Wednesday, March 1st 2017

NVIDIA's AIC Partners to Launch GTX 1080, 1060 With Faster GDDR5, GDDR5X Memory

At their GDC event yesterday, NVIDIA announced a change to how partners are able to outfit their GTX 1080 and GTX 1060 6 GB models in regards to video memory. Due to improvements in process and scaled-down costs, NVIDIA has decided to allow partners to purchase 11 Gbps GDDR5X (up from 10 Gbps) and 9Gbps (up from 8 Gbps) GDDR5 memory from them, to pair with the GTX 1080 and GTX 1060 6 GB, respectively. These are to be sold by NVIDIA's AIB partners as overclocked cards, and don't represent a change to the official specifications on either graphics card. With this move, NVIDIA aims to give partners more flexibility in choosing memory speeds and carving different models of the same graphics card, with varying degrees of overclock, something which was particularly hard to do on conventional 10 Gbps-equipped GTX 1080's, which showed atypically low memory overclocking headroom.
Add your own comment

36 Comments on NVIDIA's AIC Partners to Launch GTX 1080, 1060 With Faster GDDR5, GDDR5X Memory

#26
RejZoR
newtekie1
RE7 on very high uses about 7.5GB@1080p. But that's largely attributed to the shadow cache. Turning that off just about halves the memory usage.
Halves the memory usage at NO performance or visual penalty. In fact I don't even know what's the purpose of this stupid setting. If you leave it on, it makes things so much worse on 4GB cards because it has NO VRAM usage check. The stupid cache will use all memory and it'll just stutter like insane. Turning this off on 4GB card and you can still have EVERYTHING else on max and it'll work smoothly.
Posted on Reply
#27
GoFigureItOut
Whatever happened to HBM? I thought NVidia was working on alternative to compete against it.... or may be GDDR5X is it??
Posted on Reply
#28
Fluffmeister
GoFigureItOut
Whatever happened to HBM? I thought NVidia was working on alternative to compete against it.... or may be GDDR5X is it??
They already use HBM2 on their lower yield high-end HPC parts such as the GP100, but I guess they don't consider it particularly viable or rewarding yet bringing it to consumer parts, maybe HBM2 production is one reasons why Vega hasn't shown up yet.

GDDR5X is a nice middle ground approach especially when combined with their shiny memory compression tech.
Posted on Reply
#29
londiste
GoFigureItOut
Whatever happened to HBM? I thought NVidia was working on alternative to compete against it.... or may be GDDR5X is it??
memory standard isn't really what either amd or nvidia does. amd had a(n exclusive) partnership with hynix for hbm. at the same time, another potential similar standard in form of hybrid memory cube was being developed/considered by a bunch of memory makers. at this point, hbm seems to have won and is a standard for stacked, tsv-connected kind of memory.

gddr5x significantly increases the bandwidth over gddr5 while remaining a cheaper option. memory configurations on different gpus appear to indicate, that hbm-s bandwidth bonus is not as significant as expected when compared to what can be done with gddr5x. it'll get there, it just will take some more time and refinement.

hynix's specs were in the news recently, hbm2 bandwidth is 1.6 Gbps (204.8 GB/s of memory bandwidth) per stack plus they confirmed this will be used in vega. so far, the pictures we have seen of vega show it has 2 stacks of hbm, that's a bandwidth of 409,6 GB/s. vega might turn up with 4 stacks of hbm but at this point in the game it doesn't seem realistic.

this is an impressive number but smaller than fiji's 512 GB/s and not too much ahead of 390x's 384 GB/s and actually behind titanxp/1080ti's 480 GB/s.
Posted on Reply
#30
Kanan
londiste
memory standard isn't really what either amd or nvidia does. amd had a(n exclusive) partnership with hynix for hbm. at the same time, another potential similar standard in form of hybrid memory cube was being developed/considered by a bunch of memory makers. at this point, hbm seems to have won and is a standard for stacked, tsv-connected kind of memory.

gddr5x significantly increases the bandwidth over gddr5 while remaining a cheaper option. memory configurations on different gpus appear to indicate, that hbm-s bandwidth bonus is not as significant as expected when compared to what can be done with gddr5x. it'll get there, it just will take some more time and refinement.

hynix's specs were in the news recently, hbm2 bandwidth is 1.6 Gbps (204.8 GB/s of memory bandwidth) per stack plus they confirmed this will be used in vega. so far, the pictures we have seen of vega show it has 2 stacks of hbm, that's a bandwidth of 409,6 GB/s. vega might turn up with 4 stacks of hbm but at this point in the game it doesn't seem realistic.

this is an impressive number but smaller than fiji's 512 GB/s and not too much ahead of 390x's 384 GB/s and actually behind titanxp/1080ti's 480 GB/s.
At least HBM2 still has less power consumption compared to G5/G5X.

I think the original plans for HBM2 were to run with 2 Gbps, but there were issues and it was reduced to 1.6 Gbps. Well, but maybe it will run with 2 Gbps after all, so that Vega has enough bandwidth.
Posted on Reply
#31
Vayra86
microsista
there are alredy games that use around 6gb, and some use even more than that, its not that hard to imagine that in a year or so we will be getting games that require close to 8gb, especially because xbox scorpio has 12gb of memory so developers will have to take advantage of that creating high-res textures for games
Meanwhile I'm still running everything on my 3GB 780ti... and it works perfectly.
Posted on Reply
#32
newtekie1
Semi-Retired Folder
RejZoR
Halves the memory usage at NO performance or visual penalty. In fact I don't even know what's the purpose of this stupid setting. If you leave it on, it makes things so much worse on 4GB cards because it has NO VRAM usage check. The stupid cache will use all memory and it'll just stutter like insane. Turning this off on 4GB card and you can still have EVERYTHING else on max and it'll work smoothly.
Yep, that's why I said I still think 8GB is overkill, and I'd be just as happy on a 4GB RX 480 as a 8GB.

My point was that game devs are starting to do things that use a large amount of VRAM that aren't overwhelming the GPUs. It used to be that you had to crank up settings to the point that the game was a slideshow to overwhelm 4GB of VRAM. But that isn't the case anymore.
Posted on Reply
#33
EarthDog
P4-630
I think 8GB vram on a RX480 is overkill for gaming since it's an 1080p card.
I don't. There are already several titles (more monthly) which can use 4GB of vram at 1080p. Imagine in 2-3 years at the 'end' ofits life?
Posted on Reply
#34
Vayra86
newtekie1
Yep, that's why I said I still think 8GB is overkill, and I'd be just as happy on a 4GB RX 480 as a 8GB.

My point was that game devs are starting to do things that use a large amount of VRAM that aren't overwhelming the GPUs. It used to be that you had to crank up settings to the point that the game was a slideshow to overwhelm 4GB of VRAM. But that isn't the case anymore.
8GB RX480 does indeed show tangible performance boosts in quite a few games.
Posted on Reply
#35
Tomorrow
something which was particularly hard to do on conventional 10 Gbps-equipped GTX 1080's, which showed atypically low memory overclocking headroom.
How so?

Most 1080's i've seen can achieve 11Gbps. I'd say that's pretty good. Only way the new stock 11Gbps G5X is better if this too can go to (say for example) 12Gbps.
Posted on Reply
#36
newtekie1
Semi-Retired Folder
Vayra86
8GB RX480 does indeed show tangible performance boosts in quite a few games.
Yes, but only because of a few settings that drastically increase VRAM usage and give no real visual benefit.
Posted on Reply
Add your own comment