Yeah. Half the power with 1/3rd the memory. Fascinating.
Half the power with 50% higher bandwidth. If you raise the bandwidth of the GDDR-5 to the same level as HBM by incising the bus width and or memory frequency, you will get close to double the power consumption for the same Bandwidth.
Now tell me that this is not a big drawback for the mobile segment.
Do either of you have sources for your power numbers?
Yeah. Half the power with 1/3rd the memory. Fascinating.
Has anyone decreased the memory bandwidth of a Fury card to see how it effects performance? Curious to see if the increased bandwidth is being utilized...
Is that even possible?
Fury X benefits noticeably from overclocking the HBM. It's actually limited by the bandwidth, though I think that's because it needed a higher ROP count.
You're attempting to change the scope of the point.
I thought you were comparing power consumption, not bandwidth differences?
You're attempting to change the scope of the point.
I thought you were comparing power consumption, not bandwidth differences?
HBM uses half the power but needs to address MUCH less address space as opposed to 12GB on the TitanX for example. And HBM doesn't achive higher bandwidth through insane power consuming clocks but by the bus width instead. So now, pretend that HBM could have had 3 times the density as it actually does in FuryX. 12GB. I think it might require some more power to feed that extra 8 GB and all it's physical connections.
And since you brought up bandwidth, it doesn't seem, at least in this incarnation of HBM, that it's higher bandwidth is tremendously beneficial..... Don't get me wrong, I'm faulting the GPUs and not the HBM technology.
Maybe next time and I'm actually hoping GPU's can make use of that higher bandwidth next gen.
Well, for the given capacity, which was never a problem - see professional cards.
GDDR5/X memory lacks bandwidth which in turn requires a wide memory bus, for example 256bit wide. Now, that 256 bit bus is actually made out of 8 32bit controllers, each paired with GDDR5/X memory module. 8 modules minimum for 256bit bus. Increasing density doesn't change it.
Lets stick to 28nm becasue it is considerably cheaper, easier, and have better performance. 14/16nm go home.
We've been hearing from forum posters for a while now that graphics cards are not bandwidth limited. It's just people trying to counter that their favorite brand has less bandwidth. If it wasn't needed then both companies wouldn't be working so hard to increase memory bandwidth. It's pretty clear that both AMD and nVidia consider increasing bandwidth a priority.
People say that because it's true. Find me any 980 Ti/Titan X game benchmark that show large performance increases from increasing memory speed. I'll save you the trouble, there isn't any. I don't even bother with overclocking my memory because it yields no gains in performance. For this generation of 28 nm chips, HBM is overkill even at 4k. I do think HBM is the future for top end cards and eventually should fill the entire product line due to the power savings it offers but for someone like NVIDIA, power draw isn't really a problem right now unlike AMD who needed it for Fury.
So, you're implying that Fiji's power efficiency improvements relative to Hawaii are actually almost entirely due to architecture improvements, meaning that current HBM is a failed technology with only tiny benefits? If those numbers are actually accurate, AMD would have been better off releasing a 300W TDP card with 8GB GDDR5 for $50-100 less and would have had the same performance.
Though, this does assume that the numbers are accurate in spite of not making logical sense...
Why would you assume that the only way to increase the capacity of HBM is to increase the bus width? Even from the link you provided, the 290X and Titan X have similar bandwidths and DRAM power budgets, even though the Titan X addresses three times the memory. The address space doesn't determine the power used; bandwidth does. Fury X is limited to 4GB by the use of 256MB dies, but that's not a limit of HBM itself.
If HBM is overkill and has more bandwidth than is needed, then why does overclocking the HBM on a Fury increase performance? That doesn't seem to make sense to me. I would think that if there was more bandwidth than was being used, then increasing it wouldn't make a difference.
If HBM is overkill and has more bandwidth than is needed, then why does overclocking the HBM on a Fury increase performance? That doesn't seem to make sense to me. I would think that if there was more bandwidth than was being used, then increasing it wouldn't make a difference.
Overclocking just the HBM increases performance in the low, single-digit range, and therefore has little effect.
Yeah. I didn't do that. Not even a little.
You're attempting to change the scope of the point.
I thought you were comparing power consumption, not bandwidth differences?
HBM uses half the power but needs to address MUCH less address space as opposed to 12GB on the TitanX for example. And HBM doesn't achive higher bandwidth through insane power consuming clocks but by the bus width instead. So now, pretend that HBM could have had 3 times the density as it actually does in FuryX. 12GB. I think it might require some more power to feed that extra 8 GB and all it's physical connections.
And since you brought up bandwidth, it doesn't seem, at least in this incarnation of HBM, that it's higher bandwidth is tremendously beneficial..... Don't get me wrong, I'm faulting the GPUs and not the HBM technology.
Maybe next time and I'm actually hoping GPU's can make use of that higher bandwidth next gen.
Then I must have misunderstood you, because it really does seem like that's exactly what you implied.
There's no reason that HBM can't use denser dies to achieve more total VRAM, without increasing the number of I/O, in a similar way that the TitanX addresses more VRAM than the 290X in part due to using 512MB GDDR5 vs the 256MB GDDR5 in the 290X, without using more power.
Teal, you're missing the point. How can you (not you personally, but anybody) compare the power consumption advantage of one memory technology vs. another one when the total memory of one is 4GB and the other 12GB?
That is the straight, laser focused point, without any segways, or derailments or any other BS that I was attempting to claim.
And please don't say the bolded above again. I never said this. It's what you've imagined I've implied. I never did. It's what you inferred.
It might just be the way Fury is designed, that doesn't broadly apply to all architectures though. That's why I specifically mentioned 980 Ti/Titan X, they don't get any tangible benefits in games from memory overclocking, even at 4k or with a lot of AA turned on.
Edit: Looked into the Fury X memory OC and it was only 1.6 fps:
I would hardly consider that worth the effort.