• 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 Could Feature Next Generation HBM Memory on Volcanic Islands 2.0 Graphics Cards – Allegedly Laun

Boomstick777

Boomstick777

Boomstick777

Soldato
Joined
2 Jan 2012
Posts
12,445
Location
UK.
While NVIDIA plans to launch their high-end Maxwell architecture in the second half of 2014, AMD has also plans to introduce their Volcanic Islands 2.0 graphics cards during the same time. Overclock.net Forums have got some interesting details to share regarding the upcoming generation of AMD GPUs and from the looks of it, AMD could be the first graphics company to adopt the HBM Memory design.

Read more: http://wccftech.com/amd-feature-gen...ds-allegedly-launching-2h-2014/#ixzz31Rt01AiS
Click to expand...

eK9A71w.jpg
 
Something I've been wondering a long time, how much RAM can they stack? Going by that pic it seems only 1GB (2Gb x 4 = 8Gb = 1GB). Will that become like L2/3 cache and the rest of the modules on PCB like now?
 
Last edited:
AMD to launch new flagship Radeon graphics card this Summer

I recently had a very interesting conversation with one of the add-in-board partners of AMD. They told me that Radeon R9 290X will soon be replaced with a new graphics card. Graphics card that would take the performance crown back from GTX 780 TI.
Click to expand...

http://videocardz.com/50472/amd-launch-new-flagship-radeon-graphics-card-summer


28nm TSMC 1H 2014 : Hawaii, VI 1.0
28nm GlobalFoundries 2H 2014 : Iceland and Tonga, VI 2.0
28nm GlobalFoundries 1H 2015 : Maui, VI 2.0
20nm GlobalFoundries 2H 2015 : Fiji and Treasure, PI 1.0
20nm GlobalFoundries 1H 2016 : Bermuda, PI 1.0
14nm GlobalFoundries 2H 2016 : Mid-GPU and Low-GPU, PI 2.0
14nm GlobalFoundries 1H 2017 : High-GPU, PI 2.0
 
Last edited:
Boomstick777 said:
AMD to launch new flagship Radeon graphics card this Summer



http://videocardz.com/50472/amd-launch-new-flagship-radeon-graphics-card-summer


28nm TSMC 1H 2014 : Hawaii, VI 1.0
28nm GlobalFoundries 2H 2014 : Iceland and Tonga, VI 2.0
28nm GlobalFoundries 1H 2015 : Maui, VI 2.0
20nm GlobalFoundries 2H 2015 : Fiji and Treasure, PI 1.0
20nm GlobalFoundries 1H 2016 : Bermuda, PI 1.0
14nm GlobalFoundries 2H 2016 : Mid-GPU and Low-GPU, PI 2.0
14nm GlobalFoundries 1H 2017 : High-GPU, PI 2.0
Click to expand...


What?!?!?! already? i just bough a 290.
 
Orangey said:
Something I've been wondering a long time, how much RAM can they stack? Going by that pic it seems only 1GB (2Gb x 4 = 8Gb = 1GB). Will that become like L2/3 cache and the rest of the modules on PCB like now?
Click to expand...

You aren't limited to one stack ;)

However the issue becomes bandwidth per stack, 128GB/s per stack(closer to 256GB/s in a couple years) means 4 stacks for 512GB/s bandwidth. Now latency I'd expect to improve and simpler communication with the memory. Shorter traces and less power mean less overhead, simplified memory controller which should help with some better access to the memory.


Cache would increase power usage effectively as you'd be adding an on package chunk of memory, still require off package memory and a more complex memory controller which uses more power. It's the wrong direction so very unlikely it will happen.

If AMD does it is a complete unknown, these slides aren't new and there is a lot of guessing. Currently volume might be too low to make it viable cost wise. We might well see it on firepro's or something for a generation and see something like £1.5k cards in apple mac's subsidise early production till volumes/yields get good enough to be viable in desktop gpu's.

Where it gets a little interesting is interposes are actually being made at 65nm, because they have the capacity available all over the world and interposers (effectively a circuit board but at silicon/nm trace widths rather than pcb + solder scale, so 65nm is still a magnitude smaller, and crazy cheap). The HMB can be created on 20 or 14nm and is packaged to be stuck on an interposer, and any chip can effectively be stuck on the interposer with it.

