RX 580 motherboard
30 Mar 2019 6:54:26 UTC
Topic 218532
(moderation:
Hi,
Planning to make a machine with 4 RX580s. Anyone can suggest some choices for motherboards? I have done some research but so far the results are not very clear, mostly due to the fact that many online shops lack details in the description or the sellers are not qualified enough to answer detailed questions. So, I hope someone experienced can share a few thoughts.
I am looking to have a comparison, but ALL must have AMD CrossFire/CrossFireX. I assume the RAC is (slightly?) higher if you use CrossFire, instead of just simply inserting them on the MB.
And what would you choose?
1. 8 or 16 gb RAM?
2. video cards with 4 or 8 gb Ram? Does it really make a difference? all are 256 bit. I found some big differences in price. For example, i can buy today (30.03) a "XFX Radeon RX 580 GTS XXX, 8GB, 256-bit" for 210eur/236usd. Or, I can choose "ASUS, ROG Strix Radeon RX 580 T8G, Gaming Top OC Edition GDDR5, 8GB" for 492eur/553usd. So, you can see the price varies big time. But not the performance, i assume.
Something in the middle would be "Asus Radeon RX 580 Dual, 8GB, 256-bit", going for 283eur/318usd. So it's a bit confusing, considering all are on 256 bit.
3.The PSU, what do you think about this: Corsair RM-X Series RM850x, 850W, 80 PLUS Gold?
Thank you, happy crunching!
John wrote:Hi, Planning to
)
Nope cross-fire only works in gaming not in crunching!!
8gb of ram is plenty but 16gb let's the machine breathe better.
The higher bit rate means a faster crunching rate meaning a 384bit gpu will crunch the same wu faster than a 256 bit rate gpu.
You need to know what the machine uses with one 580 in it, then you can figure out how much it will take with 4 580's in it. The side of the box should tell you how many watts a 580 takes, you will have to multiply the by 4 to figure out the minimum just for the cpu's then add in the machine and some extra so you aren't stressing this expensive machine. An Intel cpu can take more power than an AMD cpu depending on the model, then you've got quad core, 6 core, 8 core, 24 core cpu's etc etc to take into account for this new machine.
You too
mikey wrote:The higher bit
)
But that is purely a theoretical comparison. I don't know if there's ever been a GPU model with versions having a different bus width available. On the other hand, there are for example 384 bit GPUs that are much slower than some 256 bit GPUs. Crunching speed depends on so many other factors.
4GB or 8GB GPU... well, optimal amount of GPU tasks running concurrently these days has been 2-3 tasks... and that amount requires under 3GB of GPU memory total. So 4GB is easily enough for that. Who knows about the future, but I think the developers on this project aren't willing to raise the mem requirements aggressively. Big chunk of the voluntary computation power is probably based on such hardware that raising the requirements could drop plenty of computing power out of the grid.
Hi John, What are your
)
Hi John,
What are your goals, why are you wanting a quad gpu system and all the hassle that will bring?
As Mikey said, crossfire is not going to help unless your primary concern is gaming and even then I would ask the question, is it worth it? For crunching (here at Einstien at least) crossfire is not going to be your friend.
To answer your questions...
1). 8GB will be more than sufficient for a crunching machine and daily driver. I wouldn't go for more RAM unless you have a real need.
2). Go for a video card with 8GB of ram over one with 4GB, memory size does matter but the banwidth calculations can be very complicated! As a rule of thumb a 256 bit card with 8GB of memory should be quicker than the 4GB variant but the difference may be marginal depending on card.
3). The Corsair RM series psu's are a great choice but I personally wouldn't use the 850W to drive 4x RX580's mainly because I don't know the details for the rest of your system and at full crunch I suspect you will be cutting it fine...
Also consider the spending for 4 RX 580's and their long term power consumption, 1 or 2 Vega64 or Radeon VII crads may initially cost more but will pay dividends in saved electric costs and RAC over your proposed 4 card setup.
Gav.
