uhn... just look at the tdp of recent midrange gpus (aka 4850) -> 130W now combine it with a dualcore and your beyond 180W -> have fun with overheating.Quote:
Originally Posted by Eson
Printable View
uhn... just look at the tdp of recent midrange gpus (aka 4850) -> 130W now combine it with a dualcore and your beyond 180W -> have fun with overheating.Quote:
Originally Posted by Eson
dude.. dont flame and troll.. im not drinking and posting nor do i do drugs.. and even if i do its none of ur business.. im discussing this decently.. i suggest you keep within your limits.. and before you call this 'random' invention.. think a little.. this was the primary reason amd bought ATI.. now they dint randomly buy ati.. besides.. a gpgpu concept has been there for a long loong time... i asked you to answer my quesitons in the last post.. take each point and shoot it down with technical answers and ill totally agree with you... dont flame me.. :shocked:Quote:
Originally Posted by Shintai
precisely... :up:Quote:
Originally Posted by Eson
Besides the obvious thermal issue as another person kindly described. There is also a few small bandwidth issues with it. If you wanted say a 4850 would most likely need something along with 12+64GB/sec...using DDR3 dimms? 6 Channel IMC?Quote:
Originally Posted by Eson
And what would the size of such a CPU be if it was to integrate say a 4850?
So ye...im pretty confident. Its not like Fusion or value Nehalem will revolutionize anything. Its just IGP moved ondie.
IGP moved ondie with a killer feature. Direct access to CPU and fast tracks to RAM (though not totally fast but DDR3 can resolve that)Quote:
Originally Posted by Shintai
Not to mention that current IGPs are not really dense yet, if we see a HD4450 40nm derrivative into Fusion 2nd/3rd gen it should be better (ATI's ALUs are small for the performance delivered)
/me hands out cookie.Quote:
Originally Posted by Macadamia
I fully agree. And thats a realistic option. Not some 4600 series or even 4850 class GPU and performance.
CPU access doesnt really matter. 50ns or 500ns..no difference. But the memory part matters.
Are you kidding me??? That's awesome. Competition is good. Look at what it's doing to prices!Quote:
Originally Posted by hecktic
Are your assumptions based on todays highend GPUs or tomorrows? When Fusion hits the marketing department i think there are more powerful GPUs then the HD4870. So please specify what kind of high end GPU we wont be seeing in Fusion because HD4870 will most likely be of IGP class by then. So you are both wrong and right.Quote:
Originally Posted by Shintai
Well, you have to consider another part that when the GPGPU memory interface is substituted by the IMC, you have some extra space to manuever.Quote:
Originally Posted by Shintai
240 seems to be the sweet spot for the bandwidth you can grab and decent performance. Decent enough to make the lowend nVidia cards gone, and lowend AMD cards OEM/upgrade only.
4850- 800SP, 40TU, 256-bit MI 256mm^2 (351600 pixels in Photoshop)
http://www.rage3d.com/reviews/video/...e/pics/die.jpg
Fusion MAX- 400SP, 20TU, NO MI, should be smaller since I didn't remove extra register space or display controllers. 205495 pixels in Photoshop, ~150mm^2. Predicted size @ 40nm = 108mm^2
http://img67.imageshack.us/img67/3824/51684309ic2.jpg
Fusion PROBABLE- 200SP, 10TU, NO MI, should be QUITE smaller since register space/controllers is not removed as much as it should. 168337 pixels in PS, 122mm^2, Predicted @ 40nm = 88mm^2.
http://img77.imageshack.us/img77/7042/14857385dq1.jpg
So yes, it can definitely reach current midrange (3650/ 120SP) levels, probably in the future exceeding it. With careful planning 3800 levels can be achieved. But as in units, not clocks- these will be clocked low for TDP. :rofl:
It wont change much till Fusion or Nehalems with GPUs to come. A 4870 will certainly not be IGP class, unless Fusion is to hit in 2013+. They are both somewhat in the range of 1 year away. Selecting a 4450 is even way above IGP class today.Quote:
Originally Posted by Eson
If you think it will be 4870 class, specially without the 116GB/sec memory interface...feel free. But to me, a double performance of X3500 and 790G would be a miracle by itself.
