I still think someone needs to solder GPU chips directly to a motherboard so we don't have to work through a PCI bottleneck. But I guess they won't do that either... Capitalism and Markets - you gotta love how they end-run around science...

EDIT: Aaaaannndd... 'ang on a minute. What is the error correcting properties of that 128MB? I spent a lot of money on ECC Registered DRAM and get recovered errors something like once a week, on the average, on big, complex simulations (24GB grids and larger). So type of memory does matter.

AMD is getting close to that, with their recent hybrid CPU-GPU units, such as the A10-7850k and the AMD "Jaguar" processors in the latest PS4 consoles. The integrated GPU is pretty powerful and definitely will push a demand for faster RAM! If you look at what they did on the PS4, should give a good brief view of what will happen in the near future... The PS4 sports 8GB GDDR5 @ 5500 MHz embedded onto the motherboard!

RAM so close to the CPU should not have any errors, aside from whatever is used for consolidating the current CPU integrated cache mechanisms... if there are any additional errors, then the CPU needs fixing ;)

--------------------

But there is a characteristic/possible detail regarding why they haven't release such hardware sooner:

1. Would be excessively expensive and possibly not tested enough. Intel's IA64 comes to mind.
2. Would create a massive disturbance in the economy of the world.
   • Blue-ray didn't give much of a fight, given the broadband and SSD technologies being pushed for.
3. It's very risky for a company.
   • For example, AMD is taking a great risk with HSA, because software development is having a hard time catching up to this futuristic hardware development; their current mainstream CPUs aren't considered very powerful with current software, but if the software is properly optimized, these AMD CPUs with HSA capabilities can be more powerful than most of the current CPUs on the market.

On this line of thought, my favourite example so far is the Thunderbolt technology: http://en.wikipedia.org/wiki/Thunder...28interface%29 - this technology can potentially extinguish all other data communication cabling: SATA, USB, Ethernet, Infiniband, HDMI and whatever else there is on wires. Can you imagine the sudden economic crash that could occur if this technology could nearly and instantly replace all of these cables/protocols? Simply because a lot of factories would either close or be converted to producing this single technology. Worst even is if this could also provide high capacity DC power... it could almost render Wifi/Bluetooth useless, unless for portable devices... but then again, it can also expose us to greater insecurities, hence it still being in development... if we don't count on Apple selling it with their Macs.

So Intel has finally released a Xeon with the integrated GPU and on-socket RAM. I'll purchase one once they are available and do some real benchmarking to see if the 128MB L4 cache actually helps. Here is some more information on the chips: AnandTech

Does anyone know what would be required to actually use the GPU in OpenFOAM? From what I understand the real problem with using GPUs is the transfer of data across the PCI bus, which obviously wouldn't be an issue here.

AFAIK, there is no straight forward way for OpenFOAM to use the GPU. Most implementations I know of are tied to CUDA and Intel's GPU technology doesn't rely on it.
Either way, you can find several implementations listed here: http://openfoamwiki.net/index.php/GPGPU - from there, the closest one I can think of is the one by Paralution, because last I saw it had OpenCL implementations.