- Jul 1, 2005
- 5,529
- 0
- 0
That's silly. My 8800GTS is around the same size as an MATX motherboard.
There is no way they will solder the memory to the motherboard. That's only done due to space constraints in devices like consoles. They do it on graphics cards because there is usually no need for the user to upgrade the memory (and it's not like they sell GDDR4 in stores).
I'm certain that the memory bus width is very 'low tech' at this point and will be one of the simplest issues to overcome. The only 'cost issue' that I forsee is having to buy GDDR4 for your entire rig, not just the graphics card. That should be offset by the fact that you're sharing the memory and not having an add-in board. IMO high end systems will be $100 to $200 *cheaper* than current high end setups because of this type of implementation.
You still have not refuted my comment re: the Xbox 360. That's an integrated GPU, and it was faster than any graphics card on the market....in 2006.
apoppin
Lifer
Originally posted by: SexyK
Originally posted by: apoppin
isn't it all really speculation - at this point?
there is really very little hard info available ... i expect everything to be much clearer by the end of this year
Agreed apoppin - it is all speculation, but I think people predicting that Fusion-based GPUs will outperform discreet add-in boards, or even the demise of discreet add-in GPUs entirely are off-base. Neither of those scenarios will be happening any time soon.
define 'soon' so we have a common base
i am saying 5 years will start to eliminate discreet GPUs as the integrated solutions surpass the current PCIe2 technology
That's fair, but I think it's more likely that nVidia will either die or be absorbed by a CPU company than it is to think that AMD will go out of business within the next 5 years.Originally posted by: SexyK
Originally posted by: apoppin
isn't it all really speculation - at this point?
there is really very little hard info available ... i expect everything to be much clearer by the end of this year
Agreed apoppin - it is all speculation, but I think people predicting that Fusion-based GPUs will outperform discreet add-in boards, or even the demise of discreet add-in GPUs entirely are off-base. Neither of those scenarios will be happening any time soon.
My point is the argument "i don?t see any CPU with high bandwidth to memory subsystem NOW" is flowed because CPU don?t NEED it.
If CPU NEEDED high bandwidth, you would have seen high bandwidth memory subsystem.
Saying that it doesn't exist NOW is a moot point. It isn?t needed NOW.
When it would be needed, then you will see it. It is that simple.
Because integrated GPU NEED it you will see it happen.
It is technically possible!
You don?t see that in desktop system with a 1K$ price tag on them because it isn?t needed there.
If CPU NEEDED high bandwidth, you would have seen high bandwidth memory subsystem.
Saying that it doesn't exist NOW is a moot point. It isn?t needed NOW.
When it would be needed, then you will see it. It is that simple.
Because integrated GPU NEED it you will see it happen.
It was, like I said, just an example.
It is technically possible!
You don?t see that in desktop system with a 1K$ price tag on them because it isn?t needed there.
As for the xbox 360, the GPU has access to ~25GB/s of memory bandwidth, however the memory is not in DIMM form - it is attached to the motherboard making the xbox360 graphics solution more like an add-in card than an integrated solution.
I also am not disagreeing that in 5 years we may have a fusion-type product that can compete on the high end. However many people around here seem to think that the first-gen fusion products will make discreet add-in cards obsolete overnight, which will not be the case.
Cache and prefetch are example of creative solution that engineers come up with when they reach hard technological limitations (latency).
Well, I was just reading that AMD is working on 4 versions of Fusion:
- data centric
- graphics centric
- media centric
- general use
What I take from this is that the 'graphics centric' version will probably be a pretty good performer.
They're saying that AMD's main ambition for making Fusion was to create laptops with longer battery life and much better graphics performance, so you may well be correct to a degree.
I'm thinking Fusion will be better than most midrange graphics cards when it is released. That's not to say that it won't be possible to beat out the high-end; it's just that very few people require that level of graphics horsepower.
apoppin
Lifer
true .. but then most of your threads ...