Richie wrote:mikey wrote:The
)
I can speak to this. I've compared crunch times between the 2 types. The higher bit does mean that it will crunch the same work unit faster. Mikey was correct in his statement.
I also have several Quad Machines with 1080Ti. All my machines have 1600W PSU in them. EVGA Platinum series. And yes, the electrical cost of running them is expensive. But there is also the consideration of the power outlet and the circuit breaker. I found out the hard way about pulling too much power from them. Burned out 2 outlets and 1 Surge protector (lucky I didn't burn down the house) and required a higher capacity circuit breaker to be put in. I'd recommend a large UPS to protect both your machine and house.
Zalster wrote:I can speak to
)
OK. What GPUs / cards / models are you talking about exactly?
EVGA 1080Ti, 1070TI ,
)
EVGA 1080Ti, 1070TI , 1070FTW, 980TI, 980, 970s, 780, 770, 760, 750, 740, 730. Only reason I don't do any Turing is due to the issue different projects have with the new architecture.
Zalster wrote:EVGA 1080Ti,
)
None of those cards have identical architecture with each other in such a way that the bus width would had been the only variable.
GTX 780 has 384 bit bus width and 1070Ti has 256 bit bus width. Still 1070Ti would crunch a task much faster than a 780.
https://en.wikipedia.org/wiki
)
https://en.wikipedia.org/wiki/GeForce_10_series#GeForce_10_(10xx)_series
https://en.wikipedia.org/wiki/GeForce_700_series#GeForce_700_(7xx)_series
Usually you'll find different bus width & GDDR generation cards in the lower region of each Geforce generation.
Eg. GTX 1050 (2 GB) 128 Bit vs GTX 1050 (3 GB) 96 Bit
or
GT 730 DDR3 64bit
GT 730 DDR3 128bit
GT 730 GDDR5 64bit
I had such a GT730 DDR3 64bit once upon a time and it benefited quite a bit from overclocking the memory doing E@H.
This might not be the case for RX580 to the same extend, still I lost performance going down from 2075MHz vRAM to 2000Mhz. In my opinion bandwidth does matter.
John wrote:2. video cards
)
I you're planning to run only E@H, then you will definitely want to go with either that XFX or a Sapphire NITRO+ RX 580. Both have a dual BIOS switch that provide a mining BIOS, which will increase speeds and decrease power consumption. See the discussions over at https://einsteinathome.org/content/walk-amd-side. These appear to be the only two brands that provide a pre-loaded mining BIOS.
On newegg.com right now, a XFX Radeon RX 580 XXX Edition is going for USD$190 and it will perform just as well as the GTS model you mentioned (the main difference being the hot-swappable LED fans with GTS). On that same site, a Sapphire NITRO+ RX 580 is going for USD $200.
As far as I know, you will have to use a cryptomining type of motherboard to get more than three GPUs in one rig. Do a search on "cryptomining best motherboards" to see your options. Mining motherboards have chipsets that can support four or more GPUs.
I'm about to build an E@H-dedicated PC that will run a couple of RX570s, and maybe also a RX560. For that motherboard, I bought an ASUS Prime Z370-A II, which has three PCIe expansion slots. Most Z-series motherboards seem to have three PCIe x16 slots - I haven't seen any with four.
For a PSU, if you're going to be crunching all the time, I would go with a 80+ Platinum, which is what I bought for my new build. There are a series of recent posts about PSU considerations, preceding and including this comment https://einsteinathome.org/goto/comment/169263.
Ideas are not fixed, nor should they be; we live in model-dependent reality.
koschi wrote:Eg. GTX 1050
)
Yeah, I agree bandwidth would matter, but we wouldn't be able to compare those cards here in terms of what specification would cause what. Those 1050's have different "core configs" and the 128 bit version has actually slower single precision speed (1733 ) than 96 bit version (2138).
Two 730's with DDR3 had waaaay different cores with each other and because of that the 64 bit version looks to be about 5 times faster than the 128 bit version. I don't know how them would compare in real speed though. Then DDR5 vs. DDR3 will naturally change the bandwidth even with same bus width, but that's another thing.