SK Hynix showcases first GDDR6 - Double The Bandwidth
Click here to post a comment for SK Hynix showcases first GDDR6 - Double The Bandwidth on our message forum
SirDremor
Nice!
Is it only me, or does HBM(2) start to look more and more like an infamous "RDRAM" of modern days?
Hulk12
GDDR6 is for Hynix only!?? 😕 I don't like this brand "bugged" because EVGA GTX 580 3GB got freeze in some games while EVGA GTX 780 Ti Classified got render failed at many times with every minutes or hours.
Stormyandcold
"uneiled" face-palm.
Anyway, it's looking great.
labidas
SirDremor
PrMinisterGR
The Commenter
HBM looks really impressive to me. This video shows the true benefits of it.
youtube.com/watch?v=85ProuqAof0[/url]
Denial
http://i.imgur.com/ffXpyBf.jpg
Being able to have a small contained GPU core with memory onboard allows it to be stacked/configured far better than a giant PCB with GDDR.
HBM/Stacked ram is definitely the future. May take a while for all consumer boards to have it but it will come eventually.
Yep, latency is a huge factor, especially with deep learning/compute workloads. One thing people forget is size too:
NewTRUMP Order
Loophole35
HBM is the future but until it is more mature GDDR will be a main stay in the consumer level for a few more years expecialy with the bandwidth improvements we see here. Now I understand latency is higher on GDDR6 than HBM however considering we tend to focus more on gaming than deep learning it is almost a non factor.
HBM will be on professional and enthusiast products exclusively for I'd say about 3-4 more years before we see a mid range product with it ($300 and lower).
Exige245
I would think that the design concept of HBM would allow it to be implemented on CPUs in the future (not just GPUs) and thus change the landscape drastically as it would integrate system ram on the actual CPU chip itself...
schmidtbag
Hynix with GDDR is to HBM as Western Digital with HDDs is to SSDs.
GDDR6 is appealing because it [probably] is a cheaper option to HBM without being inadequate, just as HDDs are a cheaper option for extra capacity compared to SSDs. Both GDDR6 and HDDs are technically worse, but they still have good reasons to exist. It'd be interesting to see two models of the same GPU with GDDR6 vs HBM, though I don't expect we'll ever see that. The fact that Nvidia wants access to HBM shows promise in the technology.
TieSKey
schmidtbag
Fender178
Denial
http://electroiq.com/insights-from-leading-edge/wp-content/uploads/sites/4/2016/03/Sys-plus-1.jpg
This kind of shows the cost breakdown - but it doesn't break out the cost of the HBM module itself and the manufacturing process. Vega, at least the ones we've seen only have 2 stacks - so manufacturing should be a little cheaper than the 4 for Fury/GP100. But it's HBM2 and we don't know if the cost of those modules are higher/lower or what.
AFAIK the most expensive part is the TSV mounting process. They essentially grow tens of thousands of crystals through each stack and if a single one fails they essentially lose the entire stack or bin it as a lower memory chip. The process itself is far more complex than mounting GDDR so no matter what it's going to be more expensive.
HBM is a leap in performance for specific workloads but it doesn't really do much for gaming as far as we can measure - at least as long as there is sufficient memory in the first place. Vega has the cache controller which uses it's HBM a little differently and can boost minimums by a pretty awesome amount - but it only does that when it's out of memory in the first place. There could be future gaming implementations that utilize the latency reductions though. Bandwidth, given Vega's stack limit, isn't that much higher then what GDDR5x offers. Compared to the Ti it's not that much more. Power consumption of HBM2 scales differently too based on bandwidth:
https://www.extremetech.com/wp-content/uploads/2016/02/NV-HB.png
This is a little old - but depending on the implementation HBM can actually use more power overall.
So there are some design/cost tradeoffs.
TieSKey
Interesting info.
I didn't get the point with the power/bandwidth, HBM seems to scale a lot better so if u use more power with it, is because you are getting an insane bandwidth (unless gddr has changed it's projected curve since that graph was introduced.)
The cost breakdown is a little weird (but might be ignorance on my side). All that extra pcb, cabling (lanes/whatever) and the work that needs to be mounted costs only 10? I would think just the extra copper for memory heatpipes would cost at least that :P
Any idea why HBM is not being used in phones (premium ones at least)?
My whole point is just "I think companies are not investing enough in HBM as they are comfortable with this (really) slow road map."
(As a side note, I think the same is true for VR Headsets which still cost double of a whole console.)
schmidtbag
tsunami231
TieSKey