[most of all threads]
... don't get any attention at all
if we stayed on topic it would have been 1 or 2 pages ... max
you DO have a gift
That's interesting information on fusion, do you have a link? Anyway, you may very well be right about what performance to expect. Considering the amount of information available now, it's really impossible to know. I am just thinking out loud about the hurdles i foresee AMD/ATI having to overcome in order to make fusion work. Trust me, I would love to see them succeed. Although I still think initial system cost would probably be higher with fusion, the cost of upgrading a socket-based GPU would presumably be lower than purchasing a whole new add-in card.
I think he is referring to this one:
http://www.amd.com/us-en/assets/content...406_AMD_ATI_Strategic_Vision_FINAL.ppt
http://www.amd.com/us-en/assets/content...406_AMD_ATI_Strategic_Vision_FINAL.ppt
LoL!!
Sorry, but the 360's Xenos GPU was not faster than any card on the market in 2006. Hell, the G80 launched in 2006.
Even before the G80, the X19k series and G7 series could play some of the same games with higher resolution higher IQ and still get playable frames.
I still have yet to see a single Xbox 360 game that actually uses decent levels of AA and AF. (Granted, there could be one out now. I haven't played the 360 for a few months)
Part of the reason why the probably don't use higher IQ in their games is because of the bandwidth constraints.
- Jul 1, 2005
- 5,529
- 0
- 0
The 360 does not even come close to matching the image quality or resolution of a separate video card. Not too mention that all it has to do is run games customized just for the GPU and a minimal operating system.
I don't think a CGPU or a integrated GPU will ever replace a separate graphics card. Why else would the R600 be so huge if they were moving in a direction of making em smaller.
The only benefit of a combined CPU\GPU is cost.
Yeah, TODAY there are better graphics cards than the Xbox 360`s GPU. When it was released, there were not.
The X1900 cards weren`t out yet, and I`m not certain if even the X1800 cards were.
Xenos is no where near as capable as you people make it out to be.
You guys are fooled into thinking it performs better than it does because it uses it's own custom API that is tailor made for it without any OS overhead.
Look at the freakin RSX GPU in PS3. That's nothing but a 7800GTX and as far as I'm concerned, PS3's graphics easily match the 360's (I'm a 360 user and I love it).
As for the whole C/GPU issue, I'm not going to get sucked into a flame war at the moment. I have a couple more classes today and I dont want to start something I cant finish. Besides, this thread is already off topic enough.
kobymu- It's obvious that I'm not going to change your mind and you wont change mine. I'm sticking to my opinion that C/GPUs wont overtake discreet graphics till 2011. Your opinion is just as valid as mine because they're both based off of pure speculation.
On a side note, if AMD is going to shock the world by giving us a C/GPU with greater than R600 performance next year, how in the hell is Intel planning on competing?
You guys are fooled into thinking it performs better than it does because it uses it's own custom API that is tailor made for it without any OS overhead.
Look at the freakin RSX GPU in PS3. That's nothing but a 7800GTX and as far as I'm concerned, PS3's graphics easily match the 360's (I'm a 360 user and I love it).
As for the whole C/GPU issue, I'm not going to get sucked into a flame war at the moment. I have a couple more classes today and I dont want to start something I cant finish. Besides, this thread is already off topic enough.
kobymu- It's obvious that I'm not going to change your mind and you wont change mine. I'm sticking to my opinion that C/GPUs wont overtake discreet graphics till 2011. Your opinion is just as valid as mine because they're both based off of pure speculation.
On a side note, if AMD is going to shock the world by giving us a C/GPU with greater than R600 performance next year, how in the hell is Intel planning on competing?
apoppin
Lifer
intel has their own solution
even the OP is OFF topic
:roll:
what's that about "derail" again, Wreckage?
even the OP is OFF topic
:roll:
what's that about "derail" again, Wreckage?
Originally posted by: apoppin
intel has their own solution
even the OP is OFF topic
:roll:
what's that about "derail" again, Wreckage?
