Re: OT: Drool fodder for CUDA devotees

[Erik Mickelson] "The card is Pcie3. That's the big deal. Twice the processing power of Pcie2."

PCIe 3.0 is indeed a big deal, but it means double the effective bandwidth, not double the processing power.

[Erik Mickelson] "It will not matter the slots multiplier, you will not have to worry what slots are open."

It would matter less compared to today's constraints on a Mac Pro, but we don't even know if the Mac Pro will have a successor, by any name, with slots of any kind.

[Erik Mickelson] "I don't think anything can saturate a Pcie2 bus except for uncompressed footage."

Uncompressed 10-bit 1080i29.97 is about 250 MB/sec, or half of one lane of PCIe 2.0. A 4x PCIe 2.0 slot would be saturated with 8 streams. The highest end GPUs can push past the bandwidth limits of PCIe 2.0 8x under load, but most GPUs do not. Putting lesser GPU cards in a 16x slot is for power requirements as much as it is for bandwidth.

[Erik Mickelson] "So put the raid on the x16 slot and the video card on a x8."

You'd need some seriously fast storage to pull more than a PCIe 2.0 8x slot can support, not to mention an HBA with more than 8 lanes (and I'm not aware of any PCIe 2.0 16x HBAs).

Someday soon we'll surely find ways to need all the bandwidth available in PCIe 3.0 16x and 8x slots (expansion chassis is the obvious application), but today, single cards of any sort will struggle to fill those pipes. That's a great problem to have. In a couple years, when we have 12 Gbps SAS3, we'll be glad to have PCIe 3.0 slots to saturate.

Then you can do all the stereoscopic 4K multi-cam you want!

Best,
Andy