So we're at the point where it's possible to make a 65nm interposer, stick a 28nm gpu in the middle and 4 stacks of 20nm HMB around it.

It's likely(I'm not 100% clear on the tech) that you need a different memory controller for HMB, so considering the existing chips at 28nm, the work done, it's a fairly large step to make such big changes to current chips rather than make the change in the next gen.

The issue does come in that more stacks decreases yields. The biggest issues with interposers is. Make a set of chips like gpu's and say have a 80% yield, you throw the 20% away and the rest are working. Likewise you have working HMB stacks, now stick them all together on an interposer and, while you can resolder a chip on a pcb relatively easily, you can't "re-do" an interposer, if a single chip you add fails, then the whole thing is screwed. So you can have £100 of memory, £150 of gpu, £5 of interposer, all working, stick them together and not working. More things you stick together, more chance of failure. So more stacks is an issue for sure. That is where interposers start to hurt yields, the more things you stick together the worse the yields and after 5-6 chips the yields tank horribly.

So while interposers will let them stick things together to create silly fast connections between chips, very close to as if it was all on the same die, the downside is the more bits you stick together the worse the yields get. The memory stacks themselves are only up to 4 because currently the yields aren't there to do the 8 high stacks. But in 2 years you can have 4x 8 high stacks for 1TB/s of bandwidth with insanely higher yields than you'd get with 8 x 4 high stacks, same bandwidth possible but yields would drop to make prices not viable.
 
humbug said:
What?!?!?! already? i just bough a 290.
Click to expand...

Nothing as yet is known about those two code names, they may never launch, they might be low end or midrange. They could be faster than a 290 or merely on par with a 290 but made at GloFo. AMD appears to be making it's move from TSMC to GloFo and that could be another huge advantage. 20nm at TSMC is available, it's in full production, just that Apple is buying up every wafer and making it very very hard for anyone else to get enough wafers to release a product.

A lot of people suspect Beema/Mullins is mostly a result of improved process and porting from TSMC to GloFo, if you managed 20% less leakage on a 290x, even a similar chip could clock higher at the same power usage. They aren't necessarily going to make something significantly faster than a 290x, might not be any faster, lower power, higher yields or simply higher margins for AMD.


AlamoX said:
thought no 20nm for AMD untill 2015, or is that just for cpu ?
Click to expand...

Who said 20nm in 2014?

If those dates/targets/process size are remotely accurate it says 20nm 2h 2015.

Does it mean anything, who knows, but Iceland has a HUGE area, 100,000km^2, Tonga is very small and Maui is fairly small, with Hawaii being the second largest..............
 
drunkenmaster said:
Nothing as yet is known about those two code names, they may never launch, they might be low end or midrange. They could be faster than a 290 or merely on par with a 290 but made at GloFo. AMD appears to be making it's move from TSMC to GloFo and that could be another huge advantage. 20nm at TSMC is available, it's in full production, just that Apple is buying up every wafer and making it very very hard for anyone else to get enough wafers to release a product.

A lot of people suspect Beema/Mullins is mostly a result of improved process and porting from TSMC to GloFo, if you managed 20% less leakage on a 290x, even a similar chip could clock higher at the same power usage. They aren't necessarily going to make something significantly faster than a 290x, might not be any faster, lower power, higher yields or simply higher margins for AMD.
Click to expand...

So they could be the same GPU under the same name only with an improved process.

A bit like Bulldozer to Vishera / Trinity to Richland, actually the same CPU but improved, they may even do it silently.
 
drunkenmaster said:
You aren't limited to one stack ;)

However the issue becomes bandwidth per stack, 128GB/s per stack(closer to 256GB/s in a couple years) means 4 stacks for 512GB/s bandwidth. Now latency I'd expect to improve and simpler communication with the memory. Shorter traces and less power mean less overhead, simplified memory controller which should help with some better access to the memory.


Cache would increase power usage effectively as you'd be adding an on package chunk of memory, still require off package memory and a more complex memory controller which uses more power. It's the wrong direction so very unlikely it will happen.

