• 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.

AMD Radeon R9 285 with Tonga GPU pictured

People missed the important part, the subheading

Three years later than planned but finally here
Click to expand...

Three years ago AMD showed stacked mem samples ;)

Fiji has HBM at a guess, some time next year. Don't think Nvidia are going HBM before 2016.

What will be interesting is how it's done. We might see a single stack of HBM as a cache initially somewhat like Intel's Iris pro setup but with a hugely reduced cost of not having to make one big die or very expensive edram or esram(forget which intel used of those) then have a "normal" off package memory. Like 8gb as normal plus 1-2 gb stack of really fast cache on die with lower latency.

I'd quite like to see the full thing in HBM, mostly to see how it's done but. If AMD are using HBM and saving maybe 40-50W compared to Nvidia on the same process node then AMD is going to have a very big advantage in terms of spare power to push up clocks and/or the ability to use a lot faster bandwidth and more memory.
 
drunkenmaster said:
People missed the important part, the subheading



Three years ago AMD showed stacked mem samples ;)

Fiji has HBM at a guess, some time next year. Don't think Nvidia are going HBM before 2016.

What will be interesting is how it's done. We might see a single stack of HBM as a cache initially somewhat like Intel's Iris pro setup but with a hugely reduced cost of not having to make one big die or very expensive edram or esram(forget which intel used of those) then have a "normal" off package memory. Like 8gb as normal plus 1-2 gb stack of really fast cache on die with lower latency.

I'd quite like to see the full thing in HBM, mostly to see how it's done but. If AMD are using HBM and saving maybe 40-50W compared to Nvidia on the same process node then AMD is going to have a very big advantage in terms of spare power to push up clocks and/or the ability to use a lot faster bandwidth and more memory.
Click to expand...

Would be awesome if this happens next year.
 
Could certainly be interesting, It would be nice to see one of these new technologies actually arrive and hopefully make a considerable difference.

whether it will mean that AMD has a serious upper hand for a while or not we will have to wait and see.

Sounds good though.
 
I'm not entirely sure when the 8-hi version is coming, afaik they are currently only doing 4-hi stacks, as in 4 or 8 2Gbit chips meaning it's currently 1GB capacity with 128GB/s bandwidth with 2GB stacks when 8-hi ones are being made. There are higher density versions but I presume cost goes up significantly.

Stacks included... I don't know the limit, well I know you can use as many stacks as you can fit on an interposer, but I don't know if there are current limits on interposer size.

IE interposers/cost/yields might mean you can't use more than 4 stacks currently which could limit it to 4GB and 512bit, or it might be easy as hell to fit 8 stacks and achieve 8GB and 1TB/s already, I don't know.

It's also possible we'll see massively different packaging, think those Pascal mock ups we saw where it's like a tiny card with very simple lay out, simplified power componentry simplified lay out. The massive majority of traces and reason for size of a graphics card PCB is the memory traces, so move then on package and you can like Pascal basically put a high end GPU into a package less than half the size. In terms of heatsinks it means creating a socket type design for which heatsinks attach is simplified with almost all the major heat sources in the same location and having all the mosfets in one location.
 
You must log in or register to reply here.
Back
Top Bottom