[ToTTenTranz]

The GK107 "Mini-Kepler" has been taking design wins on laptops by storm, since its performance/power consumption is really great.

However, most of these laptop models (and actually the only desktop implementation so far) come with DDR3 memory @ 1600MHz using a 128bit bus.

This results in 28.8GB/s total bandwidth, which is a bit suffocating for a GPU with a performance-per-clock equivalent to AMD's Juniper (800sp, 32 TMUs, 16 ROPs).

I have one of those in my laptop, and for its 1366*768 resolution, the sub-par memory bandwidth doesn't make that much of a difference.

Plugging it to a 1080p TV, it's a different story. I've tried overclocking the memory to ~2200MHz, increasing the bandwidth to ~35GB/s, and the performance bump is almost linear with the bandwidth increase (not a surprise, I know).


That said, I also found out that most of these Mini-Kepler GPUs are coming with Ivy Bridge systems and PCI-Express 3.0 connections.

Now, a PCI-Express 3.0 16x link does a maximum of 16GB/s. Adding those bandwidths up would result in almost 45GB/s, which is a solid 55% increase in memory bandwidth.

And even if the PCI-Express link could only be used at half its theoretical speed, we'd still have some 27% bandwidth increase.


So why isn't nVidia implementing Turbocache for this GPU? Could it be implemented with a simple driver update?
I know this would introduce lots and lots of horrible latencies, but couldn't the driver "choose" to only load assets that are less latency-sensitive through the PCI-Express bus?

[swaaye]

It's not a new idea.
http://forum.beyond3d.com/showthread...memory+enabled

I have a feeling that this tech is really only practical for those ultra low end budget cards it created.

[ToTTenTranz]

Originally Posted by swaaye

It's not a new idea.
http://forum.beyond3d.com/showthread...memory+enabled

I have a feeling that this tech is really only practical for those ultra low end budget cards it created.

In 2005 we had PCI-Express 1.0 (4GB/s @ 16x) and DDR2-533MHz (8.5GB/s @ dual-channel).

Both bandwidths are now ~4x faster in most recent PCs and the truth is that if the GK107 with DDR3 could use the extra bandwidth from the system memory effectively, there could be a significant performance advantage.
Both nVidia and AMD could save some money on using dirt-cheap DDR3 with their mid-range GPUs without losing too much performance.