If AMD does it is a complete unknown, these slides aren't new and there is a lot of guessing. Currently volume might be too low to make it viable cost wise. We might well see it on firepro's or something for a generation and see something like £1.5k cards in apple mac's subsidise early production till volumes/yields get good enough to be viable in desktop gpu's.

Where it gets a little interesting is interposes are actually being made at 65nm, because they have the capacity available all over the world and interposers (effectively a circuit board but at silicon/nm trace widths rather than pcb + solder scale, so 65nm is still a magnitude smaller, and crazy cheap). The HMB can be created on 20 or 14nm and is packaged to be stuck on an interposer, and any chip can effectively be stuck on the interposer with it.

So we're at the point where it's possible to make a 65nm interposer, stick a 28nm gpu in the middle and 4 stacks of 20nm HMB around it.

It's likely(I'm not 100% clear on the tech) that you need a different memory controller for HMB, so considering the existing chips at 28nm, the work done, it's a fairly large step to make such big changes to current chips rather than make the change in the next gen.

The issue does come in that more stacks decreases yields. The biggest issues with interposers is. Make a set of chips like gpu's and say have a 80% yield, you throw the 20% away and the rest are working. Likewise you have working HMB stacks, now stick them all together on an interposer and, while you can resolder a chip on a pcb relatively easily, you can't "re-do" an interposer, if a single chip you add fails, then the whole thing is screwed. So you can have £100 of memory, £150 of gpu, £5 of interposer, all working, stick them together and not working. More things you stick together, more chance of failure. So more stacks is an issue for sure. That is where interposers start to hurt yields, the more things you stick together the worse the yields and after 5-6 chips the yields tank horribly.

So while interposers will let them stick things together to create silly fast connections between chips, very close to as if it was all on the same die, the downside is the more bits you stick together the worse the yields get. The memory stacks themselves are only up to 4 because currently the yields aren't there to do the 8 high stacks. But in 2 years you can have 4x 8 high stacks for 1TB/s of bandwidth with insanely higher yields than you'd get with 8 x 4 high stacks, same bandwidth possible but yields would drop to make prices not viable.
Click to expand...

Thanks, great info as always :)
 
I thought this was going to be ext year, now im worried if I a new card and something better comes along a month or so down the line.
 
It's pretty rare for a product to launch out of nowhere. We usually get some info on a tape out 3-6 months before a product is due(tape outs take 6+ months but don't always hear about them straight away). Then you usually get more and more leaks/rumours leading up to a launch, probably 2-3 months out is when you start to get a hint at specs/realistic dates.

I'd be pretty surprised at this point if anything was going to come out before Aug/sept at he earliest.
 
AMD to launch new flagship Radeon graphics card this Summer

AMD to launch new flagship Radeon graphics card this Summer

I recently had a very interesting conversation with one of the add-in-board partners of AMD. They told me that Radeon R9 290X will soon be replaced with a new graphics card. Graphics card that would take the performance crown back from GTX 780 TI.

If you remember Radeon HD 7970 GHz Edition, then you probably know where this is heading. There are basically two options, yet another Hawaii variant, or brand new GPU. It’s worth noting that there are some traces of a new GPU called Iceland. Could it be the new high performance heart of new flagship?

We have our own theories about this card. It’s not even a year since R9 290X was announced, so it’s highly unlikely we are expecting R9 300 series this soon. The easiest, and probably the most logical approach, would be to launch R9 295X (note the missing 2). Although the new naming nomenclature is not among our favorites, it has some room for adaptation and AMD can always add an extra ’5′ if needed. Of course this happened already, AMD launched first 2×5 graphics cards few months ago, including 235, 255, 265, 275M and the latest addition R9 295X2. That said, it is possible AMD will simply launch R9 295X, instead of whole new series (R9 390X).
Click to expand...

Source
http://videocardz.com/50472/amd-launch-new-flagship-radeon-graphics-card-summer

I wonder if this could be a full fat Hawaii core. There were rumours that there is part of the XT chip missing and the full amount will take it to around (roughly) 3000 shaders. If this is true, the reference cooler won't cut it so unless they've designed a new cooler expect this to be a AIB cooled card only. My two cents.
 
You must log in or register to reply here.
Back
Top Bottom