If I had the money I'd get a full threadripper workstation, or at least a ryzen 5950x, but for now if I am to actually consider building a PC, it can't be that expensive and thus I wanted to ask what generally appears to be the most cost-effective solution I can look to save up for.

Ideally I would have a computer that can not only solve quite complex game trees, but also do so in a short period of time. As established, this is a luxury, and thus I would probably just be hoping for a computer that can solve relatively deep-stacked SRPs in a reasonable timeframe, and then be able to knock out some very complex trees in some not completely unreasonable amount of time.

Does anyone have any general advice for what I should expect from such a rig, what components to look for/ threads & clock speed etc., and a reasonable price range I should be looking to meet? Much appreciated.

Sent from my SM-G973F using Tapatalk

can't afford a R9 3900x? get a R7, dip into the sh market and try to find an earlier gen R7, 2700x or even a 1700, that's even better value if you can find one for ~100$. short of that, dual xeon cpu workstations if you can find one for a good price, nothing else offers similar price to performance like R7/R9 or 2x 12c/24t xeon refurbs

Go down this list until you find a CPU that you can afford. The price and multi-core score should be the deciding factor for your solver build

Cheers, I'll look at that. The thing I'm having difficulty with though is that I don't have a fixed budget, and thus I'm struggling mainly to decide between getting something very cheap, or looking to shell out a little bit on. Like, say I drop $1500 on a PC, if spending another $500-$1000 would significantly unlock solving abilities (such as being able to solve SRPs within 30 minutes, being able to solve 3-bet pots really deep with upwards of 3 bet sizings etc.), then I'd probably look to do that...

Sent from my SM-G973F using Tapatalk

Can you really get a xeon setup cheap enough to be worth it?

Sent from my SM-G973F using Tapatalk

https://www.amazon.com/Workstation-P...3462861&sr=8-6

something like this would be similar to a R7 2700x on paper in gto+

If I may hijack the thread. Has anyone tested how much additional cores help with performance?

Right now I have a 9900k with 64GB and I am pretty sure it will manage well with most solver work. I am a computer geek though and I can't help but wonder what level of improvement I would get if I had a 5950x with 128GB of ram. 10%? 20%? 50%?

PS. FWIW, if you disregard what pokertracker says and install one of the later version of postresql that supports multithreading, you do get improved performance which does come handy when you are trying to switch from different reports containing thousands of hands. I am using postresql 11.2 and other than a few annoying messages, I haven't encountered any discernible problem yet (touch on wood). I also wonder how much improvement I will get in those use scenarios with a processor like the 5950x.

PS2. I am definitely not getting a 5950x. At a minimum, I am waiting until the fall to see what Intel's 10nm process does and maybe even further until DDR5 becomes a thing.

PS3. I am shocked that it's hard to squeeze any performance from even better SSDs. Higher sequential speeds don't matter at all.

gto+ dev said there are diminishing returns once you get passed a certain amount of threads, I don't recall the number, either 24 or 32; you may ask him in the crEV support thread
ram is just used to store the trees, ram speed does not affect performance
storage speed will not affect solve times (unless you're working on a database with the autosave feature on - it auto saves after every flop it solves). faster ssds will also open large files quicker (databases)

now for the hud, faster storage should make a significant difference when loading reports and running filters, ram speed prob won't matter, cpu will but I think you're looking for higher clocks rather than higher thread count here as I doubt it can leverage all threads

pretty sure Intel is done for the foreseeable future. comfortably second. I think it won't be long til Nvidia won't be able to compete with AMD either. all hail our AMD overlords!

Intel 10nm process is a dog, that's why they're still shipping so many 14nm+++++ processors. It just doesn't clock well, they won't do a re-engineer on the process at this point, best case scenario is that they apply lessons learned to 7nm.

9900K is a snappy proc, but doubling the physical core count would move the needle. DDR5 at launch is likely to have a large price premium over DDR4.

Speaking from a SQL perspective, RAM RAM and more RAM. Outside any licensing/configuration limitations, SQL will use as much RAM as you can throw it's way, and RAM is a couple orders of magnitude faster than even SSDs. Storage speeds still matter for reading data into RAM and writes back to disk, but queries against tables already in memory will run way faster.

@OP:
If your solver software's licensing allows for it, consider running in a cloud VM on Azure or AWS. Renting compute may be a reasonable expense, for standard IAAS VMs, you only pay for compute while using it and any permanent storage you're using. Make sure you shut down the VM when done, you may also want to set billing alerts to keep an eye on usage.

Thank you for your replies, guys, much appreciated.

Some random thoughts.

1. What I knew about the 10nm process was that Intel couldn't produce cost efficient yields, not that it didn't clock well. Obviously, who knows, but hopefully, we will find soon enough. Either way, the 6950x is bound to be a decent upgrade for AMD, so....

2. How much time does a typical solution take? From what I read here and there, you can do a flop in 10 min or something, but if you solve for multiple hours, you leave it up all night?

3. SSD performance and postgresql. When I upgraded from my 4770k to my 9900k, I also got an NVME drive as an upgrade to my sata SSD. Both were high end models from Samsung. I tested the 9900k with both. It didn't make a lick of difference in loading reports.

On the surface, that's surprising, because you see that NVME drives are 5-6 times faster than SATA drives. But it doesn't matter, because those are sequential speeds and apparently database reading doesn't involve sequential reads. If I understand correctly, an SSD will make a difference from another SSD if it's faster on lower queue depths. In the regard, an NVME drive isn't faster than a SATA drive.

A few years ago, I went ahead and I bought an Intel Optane 900p which was at the time the fastest drive around. It shaved 15 seconds or so out of a report that took 105-120 seconds to load. But it was very expensive and not sure it was worth the cost.

4. Like I said before, upgrading to a later version of postgresql helped a lot with performance. The versions supported by Pokertracker do not utilize multiple cores. The latest versions do, so that's why it makes as difference.

5. RAM. I have noticed that when the size of my PT4 database exceeds the size of my RAM, there's a drop in performance. You really have to play with the settings however.

As far as RAM speed is concerned. 10 years ago, we weren't concerned just with loading reports faster. We were concerned with having the HUD updated in a timely manner when we were playing multiple tables. At some point, even a video card upgrade made a difference because a slow GPU acted like a bottleneck!

Anyway, back in the day, I was doing research on what to look for in an upgrade and I ended up in the postgresql forum, where the guy who wrote postgresql -who turns out used to play poker as well- told me to get as fast a memory as I could.

Nowadays, the HUD is updated pretty fast, so I don't know how much RAM speed affects database performance in other ways.

I do pledge that if I bing the SCOOP main event, I will upgrade my computer and tell you which components make the bigger difference. If I don't bing however, you are SOL.

2.
vastly depends on the complexity of the spot, i.e. how wide ranges are, spr, number of bet and raise size options in the tree
but smth simple like a 4b pot at 100x eff with a few bet sizes everywhere just takes 15-20 seconds
I have a dedicated R7 2700x pc for solves and solving an average srp spot (all 1755 unique flops and a practical tree) took ~6-7 days of running 24/7
ram wise you're looking at 32gb for all your practical needs and super complex trees get to 128+
for solving preflop you kinda need 64+ to get accurate results for 6m and a tree takes a day or 2. not the case if you're solving for fringe tourney spots, fewer players, not 100 deep, those are going to be quick and light on resources.
3 way solves take quite a long time too and need more ram than hu, as you'd expect

I've learned a bit about sql, 10x guys