We can only hope
-
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.
Nvidia GTX Titan II Flagship with GM200 Core Is Launching Sooner Than Expected – Ready Within 2 Mont
- Thread starter Gregster
- Start date
We can only hope
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
That isn't how HBM works. It's difficult to describe concisely(and not my forte either
).First off HBM needs a large bus to work, it isn't magic by any means. A 290x has a 512bit bus and gets 320GB on 4GB of memory. As HBM currently stands it's stacks of 4 chips making 1GB per stack and giving 128GB/s of bandwidth per stack. The memory controller has to handle the data coming in, in a similar way. Ultimately I'm not sure how it's handled in terms of what they call the bus width on the gpu, but handling what could be 4x HBM stacks with 4GB and 512GB/s of bandwidth, the controller will need to be as big or bigger than it is currently.
Where HBM wins over gddr5(and future versions) is the packaging/connection.
The chip is the silicon bit, the package is the green(usually) bit the gpu sits on, under this you have hundreds/thousands of pins which are connected to various voltage inputs, video outputs and all the memory connections.
The issue with making a bigger memory bus is it means more pins out (and there is quite literally a limit to how many you can fit on a sensibly sized package), how small those pins can be(another thing limiting total number) and the number of traces required to the memory. More traces means longer routes, more layers, more complexity, higher failure rates but most importantly more power.
Where HBM comes in is, the stacked memory part really isn't that important, it's that it's connected on the package level, not the PCB level. Instead of traces that are say 1-2mm wide each and can't get smaller, you make these connections on a piece of silicon, cheaply and at a fraction of the size(ie 40nm on a 40nm process). Effectively a interposer is a piece of silicon where instead of making a full chip, they basically lay out the connections between the gpu and memory stack, then you stick an HBM stack and the gpu down on top of this interposer, then this sits on the package.
In doing this you reduce the length of traces from say anything between 10-50cm to 1-2cm(as traces go up and down layers on the pcb they can loop around for a huge distances before getting to the memory chip), this reduces the power usage by pretty much 60%. It also removes the need for massive numbers of pins off the package, hugely simplified and cheaper PCB's and easier power circuitry because all the power goes to the same place(again reducing traces as each memory chip needs to be connected to a power source, now just the package gets connected).
HBM is effectively taking it's connections and communication down a level from PCB/traces to effectively on die(in this case an interposer).
You can do a 1024bit bus now without it, it would just use so much power, so many traces, such a expensive pcb, chip and ultimately within the same 250W power target, if you use a larger percentage for memory, you have less to use for the rest of the gpu.
HBM lets you save power, increase bandwidth and so have more shaders/tmus/rops.
Soldato
- Joined
- 24 Aug 2006
- Posts
- 10,113
- Location
- Gibraltar
i heard if you buy the new titan you get a free house with it
Im happy paying £20 for a house, if thats the type of free you are on about
http://www.overclockers.co.uk/showproduct.php?prodid=GX-353-SP
290X with 8GB exists.....at the same price as the OC UK Titan Black (cheaper than the other Blacks also) .
That 8GB card is a non Ref card made by a vendor own choice and was limited stock if you read Gibbos post the other day.
Also F.A.O Kaap, traditionally amd/ati use lower speed VRam.
I think Nvidia should have 512bit Bus on next flagship, they have done the R&D as they had it on the 280/285 (R&D cost was excuse before those cards when they were asked about 512bit Bus by a hardware site) and GDDR5 is Quad Pumped so does not need same Bus width as DDR3.
Also F.A.O Kaap, traditionally amd/ati use lower speed VRam.
I think Nvidia should have 512bit Bus on next flagship, they have done the R&D as they had it on the 280/285 (R&D cost was excuse before those cards when they were asked about 512bit Bus by a hardware site) and GDDR5 is Quad Pumped so does not need same Bus width as DDR3.
Last edited:
Soldato
- Joined
- 22 Jul 2012
- Posts
- 16,145
- Location
- London
Yeah, but you can't get blocks for it and therefore can't fit 4 cards.
Soldato
- Joined
- 5 Sep 2011
- Posts
- 12,883
- Location
- Surrey
I think EK are actually in the process of making them. However I'm unsure if this fits the 8GB model. I'd imagine it would...
http://www.ekwb.com/news/500/19/New-water-cooling-gear-in-the-works-July-2014/
That said, bit late now arguably. On air though they're a complete waste of time in Crossfire. The three slot design makes them cumbersome. Titans definitely still the best choice for air users!
Bit annoyed actually as I contacted EK directly and they said they had no plans to make any. Obviously they saw the demand not long after.
Thanks EK.
http://www.ekwb.com/news/500/19/New-water-cooling-gear-in-the-works-July-2014/
That said, bit late now arguably. On air though they're a complete waste of time in Crossfire. The three slot design makes them cumbersome. Titans definitely still the best choice for air users!
Bit annoyed actually as I contacted EK directly and they said they had no plans to make any. Obviously they saw the demand not long after.
Thanks EK.
Soldato
- Joined
- 22 Jul 2012
- Posts
- 16,145
- Location
- London
I could be mistaken, but I can't see anything there regarding the 8GB vapor X and I'd be very surprised if they did make one considering how limited the run was and that it was only ever available from OcUK.
Soldato
- Joined
- 22 Jul 2012
- Posts
- 16,145
- Location
- London
Nice! I stand corrected 
Soldato
- Joined
- 5 Sep 2011
- Posts
- 12,883
- Location
- Surrey
The back plates on the 8GB get seriously hot. Practically too hot to touch in fact.
Well as the Vapor-X range had no blocks up until even now - I pointed this out to AMD at the time (one was generously given to me) as it was a big sore point. I was told "we'll see what we can do" but I didn't expect anything to come of it
Well as the Vapor-X range had no blocks up until even now - I pointed this out to AMD at the time (one was generously given to me) as it was a big sore point. I was told "we'll see what we can do" but I didn't expect anything to come of it
Last edited:
I think the R9 290 brought them plenty to think about and I am sure AMD will do it again...
It did but not a great deal, in nvidias eyes it took amd over 6months to even match the original titan, even then a 780ti and titan black later and they nudged back ahead...If amd want to give them something to think about they need to be the ones setting the bar
Soldato
- Joined
- 28 May 2007
- Posts
- 10,222
It did but not a great deal, in nvidias eyes it took amd over 6months to even match the original titan, even then a 780ti and titan black later and they nudged back ahead...If amd want to give them something to think about they need to be the ones setting the bar
Amd matched the Titan at half the price though. I agree mostly with what you are saying and Amd need to come out of the blocks with a complete package. The 7970 should have came out with Ghz clocks and the wonder drivers to go with it. This would have given Nvidia something to really think about. The 290x should have came out with a much better cooler. I think Amd are on the right track with Gcn architecture though as they can compete on performance again.
+1 Totally agree