Next gen memory bandwidth - any impacts on PC?

[escrow4]

So it seems that the Xbox One is running 2133MHz RAM giving it close to 70GB/s memory bandwidth. Will this make future next gen games more dependent on memory bandwidth? Makes me wonder, as I currently running a dual channel 16GB 1600MHz kit on a LGA 2011 platform, so I'd hit around 18GB/s.

Opinions?

 

[imported_bman]

That 70GB/s of bandwidth is for both the cpu and gpu. I suspect that APU and IGPU based systems will be less bottlenecked by memory bandwidth in the future. Quad Channel ddr4 will come close to matching the bandwidth offered by first gen of gddr5 cards (~100GB/s). A few sites have done benchmarks for APUs and IGPUs with different memory configurations, there can be significant gains in frame rates with greater memory bandwidth for this type of configuration. As for desktops with discrete GPUs system memory is not a bottleneck.

 

[Cerb]

The 70GB/s of the Xbox One is to cut costs. They are making up for it with the 32MB of much faster (I don't think specs have been given, yet) on-die local memory. We've been used to having around double the XBO's bandwidth for some time, now.

Our CPUs already benefit little from higher memory bandwidth, even AMD's, which are weaker in that regard than Intel's. In the case of a memory-sensitive AMD setup, that isn't an APU, you also don't want to go increasing bandwidth at the sacrifice of addressing latencies. Bandwidth is most useful for your CPU in that higher RAM bandwidth, at the same wall-time latencies, reduces the time taken to transfer data, thus allowing more time to be spent changing addresses, which is what DRAM is.

As to games and bandwidth, they keep using more for video, but less so than they can use more processing power. If new texture compression formats (hardware) and techniques (GPGPU) take off (more like Civ V's than Megatexture or the like), and tessellation takes off (which could also serve to reduce the need for big normal maps), that trend will only continue.

For instance, the 192-bit 660 is perfectly fine, until you turn up the AA, at which point the similar-cost 256-bit Radeons pull ahead, mainly due to more memory and texture processing bandwidth. The performance of the GPU itself has been able to increase at a much faster rate than the bandwidth available. As it is, affordable memory bandwidth is typically increasing as fast or faster than the games and GPUs need it to.

Typical CPU and GPU work are fundamentally different. Typical CPU work is worried about reading or writing small numbers of 64 byte (cache line size for most x86 CPUs for the last 10+ years) pieces of data at a time, from and to different locations. Typical GPU work is reading and writing big sequential arrays, typically ranging from hundreds of kilobytes to tens of megabytes in size. While physically vastly different, a hard disk drive would be a good analog, in that big sequential is faster, due to not spending much time seeking, while random is slower due to spending time seeking (except that 'seeking' is charging and discharging RC circuits attached to capacitor arrays).