UniversalTruth
Veteran
As far as I understand, GDDR5 should be used as system memory. It has nothing to do with Globalfoundries.
Haswell vs Kaveri
- Thread starter AnarchX
- Start date
Here is some additional backup material for that BSN's GDDR5 story.
http://www.brightsideofnews.com/news/2013/3/6/analysis-amd-kaveri-apu-and-steamroller-core-architectural-enhancements-unveiled.aspx
http://www.brightsideofnews.com/news/2013/3/6/analysis-amd-kaveri-apu-and-steamroller-core-architectural-enhancements-unveiled.aspx
D
Deleted member 13524
Guest
This spoils all the fun.
Yeah, that doesn't make much sense from a cost / performance perspective.
I don't know - maybe the GDDR5 functionality in memory controllers doesn't add that much cost? Don't they have several GPUs with support for both DDR3 (without the G!) and GDDR5, which from they could "loan" the designs?
Yeah, if there's any truth to this, Kaveri's memory controller can handle either DDR3 or GDDR5 over the same 128-bit interface, like many recent low-end AMD GPUs.
Seems like a misinterpretation. The GDDR5 only makes sense as on package video ram, but you'd still be using ddr3 as system memory.
Sony plans to use GDDR5 as system memory - sure it's not running Windows but that doesn't mean that Windows or Windows applications couldn't work just fine with GDDR5 main RAM, too?
Well I was actually expecting that they just flat-out move to unified GDDR5 / DDR3 (high end / low end) for the entire system based on their Orbis APU plans. This might be a bridge-generation to that era when PC / tablet main-boards look more like graphics card layouts.
UniversalTruth
Veteran
Guys, I think we have to clarify one essential thing:
You can't use GDDR5 as
and DDR3 as system memory. It's either this or that.
You can't use GDDR5 as
and DDR3 as system memory. It's either this or that.
Okay, I take this to mean that the part has one 4x32bit wide controller that can either be yoked up to on-board GDDR5 or DDR3.
D
Deleted member 13524
Guest
I guess that supporting DDR3 or GDDR5 would be the only way to not render the current socket FM2 useless after a year and a half.
The only dual-memory setup I could see working would be a MCM or several layers of PoP (unlikely because of the heat)..
The only dual-memory setup I could see working would be a MCM or several layers of PoP (unlikely because of the heat)..
Hmm, not sure what I think of this. On PS4 it makes some sense because they are going for much higher BW numbers (hundreds) and they have a lot of relatively slow cores that hopefully are able to mitigate the latency problems more.
On PC, it seems far less justified. First, the memory and CPU would by-necessity be soldered in so no upgrades. Second, it's far more expensive than DDR3 and the kicker for me is that if they're only going for ~50GB/s, they could get that with quad-channel DDR3 and avoid all the aforementioned downsides. Doesn't seem like a very good trade-off IMO, but of course I'm sure AMD has considered the options and has some reason for the choice they've made.
On PC, it seems far less justified. First, the memory and CPU would by-necessity be soldered in so no upgrades. Second, it's far more expensive than DDR3 and the kicker for me is that if they're only going for ~50GB/s, they could get that with quad-channel DDR3 and avoid all the aforementioned downsides. Doesn't seem like a very good trade-off IMO, but of course I'm sure AMD has considered the options and has some reason for the choice they've made.
Quad-channel DDR3 is not an option for the segments these chips are meant for, and the likely direction of the market is that DIMMs are going to be a rarer option for the markets these chips are for.
GDDR5 might give enough of a boost to the upper pricing tier in benchmarks against Crystalwell, which sounds like it involves physical and packaging design AMD cannot match.
While it may be the case that a high bandwidth wide cache design might be better than brute-forcing things with GDDR5, it doesn't matter if AMD can't do it.
On the other hand, on-package (edit: or on PCB) GDDR is something GPUs do routinely.
GDDR5 might give enough of a boost to the upper pricing tier in benchmarks against Crystalwell, which sounds like it involves physical and packaging design AMD cannot match.
While it may be the case that a high bandwidth wide cache design might be better than brute-forcing things with GDDR5, it doesn't matter if AMD can't do it.
On the other hand, on-package (edit: or on PCB) GDDR is something GPUs do routinely.
Most users won't want to, or know how to upgrade. Even if he did, the cost of a pair of DIMMs is about $100, whereas totally new and better laptop or tablet is now hovering around $300-400 which makes an upgrade much less compelling. (You don't hear many enthusiasts complaining that their graphics cards don't have expandable DIMM slots which would also be slower.) Corporations will also want to merge the PCBs they're using for low end laptops, convertibles, and high end tablets, so using soldered memory will alleviate an APU performance bottleneck and keep out the need for cumbersome and slow DIMM slots un-suited for slim form factors.
Why isn't it an option? Can't be more expensive than GDDR5, so is it too big/power hungry or something? That'd surprise me a bit too as I don't recall GDDR being particularly power-friendly compared to low voltage DDR3 (but admittedly I'm no expert).
Right I think in the long term a full cache hierarchy is the way to go, but even if that's not feasible shorter-term it just seems like there are a lot of sacrifices that you make with GDDR5 to get a claimed 50GB/s throughput...
Meh, I'll wait and see I guess. Though "for the form factors that these are relevant to", I see the price as a problem.
Also Raquia I don't know where you buy your RAM but I can get 16GB of fast DDR3 for less than $100, 8GB for less than $50, etc. and either is far more than you're going to get of GDDR5... While I agree that replacable DIMMS are going to be less and less common, the underlying cost of DDR3 is indisputably lower.
GDDR5 as system memory seems like a bad idea since it's too expensive to make sense for desktops (where AMD's only real strength with APUs is cost) and too power hungry to make sense for laptops (vs using discrete GPUs).. Would having separate 128-bit DDR3 and GDDR5 channels really have cost them so much in die area to offset the system costs? Surely 8GB of GDDR5 is a lot more expensive than 8GB of DDR3 + 1-2GB of GDDR5 (can't imagine > 2GB is really justifiable for GPU only), and less memory in the end.
Seems like 16 GDDR5 chips isn't great for laptop PCB utilization either..
Or maybe they expect to only sell 4GB systems? That doesn't sound very good..
Seems like 16 GDDR5 chips isn't great for laptop PCB utilization either..
Or maybe they expect to only sell 4GB systems? That doesn't sound very good..
That's fair point; I got 32 gigs of fast DDR3 for around $130 not too long ago because my workflow could use it. For most people, having more ram than their equipment was installed with is giving much diminished returns from a few years ago; they would have a much better user experience from a more modestly sized but faster memory subsystem. Even soldered DDR3 in 128 bit configuration would give 2x the bandwidth of most slot based DIMM configurations as you mentioned. I fully expect APU PCBs of the future to resemble GPUs of today.