older Opteron boards worth it or no? (vs newer), also RAMdisk SAN from server mem

[Twice_Shy]

So I have in my head this idea after perusing things like ebay... having seen older Opteron hardware with up to 16 cores, and up to four sockets on a motherboard often available surprisingly cheaply... yet i'm wondering how those old cores stack up against modern ones. I'm sure power efficiency wont be anything near, but i'm curious about raw performance - if I had plenty of power and needed either real time performance (or some other time-critical workload) in what situations will throwing 64 older cores at something beat the 8-16 newer Xeon cores I might put together for the same price? Any ideas or insights?

I'll admit this is more a "solution looking for a problem" right now, I have ideas in my head of what i'd like to do I just don't know how much if any of it is feasible.

Another idea that is in my head is seeing very cheap and inexpensive server RAM selling for a fraction of what Desktop type ram does - along with server boards that can take much more than the 32-64gigs more common on newer stuff. This has me wondering whether there is an opportunity to create a super high performance SAN using RAMdisk instead of SSD's which might saturate even 40gig Infiniband or 32gig FibreChannel at a price-per-gig of RAM that I can't match trying to put on the desktop side.

This is separate from doing the numbers of what a given size job would cost doing it via SSD instead of RAMdisk and power usage and all that - it may well be fully feasible just not worth even the reduced ebay prices, or even free prices, due to ongoing power bills.

 

[Rand__]

CPUs:
-Per Core (peak) performance is significantly lower on older CPUs
-Most applications don't scale up in a linear way with more cores
->If power costs are negligible and your application(s) scale(s) up well enough and your hardware costs are low then yes you can do that-
=> It depends - foremost on your workload, i.e. application

Ram(disk):
Again, depends - if your working set is small enough (i.e. can fit in a single Ramdisk or you can build a network based Ramdisk via IB) then o/c you can do that:
Opteron 6300:
Four DDR3 memory channels with max memory bandwidth of up to 51.2 GB/s @ DDR3-1600
368 GB per Socket

 

[Patrick]

I think @Patriot had more quad socket Opteron 6300 systems than I did. I would skip the Opteron 6300 these days unless there was a very good reason to pick one up.

There was a deal that expired on the 31st for a new 4P E5 V1/ V2 system at $500 Newegg asus barebones server sales so there are opportunities for faster E5 systems.

With DDR4, 32GB modules are very common. While they cost more than inexpensive 4GB DDR3 modules, they also mean that in a standard 8 DIMM E5 V3/ V4 you can get 256GB with higher bandwidth.

Finally, on why I am not recommending Opteron 6300 systems, they lack features like SATA III and PCIe 3.0.

On the flip side, the Opteron 6300 series is very inexpensive and if you have free power it may be fun.

 

[Twice_Shy]

Well this would be more using school power while i'm at film school. Probably wont make some people happy but it's sort of dont ask dont tell for some things. We can plug in overnight rendering rigs and let them slave away if needed, even several of them, so it got me thinkin...

Workload would be either certain computer rendering or video footage processing stuff - that parallelizes really well - i'm just not sure what the equivalent multiplier is for earlier generation of cpu cores. Like if a single 3ghz modern Xeon core is equal to four Opteron cores at 2ghz then yeah may not work that well. Are there any rules of thumb for "performance per clock" by generation to compare older stuff to the latest and greatest?

For the ramdisk operation I guess I was just wondering if anyone else had done that. My main concern was cost per gig more than gigs per board, and most of the rest of the system almost an afterthought - mostly interested in serving up a super fast RAMdisk over IB or FC as a form of SAN node if the dataset one is working on exceeds the commonly haveable desktop RAM limits for consumer hardware. If good deals are found throwing a second board at it as another SAN node to increase the size or/and performance even, direct patched to multiport host bus adapters on the workstation. Though perhaps the latency however small still might mean it's not much better than a fast SSD and i'm curious what math i'd have to calculate to figure that out. I could be pulling a Rube Goldberg workaround here.

 

[gbeirn]

My opinion is you would need something that is highly threaded to make them worthwhile even with free power. Opteron 6300 are the same (essentially) cores as the FX desktop line. Difference being these have significantly reduced clock speed for increased core counts. Performance wise compared to an Intel system you are looking at maybe somewhere in 5600 series. So at that time competitive especially since Intel only had up to 6/8 cores.

 

[Rand__]

Passmark:
[Dual CPU] Intel Xeon E5-2670 @ 2.60GHz =>18374
[Dual CPU] AMD Opteron 6386 SE =>10764

So if you get a quad Opteron system cheaper than a Dual E5 system you are good to go

 

[linuxmanbg]

"Finally, on why I am not recommending Opteron 6300 systems, they lack features like SATA III and PCIe 3.0."
For SATA III workaround is using PCI-e card or RAID with SATA 6gb, well for PCIe 3 very important parameter for PCI-e SSD no workaround as far as I know.
"On the flip side, the Opteron 6300 series is very inexpensive and if you have free power it may be fun."
I bought Supermicro micro ATX mobo H8SCM-F accept DDR3 REG ECC i.e. server memory and Opteron 4332 6 cores (new feature CPU's power meter integrated ) to replace my home server's motherboard. So agreed cheaper than Intel solution.