You forgot the CPU part, plus it would be 45nm most likely. Same process. Same die. Then add it needs to be an economic viable size. As in 150mm2 or less for CPU+GPU. Next is your memory interface in best case would perhaps be 16-18GB/sec that needs to be shared with the CPU. If you noticed on 780G/790G etc. The memory part is a huge performance penalty and always been the bottleneck of IGPs.Quote:
Originally Posted by Macadamia
AMDs next IGP IS 3650 based, so of course it's dooable..
As for memory bandwidth, there's always sideport memory which is being used on 780G currently, which also allows a combination of sideport and shared systyem ram (to make a good comprimise)
In that case, the max is 200SP and a Brisbane, but 120SP and a Brisbane/Griffin seems more probable in the end.Quote:
Originally Posted by Shintai
As for bandwidth, no sweat. The 780G has an optional sideport where the IGP can access discrete GDDR3 chips for fast bandwidth. Fusion might need extra pins for that, but 128/256 would be far enough.
As far as OEMs are concerned, the discrete RAM is cheap and there's literally no point in not having it IF Fusion is rolled out as a platform.
Yeah, by 2009, 2 Gbit (256 MB) GDDR3 1 ns chip should be pretty affordable as the sideport VRAM for Fusion class GPU, 1280*1024 medium to low IQ setting gaming should be attainable by then using an onboard GPU+CPU chip package.
Fusion sounds great to me, and may help revive pc gaming, so many ppl I know have decent pc's but can't play any games bc of their crappy IGPs..
hmmmm.... i guess some more info is here..
“In the future, AMD envisions APUs that include a varying mix of CPU cores (scalar processing cores), GPU cores (parallel processing cores), and fixed-function accelerator cores—all on-chip.” Moorhead adds that AMD won’t leave out off-chip accelerators, preserving the right hardware combinations for every level of the market. To take advantage of these new architectures, AMD is working toward simplifying the programming environment.
link: http://www.computerpoweruser.com/edi...07%2F30c07.asp
and AMD guy himself is stating that the CPU GPU integration wont make discrete cards obsolete.. so i guess my assumptions were wrong... :)
you sure its on the die?Quote:
Originally Posted by Shintai
last thing i heard was, that intel is gona use mcm for value nehalem, and maybe amd goes the same way after the problems with nehalem.
They cant really make it MCM for Nehalem. Since the lowend (And mainstream) parts dont have QPI, FSB etc. It would have to sit as a PCIe card. Not as easy to share memory that way. And the GPU part is so tiny small it makes no sense not to integrate it. Why would you leave such a small part outside. Specially when you want to make it on the same process.Quote:
Originally Posted by Hornet331
there where several scources:Quote:
Originally Posted by Shintai
http://www.beyond3d.com/content/news/540
http://www.xbitlabs.com/news/cpu/dis...Necessity.html
thought this article is quite old, i dont recall that intel denied the mcm rumor.
Compare this qoute from the article with what you know.Quote:
Originally Posted by Hornet331
They say GPU and IMC is outside the CPU itself. They say socket 715 in the Beyond3D etc.Quote:
Intel’s Havendale processor is multi-chip module (MCM) in LGA1160 form-factor containing Nehalem micro-architecture-based dual-core CPU as well as graphics and memory controller hub (GMCH) that features dual-channel DDR3 memory controller, PCI Express 2.0 x16 interface to connect add-on graphics cards as well as integrated graphics core. It is projected that both chips on the MCM are made using 45nm process technology.
Plus you need extra interconnect for it to happen. QPI? FSB? Something else? Thats increased diesizes aswell and so on. And for the IMC part, QPI and FSB is too slow actually.
Its ofcourse possible. But I just dont see any reason why.