Doesn't the off-die bandwidth required go down as the quantity of on-die cache is increased? If we're talking about embedded platforms which have to feed just a single 1920 by 1080 sized screen then if they increased the cache on the GPU and increased the cache size available to both the GPU and CPU they could avert a decent number of bus calls and therefore save bandwidth.
Fusion die-shot - 2009 Analyst Day
- Thread starter AnarchX
- Start date
Blazkowicz
Legend
maybe that would work, but with a really huge quantity of cache - render targets are big.
I don't believe in MCM, huge edram, sideports etc. as the problem is more simply solved with a dedicated GPU.
I don't believe in MCM, huge edram, sideports etc. as the problem is more simply solved with a dedicated GPU.
The 5670 is a $100 part. I don't think anyone here expects Llano to replace a $100 graphics card, nor should OEMs, consumers, or AMD for that matter.
ITs neither here nor there because the 5670 is avalible. When is Liano coming anddo you think the 5670 will be $100 then ?
http://www.newegg.com/Product/Produ...1334&cm_re=radeon_5670-_-14-131-334-_-Product
They are already going for $85 bucks after rebates.
Regardless of what stage of the lifecycle 5670 is in, it will always have a higher BOM than the integrated GPU in Llano, and thus it's not a fair comparison. Llano will replace parts like the 5450.
I was responding to specific posts coming before me. Mainly 3dilettante but also rpg's post above mine that asks if the sub $100 parts will go away and if they do how will they continue r&d on the high end.
Ther will allways be a market. I think even $50 cards can be safe and i said why i believe so in my post
Silent_Buddha
Legend
I'm still not convinced x4xx class cards will be in danger from integrated. x3xx cards might have some trouble as they don't have quite as convincing a lead over integrated.
Regards,
SB
You doubt Llano can outperform the 5450 with its 80 SPs and 64-bit bus?
Silent_Buddha
Legend
Maybe, but yes, I doubt it. More importantly will it outperform whatever x4xx card is released around whenever Llano is released?
I'm just not convinced shared bandwidth to main memory is going to lead to it being faster than any contemporary x4xx card.
Note, this isn't to say it's impossible. I'm just not convinced it'll happen.
Regards,
SB
While there's a point comparing tomorrow's SoC performance with today's lower end GPUs (in order to see how performance could evolve), another factor we should consider is what lowest end standalone GPUs will look like in 2011.
Calling 240SPs (or replace that number with anything else if 240 isn't accurate) ALUs sounds wild on paper but there's another reality that might tell you that those are actually 48 Vec5 ALUs, which automatically puts a different perspective on the matter. In theory if any manufacturer would be interested (which I severely doubt) to maximize the IMG SGX543 MP with 16 cores your end result is at 64 ALUs and no I'm not comparing directly here anything either but assuming Llano reaches either way the TFLOP mark in 2011 and considering that the 5970 isn't too far apart from hitting nearly 5 TFLOPs I wonder where the enthusiast SKUs in 2011 max out at.
If you double the theoretical peak of a 5970 for 2011 having 1 TFLOP or damn close to that for the absolute lowest end PC solution doesn't sound as impressive anymore does it?
The ALU's in IMG SGX543 MP are scalar or vec4?
Llano needs to run it's ALU's/SPs at 2.1GHz to hit 1TFlop mark.
a) Can the GPU core run >@2GHz?
b) Can Llano keep it well fed >@2GHz?
{The cpu can contribute ~100GF max @3GHz.}
Historically, the difference in raw flops between the highest end and the lowest end has been somewhat more than OoM.
That's besides the point. More below.
Llano needs to run it's ALU's/SPs at 2.1GHz to hit 1TFlop mark.
a) Can the GPU core run >@2GHz?
b) Can Llano keep it well fed >@2GHz?
{The cpu can contribute ~100GF max @3GHz.}
Historically, the difference in raw flops between the highest end and the lowest end has been somewhat more than OoM.
That's still not the point. Mark Llano with any performance figure you please (be it X GFLOPs) or whatever else and wonder what the result you'll get is exactly how comparable to 2011 standalone GPUs. I don't care here if it contains 240SPs or more or what they're clocked at in a strict sense.
In 2011 more than just one GPU aspects will most likely double and that's the real measure you can compare a SoC with Llano inside and not what floats around today.
Blazkowicz
Legend
don't be so harsh
.the 5450 is yet another slow card on par with 8400GS and radeon 2400 pro (because it's still a 64bit ddr2 card). same bandwith stagnation as on IGP.
take the current radeon 5570 as the ceiling of what a Llano could do. a future card such a a 6450 would be the floor. (64bit bus, DDR3 and a low amount of SP / low wattage)
don't be so harsh
the 5450 is yet another slow card on par with 8400GS and radeon 2400 pro (because it's still a 64bit ddr2 card). same bandwith stagnation as on IGP.
take the current radeon 5570 as the ceiling of what a Llano could do. a future card such a a 6450 would be the floor. (64bit bus, DDR3 and a low amount of SP / low wattage)
Agreed. Llano apparently has a 128-bit bus with DDR3. Granted, it's shared between CPU and GPU, but the CPU part doesn't need much bandwidth. Just take a look at benchmarks comparing Phenoms with DDR2 vs DDR3.
Didn't know where else to put it, so here it is.
Opinion: AMD, Intel, And Nvidia In The Next Ten Years
Opinion: AMD, Intel, And Nvidia In The Next Ten Years
Silent_Buddha
Legend
don't be so harsh
the 5450 is yet another slow card on par with 8400GS and radeon 2400 pro (because it's still a 64bit ddr2 card). same bandwith stagnation as on IGP.
take the current radeon 5570 as the ceiling of what a Llano could do. a future card such a a 6450 would be the floor. (64bit bus, DDR3 and a low amount of SP / low wattage)
But here you're assuming that there will be no bandwidth increases to whatever x4xx card is out at that time. Nor, whether there will be any memory speed improvements in the meantime.
Compounding that is the assumption that Llano will be 240 SPU's when no definitive answer to that has been made yet. Adding to that, we're expecting an order of magnitude jump in integrated graphics while discrete graphics stay stagnant? Current integrated is still quite a bit slower than the x4xx cards.
Color me highly skeptical that Llano will be faster (not even going to touch significantly faster) than whatever x4xx card is released around the same time (which would most likely be a refresh of NI).
Again, not impossible I suppose, but highly unlikely, IMO.
Regards,
SB
The author's opinion of Itanium is laughable. Gotta love those slides from 1998.
Not sure what you mean by this, CUDA is popular because of its superior tools and support, something OpenCL has yet to match.