Originally Posted by Gungnir
Originally Posted by sdlvx
However I do not think the northbridge is dead. If Hypertransport 4.0 comes out and it requires significantly less die space on a die, then there is no reason why AMD could not run two HT 4.0 links to the NB and then the NB could switch those to PCIe 3.0.
This is what I think AMD is thinking of doing and they're waiting on HSA for that. If HT 4.0 comes in at 52GB/s and AMD runs two HT links to HTX slots that's 104GB/s of bandwidth between CPU and GPU.
HSA needs optimal latency as well, though, does it not? A northbridge just isn't going to beat an on-die PCIe controller for latency, and giving more bandwidth to the GPUs than they can use won't do you any good.
Besides, NB is just an antiquated design: both AMD and Intel are continually moving toward complete SoCs in the names of power efficiency and reduced MoBo complexity. Opteron might keep HT for multiprocessor communication (though probably not), but I doubt that AMD is going to make a new non-server platform with either HT or an NB.
I may be wrong, but that's how I expect AMD to move.
AMD PDFs for developers which outline HSA state that CPU is for low latency and GPU is for throughput. Meaning that it sounds like latency intensive tasks (like GUIs) will still be done on the CPU while throughput bottlenecked tasks (think encoding video) will be done on GPU. (http://developer.amd.com/wordpress/media/2012/10/hsa10.pdf
Also, look here:
What I'm getting out of that is that Hypertransport offers better latency. Hypertransport is latency most important and throughput least, StarFabric seems to be throughput most important and latency least, and PCIe is somewhere in the middle.
What I'd like to see is how much die space Hypertransport takes up compared to PCIe.
To be honest I would think we'd see Hypertransport links directly to GPUs (which would offer better latency than PCIe and hopefully take up less die space) with a northbridge to swap to PCIe for legacy products. If AMD wanted to be a real jerk they could make all their GPUs with HTX or PCIe compatibility and completely lock Nvidia out of AMD's platform unless Nvidia submitted to more AMD standards and made their chips Hypertransport compatible.
I don't buy the SoC design where everything that isn't a part of a single die is old news. Take a look at Apple, they added all sorts of off-die co-processors to the newest iPhone and it let them power gate things a lot better.
I don't believe AMD can go APU only, they have
to get HSA working on non-APU systems at some point down the line, and it'd be very profitable. If they could get Autodesk, Adobe, and office software like MS Office and Libre Office (oh hey they already did the open source office thing
) to use HSA then they could completely clobber the professional market. Who is going to care if a Xeon with 20 cores is faster than an AMD Opteron with 20 cores, when you can go with the Opteron platform with HSA and have something 5 times faster than the Opteron?
AMD making noise about professional graphics cards makes me think this is the direction they want to go in. If they don't, HSA is going to be relegated to niche workloads like custom solutions for data mining and video transcoders. It has a lot more potential than that, and I feel that AMD feels the same as well as they keep pushing HSA features that seem more targeted at consumers (gesture recognition, facial recognition, etc).
Watch what kind of stuff AMD announces at APU13 and think about the use cases for the software they're talking about and what kind of hardware you'd need to back it up.
Take a look, you'll notice Adobe shows up quite a bit between GPU accelerated video editing
as well as DRM and photo editing.
That market would be absolute gold for AMD if they got HSA working and it'd leave them in a position where Nvidia and Intel would be forced to compete with HSA by using traditional computing means.