szatkus :
blackkstar :
I think perhaps you misunderstood. I was saying we'd get a Hawaii refresh to hold us over until 20nm chips came out. Looking at GTX 970 and 980 release, it looks like Nvidia is set to stick with that for a while, with maybe a new Titan class chip for a ton of money.
I'd sort of imagine a Hawaii refresh to stop the bleeding against 970 and 980 and then a 20nm new chip that's not Hawaii to compete against a new Titan in the spring (Fiji).
Yes, that makes sense.
BTW, you still believe in 20nm Fiji?
blackkstar :
AMD has an opportunity with a 2p platform like that, specially regarding the news that Intel is merging mobile and desktop and there's rumors swirling of basically them abandoning big cores and making some middle of the road Atom/Core chip.
If that did happen, Intel would basically hand over HEDT, professional workstations, and big number crunchers to AMD and HSA.
Plus, I find it a bit entertaining to go from "AMD is going to abandon all big cores and Intel will destroy them" to "Intel is going to abandon big cores while AMD releases a 16m/32c 2p platform"
I am attracted to what seronx is saying because I hit this earlier. If AMD is building unified memory architectures between CPU and GPU, why not make unified memory between many CPUs and many GPUs? AMD will have a strategic position where they can release a platform with many CPUs and many GPUs, and they are all HSA accelerated and all you have to do is recompile the software (as long as it's using OpenMP) and then you've got a computing monster on your hands. A 2p, 8m/16c CPU rig with HSA and two Fiji GPUs would be serious competition to things like Xeon Phi and CUDA. And with what AMD is doing with OpenMP, it'd completely bypass the issue of being forced to use the wrong device for the wrong job or to incur stiff performance penalties for transferring the entire workload in memory over a "slow" bus like PCIe 3.0 x16.
HSA JPEG decoding and HSA Open Office sorting is cool, but sorting numbers and loading thumbnails faster on a laptop seems like a bit of wasted potential with HSA, no?
There are always a lot of crazy rumors after news like that.
Intel get most of their revenue from "HEDT, professional workstations, and big number crunchers". Abandoning it for gadget lovers doesn't sound like a good idea, especially if dozen of ARM cores compete on that market.
I don't know, to be honest. It probably isn't 20nm from what I've heard, but 28nm is old by now. We saw first 28nm GPUs January 9th, 2012. We are coming up on 3 years of 28nm GPUs and Nvidia seems to be willing to ride 28nm on GTX 970 and GTX 980 for most of 2015.
I'm just conjecturing at this point. AMD usually beats Nvidia to a new node. Nvidia is legal cat fighting with Samsung and Samsung and GloFo are working together now on process nodes. So that leaves Nvidia with TSMC and 20nm TSMC is no good for GPUs. But GloFo 20nm or 14nm is IIRC.
Personally, you've got fx8350rocks in here talking about process nodes for Nvidia and AMD not being the same for GPUs anymore, Nvidia doing some odd things like releasing a faster chip yet focusing on marketing on efficiency, etc. And of course you've got GloFo issues with 28nn, 22nm/20nm, and AMD working with GPUs on GloFo process with APUs. Maybe 28nm GloFo is vastly superior to 28nm TSMC and that's all it is? Who knows. Perhaps I am just upset that we are looking at the possibility of more than 3 years of 28nm GPUs.
Also, the TIM has nothing to do with Intel bad temps. The process has terrible voltage scaling and it gets hot. There have been plenty of chips that use TIM and stay cool in the past. My old Opteron 165 used TIM before I dellided it and it ran plenty cool, even at 3ghz from 1.8ghz and a lot of voltage with a little Zalman cooler on it. Solder helps, but if the chip is going to run hot, it's still going to run hot even if you delid it. TIM is rated to transfer certain amounts of heat, and Intel CPUs work better without TIM because they are exceeding specs for the TIM between IHS and die. 7850k uses TIM and is fine because it doesn't exceed those specs. Everyone busy fawning over "muh nanometers" and saying Intel is best process without realizing they created a process that runs too hot for TIM yet breaks down when applying fluxless solder between IHS and die.
EDIT:
http://www.fudzilla.com/home/item/36285-tsmc-16nm-finfet-plus-in-risk-production
TSMC just hit 16nm risk production. No one else finds it weird that Nvidia releases 28nm chips at this time? Usually when you release a new GPU, you are stuck with it for a minimum of 6 months. Unless we're going to see something like GTx 970 and 980 get rebranded as mid-range chips while Nvidia re-names entire line up (unless they actually go with GTX 10x0 naming lol) and the new high end is 16nm? Nvidia's actions seem completely odd to me.