You'll notice by the end of this post we are still no where near final card selection, and we will not be for another couple of posts, until all the factors are laid out for you to mull over.
Project List
Your preferred list (in brackets is my addition of primary GPU application[s] )
Milky-Way (AMD optimum, NVidia acceptable alternative but sloooow, needs double precision hardware)
Einstein (NVidia project, no AMD application, I think its single precision hardware but that needs checking)
PrimeGrid. (NVidia project, but has a slow AMD app alternative, single precision hardware)
First off - and this is real important - ignoring it is a "jump off the Cliff moment" :) - Single Precision (SP) or Double Precision (DP) hardware. Lots of factors involved, but what it boils down to is Double Precision (DP) hardware is accurate to 12 decimal places, and Single Precision hardware accurate to 8 decimal places. To be DP capable the graphics card needs a particular type of Floating Point Unit built into it. Its therefore a fundamental design issue by card Vendors, and a key one for you, because, you get this wrong, you limit yourself for the life of a card to no DP projects which can be a real pain. All cards are SP capable without exception, the question is which are DP capable.
AMD have the better DP capable ratio cards - but careful on broad assumptions, check detail, for now though, park that key thought in the mind. Not all cards are DP capable, and at the budget point you have, need to be real careful as it will be a strong factor in card selection - for now be aware of the issue, and start to get into the habit of absorbing if a card is DP capable or not, and whether or not a Project needs DP cards.
Single Precision Applications will run faster than DP applications - that therefore seems a no brainer, however, go Single Precision Card and (for example) you'll never crunch Milky Way ... ever ..... so pause, think, and choose wisely. Nose around ask those who crunch there, even join the board and post a question.
Card Vendor choice is going to be a crunch point for you. Its looking like AMD (there's a but coming later), as your preferred Project is Milky Way (needs DP cards), and Prime Grid on your preferred list has an AMD application albeit slow one.
Now the "but", and its a big "But". Science applications at present are really the purview of NVidia (in the consumer world area of graphics cards) due to their application software (CUDA). AMD propriatory software (Stream) has been parked up, and AMD are going Industry standard (OpenGL) to write their applications. This makes a lot of sense for AMD (they always take the Long View on Strategy), and their new architecture (GCN) on the new 79XX cards tie into this seemlessly. At your current budget point, you are looking at NVidia for Science Projects, and AMD for Maths projects as a preferred option. The trip and trap is what the AMD application is written in, if its OpenCL, then ask hard questions and be careful over performance at your budget point. For 79XX and GCN, this will be the big change for OpenCL, but the latter is outside your sphere at present, just be aware its there so you can recognise the issue and understand where it places itself to avoid confusion, and so you can sidestep the uninformed waffle about this, thats out there in BOINCLand.
PC Hardware
Following seven aspects need looking at (not exhaustive, can epeen nit-pick many other points on hardware, but this will get you 99% there):
Mainboard
Main Memory
SLI/Crossfire Capable
CPU
PSU
Case
Cooling
Mainboard
Check the PCI-E lanes available by looking at the mainboard vendor Site specifications. You are looking for capabilities expressed as: x16 and/or x8 and/or x4 and/or x1 PCI-E Lanes. Its likely expressed similar to : ..... the board has 32 lanes available split into following options ..... blah blah blah..... at a guess yours likely to be x8 slots on the board, x16 slots if your lucky. Just note down the number of card slots for each board, and post the PCI-E lane distribution for each one.
Main Memory
Not usually a factor intodays arena of large quantities of Gb memory stores, but worth checking how much main memory a GPU app needs. Most need around 50-100Mb or less per application instance, but some such as Collatz and a Clean Water application at WCG can use truckloads of memory per application 1Gb+ per application running. All your preferred list are light on memory use so you should be fine - but
find out, the knowledge is useful for other GPU related aspects later on.
SLI/Crossfire Capable
Mainboards vary as top what they natively support for dual card setups, and whilst you are only looking at one card, it would be dumb to go down a cul-de-sac ignoring possible expansion in the future - at least find out and keep your options open. While you are there, find out the power draw needs of the board in terms of Watts (its for PSU sizing later)
CPU
Not really a factor for card selection, but just be aware it is a factor for their use. A GPU is
utterly dependent on a CPU, a GPU cannot survive or work in isolation. Most times the CPU is only lightly using (say 0.05%) of the CPU capacity, but it can be much more. Therefore there is a drain on CPU capacity albeit only a very minor level, it will however slow any WUs running on the CPU slightly by 3-30 seconds depending on the CPU WU. The ones on your list are well behaved, but go and find out the size of the app in memory, and CPU needs from the GPU, and lodge that as another factor in card selection.
PSU
This topic (and the one on Heat below) is critical and under-rated by too many, who then wonder why they have issues. Graphics cards eat PSUs for breakfast, and pretty much take whats on offer with no fear or favour. Its a Showstopper topic which can take you right back to square one in terms of logic for card selection. With the PSUs you have on your PCs I would use the following planning criteria; 450w PSU plan a maximum draw of 400w keeping 50w in reserve, and the 350w plan on 300w used leaving 50w in reserve (always always always always x 1,000,000 always, leave at the very least 50w spare capacity, ideally only plan to use 60% of the capacity (impossible in your case, but keep 50w capacity back as spare). This is likely to limit card choice, might even push you into a new PSU purchase, complex topic we will deal will separately if it comes to it. Right now, look at
http://www.guru3d.com in their hardware review section for the card you are looking at, go to the power & heat section of the review, and note what they are saying. You can trust them absolutely - if they said black was white, I'd be rethinking my own logic to date as a result
![Smile :)](./images/smilies/icon_e_smile.gif)
They are good. don't get a downer about this as its likely you will hit a power ploblem - we'll deal with it when the time comes, just note the parameters and size of any problem.
Case
A real doozy of a problem as many are not used to selecting PC Case design based on needs of heat from GPUs. The problem is GPUs chuck out heat on an Industrial Scale rofl
![Smile :)](./images/smilies/icon_e_smile.gif)
You ignore this at pain of melting motherboard, graphics card, CPU and your power bill. We will deal with it when card is selected as one of the last (but arguably the most significant) issue to look through. The addition of a (quality) high pressure fan or two at crica £15 each can work wonders in this area, but more later ir can be complex to deal with sensibly as opposed to flippently - just be aware of the issue and importance. More GPU cards have melted, and household power bills been ruined because this was ignored than any other factor.
Well.... thats enough for you to think about this time round, and give you some "homework" to do rofl
Depending on your thoughts and findings, we will probably start talking actual card selection and the aspects to look for in the card specifications published by NVidia and AMD, on the next post.
Regards
Zy