I know Intel has there own solution, can't think of it's name right now but they're working on it.
However, who here thinks that Intel can come up with anything other than a low-mid range solution in their C/GPU?
If Intel really thought AMD was going to give us a C/GPU that was equal to a high end discreet graphics card as some of you do, they would be crapping their pants right now and might have made a push to buy Nvidia.
zephyrprime
Diamond Member
- Feb 18, 2001
- 7,512
- 2
- 81
Having a higher bit width would require a mother board with more layers. That would add cost. For comparison, a 8800gtx has 12 pcb layers and a motherboard has 4-8 depending on how fancy and how compact the board is.
CPU NEEDS high bandwidth! It's just not realistic (meaning much more expensive) to get it. That's why there are tricks around that (cache, prefetch, ...). The ultimate ideal state would be the whole RAM in a form of cache.
Some tasks don't need more than 1MB of memory. They run from cache and they are fast but there are many tasks that need to go main memory and they would benefit from high bandwidth.
GPUs need high bandwidth because by nature of a task at hand they need more than 1MB(2, 4, 8,... what ever cache size would be possible) of memory.
CPUs are a little bit different. They execute different type of tasks and many of them fit into cache and that's the reason why they appear that they don't need high bandwidth. But if there was no cache they would starve to death for high bandwidth.
From anand latest article:
http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=2963&p=6
All the core2duo variety had the SAME bandwidth.
All the Athlon 64 X@ variety had the SAME bandwidth.
And that is from the 3D RENDERING Performance page.
Even when you look at 2 core CPU only you see a delta of 50% if not more.
What does that tell you?
Did you ever program any real world application? Do you have any idea what the hell you are talking about?
For every application you can find that is bottlenecked by bandwidth I can find you 10 that are bottlenecked by other subsystems, 10!
apoppin
Lifer
http://www.theinquirer.net/default.aspx?article=38802
and
8800 Ultra to arrive on May Day
for $999
:shocked:
and
8800 Ultra to arrive on May Day
for $999
:shocked:
The point he is making is that without on-die cache every application would be bandwidth limited. Thus the fact that most PC applications work on small data sets paired with powerful prefetchers and the presence of a extremely high-speed cache on-die lessens the impact of lower bandwidth to system memory. Try turning off your L1 and L2 cache and see if applications are limited by anything other than memory bandwidth. The vast majority will struggle mightily. That is why janoo is arguing that CPUs do need massive bandwidth - they need it, but only for a smaller data set which can be predicted and cached in L1 and L2 most of the time.
This is in contrast to GPUs which work with much larger data sets. 1-2MB of high-speed cache on a GPU would be insufficient to hold all the data required to render even one frame at a decent resolution. One approach to mitigating this issue with integrated GPUs is found in the xbox 360 where there is 10MB of high-speed eDRAM integrated into the GPU die. Note however that the 360 only renders in one resolution all the time, and all 360's have the same amount of memory, so developers can target one set of specifications and tailor their applications to fit into the eDRAM. This approach would have a much harder time working on the PC platform because people expect to be able to use extreme resolutions with their high-end GPU, so the size of the eDRAM would have to increase significantly. Creating a much large eDRAM block would make the die huge and is probably cost-prohibitive. Note also that the even with the inclusion of the eDRAM, the 360 still uses comparatively high-speed system memory (4x the bandwidth of current PC system memory) that is soldered directly to the system board to complement the cache. As others have pointed out, getting this kind of bandwidth onto a consumer-level motherboard with traditional DIMMs would increase the complexity of motherboards many times over, and would most likely increase the cost of DIMMs significantly because the DIMM interface would most likely have to be increased from 64-bits/channel to 128 or 256 bits per channel. Even then the trace length would still have to be addressed in order to allow the memory clockspeeds necessary. We are a long way off from having a fusion GPU being the top dog. Midrage? Possible. But high-end is still a ways off.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads
- Started by Tigerick
- Replies: 21K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
-
Facebook
Twitter