Next leap in graphics?

[BenchPress]

The current GPUs are too slow for real time Ray Tracing.

The current GPUs are far too inefficient for real-time ray-tracing. They are still being outperformed by CPUs.

The GPU architecture as we know it today is really only suitable for rasterization graphics. To move beyond that we need more flexibility, as offered by a CPU architecture, while keeping the massive parallelism. Intel's AVX2 and TSX are a big leap in the right direction to combine the best of both worlds.

Discrete graphics cards are dying out, and the GPU and CPU are fusing into a superior homogeneous architecture which will allow developers to create new experiences.

 

[aaksheytalwar]

Discrete gpus will,never die

 

[Shaydza]

Discrete gpus will,never die

I disagree. The vast majority of people barely use the equivalent of an i3... We are going to have to find a use for all that extra die space as the process shrinks filter through. Graphics processors are the perfect use for that extra die space.

Once a CPU can run a game above 30 FPS on 1080p in high detail settings, there will only be a tiny market for discrete graphics cards. Without economies of scale the discrete cards will become unaffordable and most likely evaporate.

 

[njdevilsfan87]

Once a CPU can run a game above 30 FPS on 1080p in high detail settings, there will only be a tiny market for discrete graphics cards. Without economies of scale the discrete cards will become unaffordable and most likely evaporate.

I don't think that will be possible for a while, mainly because IGPs are dependent on system RAM, which is a huge bottleneck.

 

[MisterMac]

I don't think that will be possible for a while, mainly because IGPs are dependent on system RAM, which is a huge bottleneck.

Wait for the Extra l4 Cache for Haswell

Don't worry about speeds of memory, that will be the smallest bottleneck and is intel's expertise.

Building fast and sturdy ringbusses.


Now it's a Bottleneck yes, but it wont be in a year or two.

 

[blastingcap]

I don't think that will be possible for a while, mainly because IGPs are dependent on system RAM, which is a huge bottleneck.

They're working on solutions, short-term and long-term. Here is one possible short-term solution if Charlie is to be believed (always a big if, but still...): http://semiaccurate.com/2012/04/02/haswells-gpu-prowess-is-due-to-crystalwell/

 

[ShintaiDK]

I don't think that will be possible for a while, mainly because IGPs are dependent on system RAM, which is a huge bottleneck.

Thats true. But 64bit DIMMs are overdue for an update. Plus DDR4 is getting close. One can only hope for 256bit DDR4 DIMMs.

256bit DDR4 3200Mhz(middle of the DDR4 spec) would give 100GB/sec. GTX 680 is 192GB/sec to compare. Even with the current dualchannel 64bit it would bring 50GB/sec.

Another option is to see 512MB-2GB perhaps of onpackage memory to the CPU.

 

[Cerb]

Wait for the Extra l4 Cache for Haswell

Slow caches are a problem, and another level adds latency, power consumption, or both. Intel's CPUs don't seem to have any issues with 3 levels, so I would hope that 4 is a long ways off. Might be useful for some Xeons, I guess, but not 1S and 2S machines, generally.

 

[happysmiles]

I disagree. The vast majority of people barely use the equivalent of an i3... We are going to have to find a use for all that extra die space as the process shrinks filter through. Graphics processors are the perfect use for that extra die space.

Once a CPU can run a game above 30 FPS on 1080p in high detail settings, there will only be a tiny market for discrete graphics cards. Without economies of scale the discrete cards will become unaffordable and most likely evaporate.

4k.

 

[Dark Shroud]

I know everyone doesn't like Rambus but XDR memory would be quite useful for the time being.