Yeah I added him to my ignore list along time ago. I don't mind people who prefer one company over another and defend theirs, but he just comes out with so much garbage and hate I can't be bothered to wade through it.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
Micron to start mass producing GDDR5X memory chips
- Thread starter Gregster
- Start date
Yeah I added him to my ignore list along time ago. I don't mind people who prefer one company over another and defend theirs, but he just comes out with so much garbage and hate I can't be bothered to wade through it.
Game code is largely highly abstracted from the memory management/interface layer - though DX12 will change that a bit - currently everything related to the memory architecture is handled at driver and/or firmware level.
The optimisations you are talking about are largely about stretching that 4GB a bit further.
The optimisations you are talking about are largely about stretching that 4GB a bit further.
Nothing has been underwhelming except its initial capacity limit. Hence why they need to specially change memory behaviour for fury cards so they can run at high settings in the 4GB limit. None of the optimisations are for a type of memory.
It is kind of underwhelming from the perspective people were talking like it would be game changing in terms of performance while largely it has made no real difference as current GPUs architectures are far from handicapped by memory performance - and even being able to design an architecture without having to make any consideration for memory inefficiencies won't result in any significant change for atleast another generation yet.
That's not my understanding of memory management but I'm open to correction.
The caveat of the HBM on the Fiji cards is it excels at 4K and higher resolutions but truthfully, 4GB just isn't cutting it but 8GB plus on a card that is 50% faster will be sweet.
You don't software optimise for a type of memory unless you have two active and dissimilar memory systems. I.e one has better latency or bandwidth than the other.
The main optimisations as i stated for the fury cards are purely to do with the capacity limitations of the card. To allow it to run games at high resolutions and settings within the 4GB limit.
Any differences in hardware are dealt with by the memory controller on the gpu.
Also the games themselves do not deal with memory managment or even think about gpu memory management, it is the driver that does this.
Last edited:
Permabanned
- Joined
- 12 Sep 2013
- Posts
- 9,221
- Location
- Knowhere
Is that a known fact or just an opinion?
I also have a concern about whether having hbm drastically reduces a cards overclocking capabilities.
I'm currently leaning away from HBM next gen, if gddr5x acts just like gddr5 but with the extra speed I'll be watching with interest.
^^^ AMD made a statement to that effect when the FuryX was launched.
The only optimization AMD talked about with Fiji were to overcome the 4GB limitation. I expect they have to look at what resources are being used in different scenes of a game and manually pull our and put back resources without the game noticing. They might even go further than that, and notice a texture is used that is never ever visible up close for whatever reason so is only using the lower mip-map levels, thus they could scale texture sizes down.
Plus DX API and windows add memory overhead which could be optimized. With the 970 Nvidia put a lot of low priority junk in the slower 512MB.
The only optimization AMD talked about with Fiji were to overcome the 4GB limitation. I expect they have to look at what resources are being used in different scenes of a game and manually pull our and put back resources without the game noticing. They might even go further than that, and notice a texture is used that is never ever visible up close for whatever reason so is only using the lower mip-map levels, thus they could scale texture sizes down.
Plus DX API and windows add memory overhead which could be optimized. With the 970 Nvidia put a lot of low priority junk in the slower 512MB.
Associate
- Joined
- 3 Jan 2010
- Posts
- 1,379
Now your starting to unravel your own argument, you clearly have your usual childish bias against AMD and calling it a failure when we see it's actually going to go ahead and get the next iteration with HBM2 (and it's working VERY well for high end cards to help with more bandwith for 4k games) then yeah, it's stupid to call it a failure. It's been a success regardless of whether there is issues with the odd optimisation, a single flaw doesn't make it fail considering it's going to be improved upon. I smell a troll
Seriously lambchop, I wasn't going to weigh in because i figured people were jumping on you a bit much and I'll be fair and honest and say I agree you was right on gddr5x as it can be twice as much bandwith or whatnot but under pressure of argument you go and launch more tiring dribble that is just anti-amd childish sludge. Come on man, play catch up with the rest of us and we can all learn to play nicely with our toys together. When backed into a corner (even when it's minimal pressure and you have a reason for your argument) you still produce a lot of fud and end up pushing the aggressive agenda of pro nvidia bull. HBM has a lot more potential as others have stated, just get on board and you can enjoy gddr5x if you want but it does seem the lower end memory compared to HBM.
Last edited:
from what I can tell HBM 1.0 is better at moving large data ts 1 by 1, whereas GDDR5 is good at small data sets at a faster rate, but hbm2 will have higher frequency which will offset this disadvantage, I dont know for sure but I think thats whats causing the odd image quality performance scaling on fury x.
Its like a cargo ship versus a freight aircraft, they each have a different use case in which they excel
Its like a cargo ship versus a freight aircraft, they each have a different use case in which they excel
Last edited:
this should help you understand memory management of HBM, it's just for games that use more than 4Go, HBM is fast enough to swap memory around without getting in the way of the GPU, thing that you cannot do with GDDR5 instead it need to hug more memory.
and this trust you have for GDDR5 is weird, HBM is better in all aspects but price full stop, thats why even Nvidia is gonna use it for the high end from now on.
Last edited:
^^ Regardless of how fast HBM is if you are swapping off board then it is still limited by the speed of the PCI-e bus (or System RAM on slower systems). While AMD has managed to work around that with more success than I'd expected it is decidedly not a long term strategy.
Think of GDDR5 as a long train down a single track. While HBM is multiple smaller trains along multiple tracks. But all of the trains depart and arrive at the same time. They achieve the same thing in the end but in different ways.
And a major advantage with HBM is per column refresh, this lowers memory access latency as only columns instead of entire banks are refreshed.
I would think not, it is just to overcome the limitation of HBM 1 in the end. I still consider Fiji a pipe cleaner for their 14nm parts for HBM and interposers.
I think the thing is that both GDDR5X and HBM2 are supposedly not coming till much latter in the year, and we all want cards now. we are so sick of 28nm that the new stuff cannot come soon enough. I still would not be surprised to see a new GP104 based card or two shown and promised to be available soonish (2-4 weeks) at GTC. Maybe even a standard 8GB GDDR5 384bit card, with the top tier cards being HBM2 that might make sense. with the new node and Pascal's new architecture it could well be 50% faster than the outgoing 980ti.
Well one can hope anyway.
Well one can hope anyway.