparisons would be like.

I noticed that my current fictitious play implementation runs more than twice as fast on my i7 laptop, as it does on my AMD 1090T desktop, Both running 4 threads. I don't know if that's likely to be purely down the to the CPU - or issues with the desktop setup.

Edit: Or the Intel Xeon E3-1275v2 for £273.03

I've always done AMD builds in the past - so I'm fairly unfamiliar with the Intel world.

Nominal EC2 prices are high, but you can rent really cheap machines on the spot market - I'm currently renting 8 core 68 GB RAM machines for $0.15/h.

That must be linux.

Intel computers seem to perform better for number crunching. You might look up performance comparisons for the different processors you had in mind.

Yeah that's for Linux machines.
The CPUs are somewhat dated Intel Xeons (definitely slower per-core than my desktop i7-3770K), but still, for infrequent large jobs this offer is hard to beat - renting 400 cores for a couple days (that's my use case lately) costs pennies compared to their price. If the algo that OneDay wants to use is parallelizable across multiple machines, going massive with a farm of rented spot machines is definitely something I'd consider.

It's not currently parallelizable across multiple machines, and I foresee this doing a lot of number crunching for a long time. So I don't think rented power is the best way to go in the long run.

+ Once I'm done number crunching, I get my next Dev machine for free!

Speaking of getting your own machine, it's possible now to get a strong machine (CPU-wise) which is absolutely noiseless (no moving parts except the CD drive) for a reasonable price (I got mine for about $2k from quietpc.com). It's great for people who are annoyed by their botting machines making constant noises in their bedroom/living room.

Why not use a good gpu instead ? you'll get about 15x performance than cpu for number crushing.

Regards

GPUs aren't performing well if the workload doesn't fit into the memory afaik.

Regarding EC2: How does it work when using the spot instances: do you need to setup your number cruncher in the auto start of the OS?

AFAIK GPU are rather primitive creatures, for example don't have the ability for recurrence (which is essential for typical tree-traversing algorithms).

Spot machines are just like regular machines - you create them from from an image, which can be either a vanilla Linux image provided by amazon or an image you've prepared yourself (by making a snapshot of another instance earlier).

yes, but afaik you don't have direct control over the fact when your VM is running or stopped. So for instance you make a bid and the price is currently too high. Hours later it reaches the price and you automatically get the instance booked, but how to start the cruncher automatically? Do you have to take care of it via the OS mechanisms for startup? Or do you create an instance, start the process, hibernate it and pass it the spot market so its executed as soon as there's capacity available for your price?

You set a max price you're willing to pay, then just accept the current spot price. The trouble is, if somebody later buys out all the server capacity, essentially outbidding you, the instance gets shut down and no new data is saved. <-- This happened to me last week. I think you can preserve your disk instance somehow in case that happens. I'm not entirely sure how to do that, though. You can create an instance and save a disk image (AMI) with your app on it, but that's done manually.

Only if the problem is parallelizable - which his isn't

It's possible to make the disk persistent, so work isn't lost - I think it set when creating the image.

Sounds, like this would be great for computing NEQs with a high number of buckets, definitelly sounds easier than connecting 4 computers in a LAN, which was my original idea. Are there any drawbacks, that makes renting a machine with lots of RAM impractical?

Transfering a lot of data can be cumbersome (I'm seeing transfers around 700 KB/sec) , although there is a service for shipping data on physical hard drives to/from AWS S3.