You have confused me. The client that this is for has specifically stated that they wanted Quadros. FireGL cards are only an option if we can't get the Quadros to work. They're not really worried about IQ here. The RealiZMs don't appear to even be in production anymore, and are not readily available. The timeline for delivery is very short, so waiting weeks for anything was out of the question.
That's exactly why I don't allow clients to specify hardware; because the client doesn't know what is and isn't compatible, and what will and won't perform. They get bombarded with glossy print ads and benchmarketing. I work from a set of requirements - the system must do this, this, and this - and derive specs from there, for that very reason. And no, Creative raped 3DLabs, which is why I had to spend a lot on service parts. I think I'm the only source in the world for replacement parts now. Sigh.
And who does RAID-5 for performance? whether you're using onboard SAS controller, onboard SATA controller, or a dedicated RAID card, you still have to calculate parity, somewhere. They are not (for this machine) concerned with data loss all that much (a reload at this point only requires that they drop in my slipstreamed disk, and reload AutoCAD), they are concerned with machine downtime (that's why RAID10...as long as two disks in different sets don't fail), and time incurred for parity calculations (hence the RAID10 > copy on write...no parity calcs). They don't care about the data on the machine, per say (well, only so long as it takes them to do the work on the machine...then it gets dumped off).
Long thread elsewhere about it, and I'm aware of their workflow scheme. People here, generally don't know jack about RAID. They can talk a good game, but it's just a game. I've been doing it for years upon years. Parity is irrelevant when done correctly. They come up with any number they can to justify their ignorance. The reason for RAID5 is purely performance - you will not saturate any card with RAID10. RAID5 is faster, cold hard fact. The specific reason I stated LSI 8480 is that I'm familiar with it, and I've deployed it. It can do >475MB/s long sequential read on unaligned partitions with 7200RPM disks. Performance with 4+1 10k SAS, aligned, full stripe writes, exceeds 500MB/s typical. I didn't bother with long sequentials; at that much bandwidth, you're close to exceeding bus capability. Obviously, it's not an option since they insist on rushing it - another thing I don't let customers do - but it's basically the optimal configuration.
They do have a SAN, but the requirement for this project, and therefore, this machine is that it will be VLAN'd off, and cannot touch the SAN (but they do have one), or any other shared network devices. This machine has to be standalone, or have dedicated resources. $20k to these folks is a lot for a computer, so the idea that they would buy additional dual fabric config, RAID controller, and disk for this machine it out of the question.
Sigh. iSCSI is not a SAN, it's a joke. SAN is 4Gbit FC, 8Gbit FC. The point of a pre-processor in their scenario is really pretty simple. Rather than muss around with an extremely expensive and complex workstation, the majority of the heavy lifting is done by the pre-processor, which would typically be an xSeries or similar connected to FC disk, with a Tesla or Stream processor in it. This system performs the majority of the heavy lifting, allowing them to do less processing work on the workstation. In some situations, it can take a pair of $28k workstations down to a single $12k pre-processor, and two $5k workstations. But it's also generally more suited to a larger scale deployment, rather than a one-off.
The 30" screens that they went with were for the resolution, solely. they weren't at all interested in the overall IQ of the displays, response times, etc. Although they do intend to manipulate the objects in the environment on one screen, and have it rendered on the other, they don't do fluid animations, or renders that require fast response times, color gamut reproduction, etc. It's not that kind of application. They're doing Mechnical and Electrical work in AutoCAD, not video production, video editing, or animation stuff.
Uuuuuuugh, I think I know exactly what it is now. This really is a good illustration of why I work the way I do, and don't let clients specify components. You'll need to verify with nVidia that they don't still have the no dual-acceleration restrictions in place; used to be that if you wanted two accelerated displays from one card, tough, go buy another Quadro or a Quadro MVS (which was a glorified GeFarce2 MX.) The problem with MCAE is that despite application similarity, it's vastly different from CAD/CAE/CFD. Given the datasets you've mentioned, I don't know what performance will be; I haven't done any major MCAE since the REALiZM 800 was current. I can see why they would think dual Quadros; especially with nVidia's old single accelerated display requirement - but in hindsight.. I'm doubtful they would actually need them for any other reason.
Be interesting to see what nVidia and PNY say, or if they take the easy out and say that the 5520 isn't a supported chipset and you're stuck.