That begs the question:
Why now, not earlier? If there is such low hanging fruit, why not grab it? It's the same with NVs DX11 CPU overhead thing. They could have improved on that years ago but only did in light of mantle. I hate both IHVs for withholding such advanced and then when they are "necessary" due to the competition being better in a particular area, they tout it as the second coming...
From what I've read this is not the case HBM offers much better access patterns (for the lack of a better characterization) than GDDR5.Faster, higher bandwidth memory should not alleviate a memory shortfall by itself.
What AMD said was that present caching protocols waste a lot of memory and can be optimized to use less memory without a performance penalty. Once you have to use the PCIe bus, which you will have to do when the on card memory is fully utilized, will impact performance. AMD is planning on hardly ever reaching that point by clever caching routines.
If this is true then it does not matter the memory type as this technique will work with any memory type.
See, you didnt read the first quote I posted from AMDs engineerThat begs the question:
Why now, not earlier? If there is such low hanging fruit, why not grab it? It's the same with NVs DX11 CPU overhead thing. They could have improved on that years ago but only did in light of Mantle. I hate both IHVs for withholding such advances and then when they are "necessary" due to the competition being better in a particular area, they tout it as the second coming...
In fact, he said that current GPUs aren't terribly efficient with their memory capacity simply because GDDR5's architecture required ever-larger memory capacities in order to extract more bandwidth. As a result, AMD "never bothered to put a single engineer on using frame buffer memory better," because memory capacities kept growing.
They're implying it wasn't a fruit they needed to pick. That the extra memory was needed for a wider bus, and as such it wasn't crucial to worry about efficient memory usage...and the lower hanging fruit was probably the GPU itself, not the memory.
Now that they are limited to 4GB in this iteration, it is suddenly a lower hanging fruit as it would severely limit the card in 1440p and up.
Why do you call people with valid concerns idiots? What is wrong with you?
I would hardly call this arbitrary. True, there may be few games that benefit from more than 4 GB today, but they will only grow in number. Add to that CF/SLI and 4K and one can certainly understand why someone would not want a 4 GB enthusiast card. Especially not one for $849 (if that price holds true).
It's obvious what he means. If quadfire was better than quad sli on every single test, people would still go "ew 4 GB of vram" and take more vram even if there was 0 benefit on vram bound situations.
He's saying a lot of potential customers will not look at testing and will see 4 vs 8 and that will make the decision for them. The second amd publicly States it's a 4 GB card a lot of people will be done with the card no matter the perf
that may not be the case, the few people in this niche that care about memry configs above performance is a niche within a niche. performance/$ -as RS will attest to- is still the predominant factor for buying enthusiast class gpus. Memory doesnt give the top spots in the benchmark rankings.
performance/$ -as RS will attest to- is still the predominant factor for buying enthusiast class gpus.
From what I've read this is not the case HBM offers much better access patterns (for the lack of a better characterization) than GDDR5.
While that makes sense, it could have prolongued the lifespan of their older GPUs. Or Tonga, it would be great for Tonga. Oh well...I smell marketing BS, at least partially.
Can you expand on that? AFAIK latency, etc are pretty much the same.
Although HBM was built primarily to deliver more bandwidth per watt, Macri cites a host of reasons why its access latencies are effectively lower. First, he cites "very small horizontal movement" of DRAM, since the data paths traverse the stack vertically. With more channels and banks, HBM has "much better pseudo-random access behavior," as well. Also, HBM's clocking subsystem is simpler and thus incurs fewer delays. All told, he says, those "small positives" can add up to large reductions in effective access latencies. Macri points to server and HPC workloads as especially nice potential fits for HBM's strengths. Eventually, he expects HBM to move into virtually every corner of the computing market except for the ultra-mobile space (cell phones and such), where a "sister device" will likely fill the same role.
Do you think it's possible to do all possible research at once? All enterprises have limits. Time and money are always constraints.
From what I've read this is not the case HBM offers much better access patterns (for the lack of a better characterization) than GDDR5.
From the Tech Report article:
It's so much better than even that. Start at page 14. We are going from single command in DDR3 /5 to dual command & single bank refresh in HBM. The GPU can move stuff more often, which means better overall utilization. HBM is simply awesome.
http://www.hotchips.org/wp-content/...Bandwidth-Kim-Hynix-Hot Chips HBM 2014 v7.pdf
No it isn't and I hope RS wouldn't actually say that because it's so obviously completely false.
Just look at the marketshare numbers, buying habits, top seller lists on amazon etc. and you will quite quickly realize that price/perf definitely is not the top reason.
It's so much better than even that. Start at page 14. We are going from single command in DDR3 /5 to dual command & single bank refresh in HBM. The GPU can move stuff more often, which means better overall utilization. HBM is simply awesome.
http://www.hotchips.org/wp-content/...Bandwidth-Kim-Hynix-Hot Chips HBM 2014 v7.pdf
AMD did say they no longer wanted to be in the bargain bin, right? Higher prices would seem to go along with that.
Yeah, but if you're known as a bargain brand and you want to start charging higher prices, you better have all your ducks in a row. 4GB of RAM is going to be a severe impediment to charging anything over $499, and even that is pushing it.
Based on today's articles, it sounds like they're trying to work around the 4GB limit using driver hacks. Not at all impressed - these will stop being maintained when 8GB+ HBM2 becomes available next year and AMD wants to sell the next big thing. I've seen people praise AMD for continuing to maintain "old" cards in their drivers, but the only reason so much effort has gone into maintaining GCN 1.0 is that they're still selling it. Sure, you got great value for your money if you bought a 7970 back in early 2012, but that doesn't help someone shopping for a new GPU now.
If they really want to charge a "premium" price ($699 or more), then they not only have to beat Titan X by more than 10%, they have to do so in a manner that demonstrates that the RAM deficit won't be an issue. This means it has to beat Titan X not just at 1080p but on 4K with ultra high-res texture mods, and on games that are not specially optimized for AMD.
This doesn't make sense. The "driver hacks" are just as relevant no matter the actual amount of HBM because they will undoubtedly increase speed/efficiency. In fact the driver/memory are made for each other they don't get abandoned by some arbitrary amount of memory.Based on today's articles, it sounds like they're trying to work around the 4GB limit using driver hacks. Not at all impressed - these will stop being maintained when 8GB+ HBM2 becomes available next year and AMD wants to sell the next big thing.
Yeah, but if you're known as a bargain brand and you want to start charging higher prices, you better have all your ducks in a row. 4GB of RAM is going to be a severe impediment to charging anything over $499, and even that is pushing it.
Based on today's articles, it sounds like they're trying to work around the 4GB limit using driver hacks. Not at all impressed - these will stop being maintained when 8GB+ HBM2 becomes available next year and AMD wants to sell the next big thing. I've seen people praise AMD for continuing to maintain "old" cards in their drivers, but the only reason so much effort has gone into maintaining GCN 1.0 is that they're still selling it. Sure, you got great value for your money if you bought a 7970 back in early 2012, but that doesn't help someone shopping for a new GPU now.
If they really want to charge a "premium" price ($699 or more), then they not only have to beat Titan X by more than 10%, they have to do so in a manner that demonstrates that the RAM deficit won't be an issue. This means it has to beat Titan X not just at 1080p but on 4K with ultra high-res texture mods, and on games that are not specially optimized for AMD.
This doesn't make sense. The "driver hacks" are just as relevant no matter the actual amount of HBM because they will undoubtedly increase speed/efficiency. In fact the driver/memory are made for each other they don't get abandoned by some arbitrary amount of memory.