As everyone knows, I was VERY skeptical about this merger because AMD has such a close relationship with both companies.
 
But after I saw a diagram they posted on Anand I could understand the prupose of the merger.
 
I also thought about how nVidia and ATi are actually in different AMD market segments in terms of the chipset. nVidia enjoys wksta/small server chipsets along with high end SLI.
ATi is in the desktop gaming market but their chipsets are only getting good enough OVERALL to compete.
 
AT this point it is very important that AMD get a glueless chipset ready for K8L without relying on a "separate" company. They will need to scale to the 1000s and they would be better off doing the work in-house.
 
The diagram reminded me of the modular architecture AMD promoted at Analyst Day. By using cHT between the GPU and CPU it would be possible to have the pixel/vertex on the GPU with a high speed cache while the cores on the CPU handle the normal duties plus the feeding of pixel data to the pipes.
 
Of course they could still then put the full blown monster in a socket next to the CPU for enthusiasts and pro content people. The 4 socket mobo coul dthen have 2 GPUs and 2 K8Ls.
 
This is somewhat based on conjecture but a lot on the various news AMD has let out about their plans. Teh good thing is that even though they haven't merged yet, ATi can still work on AMD chipsets. The engineers may still even be able to share info.
 
At any rate with nVidia making the SLI chipsets and ATi settling into specialty GPU parts and high end video, there will be no difference in the amount of either sold today.
 
This is a bad thing for the Dell/Intel though because SLI only works on nVidia chipsets, while both the G80 and R600 IGPs will be Premium ready, while Intel is still working on GMAX3000.
 
Another thing this may improve is video on VMs. If they can partition the GPU the way they have the CPU it will be possible to have entire dev environments on one K8L 8 socket machine with 16 VMs and 4GPU modules. People can then logon and use a 32 bit video card instead of the emulated S3 there is now. nVidia could then license this under Torrenza and perhaps even improve it with their GPU knowledge.
 
Also I can see them making Sun chipsets since they will be sharing a socket. The possibilities ar eexciting to say the least.  
 
CPU/GPU diagram
 
I'm sure MS would love that.

I'll say it again AMD and ATI are going to work together to turn the ATI's R600 technology into a general purpose processing monster by the 2008-2011 timeframe to compete with the same type of technology which Intel and Sun are developing. Massively parrallel with relatively simple cores but massive processing power. CPU's like this will cure cancer.

 
 
No, this is about glueless chipsets in house for K8L. They could work with ATi if it was about GPUs.

Sry BM I meant to add to the conversation, not disagree with you.

Im excited to see what they have instore. quad core: 2 cpu + 2 gpus in 1 chip*dreams*

Thats all fine and dandy, but why the hell would you want your R600 using system ram across the mobo? Its gonna take something other than that to feed texture info into a 48 pipe pixel crunching monster.  
 
I know! They should keep the CPU in a flip chip design, then put the GPU on its OWN board with its OWN ram thats actually fast enough for such a thing. OOOh OOOh. . . they could also make a special slot thats uber high speed incase the GPU has to access the system ram! Genius i tell you! Brillant!
 
 
Oh wait, I'm not the first to think of that, am i?  

Exactly.

We're a long way from that.  

To increase latency, duh!  

 
You have an opinion, don't you? I don't expect you to agree with or be on the same page as me. I enjoy hearing opinions.
No offense meant or taken.

 
They will have to gut the GPU to do that. It will take a year at least. The idea behind integration is a balance of power between the execution units. With DirectConnect, it is possible to share the same parallel OR serial bus so by moving execution units to the CPU it would then be possible to make pixel chips that only process DX10 in HW.
 
But then I have been known to be faster than the times.
 
 
MS did already make the desktop 3D so now X64 desktops NEED more power. HD will need more throughput, games will need more throughput, etc. AMD is following a very ingenious track towards a unified CPU platform based on open standards.
 
OEMs won't be able to resist.

I'm always in tears reading your posts. This is one of your funniest yet.

 
That's where DC and DDR3 come in. With a low latency high bandwdith connection the only question is processing power. GPU tech pumps 50+ GB/s across it's internal bus, while games do not need that much, they need low latency number crunching.
 
That's why CPUs run at 3GHz and GPUs run at 600MHz. GPUs are throughput bound while CPUs are IPC bound. HT3 will have 25.6GB/s PER 16bit link. Connecting 3 in parallel will give you 76.8GHz of bandwidth.
 
Dividing CPU functionality and GPU functionality (in the case of games) will mean that careful caching will give you 50+GB/s of pixel bandwith and 25+GB/s of CPU/IO bandwidth.
 
ATi has experience in memory controllers for GDDR4 so CPUs could actually use DDR3 AND DDR4 through two different ports. HT would allow GPU RAM AND CPU RAM.
 
 
Imagine glueless functionality for some new special FX movie in two to three years. I bet nVidia will shine with a socketed GPU.
 
I'm actually impressed with what can be done.

Not this nonsense again.

That's not really true.
 
http://techreport.com/onearticle.x/10403
 

Latency isn't a problem for GPUs, it's always been raw bandwidth. In that case, an HTX implementation just isn't beneficial. You can say that you can link up 3 HT3 links which is highly unlikely. That would require all motherboard makers to have triple traces between the CPU socket and the GPU socket which is costly and unlikely due to space constraints. Besides, even with triple HT3 links a CPU's IMC can't provide enough bandwidth to fill all three links anyways unless you have a nice 6 channel FB-DIMM implementation, which isn't coming to desktop anytime soon. Again bandwidth is key.
 
What I can agree on though, is that HTX GPUs would be a great replacement for IGPs. They would probably be slightly more expensive, but noticeably faster. HTX is viable for low-end graphics cards because those don't need as much bandwidth.

To continue the lesson, GPU's only run @ 600MHz not because "They dont need to go faster" but because it has to keep parallel coherency across 48 cores on the die. That is no easy task.  
 
Please stop making up this crap and spouting it off. People come here to learn, and make informed decisions on thier future hardware purchases. Not only are you confusing them with facts that run perpendicular to reality, but you are also sulling up the good name of Tom's.

A hybrid GPU/CPU chip might work as a replacement for todays integrated graphics chips, but it wont kill stand-alone graphics cards.
 
HTX would be a viable replacement for PCI-e. Lower latency, direct low level access, and vastly superior bandwidth compared to PCI-e.
 
Sidenote: HTX is a slot, and it uses the same mechanical connector as PCI-e 16x.

At first glance i saw "AMDs plan?.....Clueless Chipsets and modular pixel pipes??"
 
This (A hybrid GPU/CPU chip?) may be good for super low budget mid performace parts...

I'm beginning to believe that all this attention towards the AMD/ATI merger has all the Intelebies and Conroe fanboys tights in a tizzy because it's taking the fizz out of Conroe's release hype.
 
Think about it, an existing company (Intel) releasing a new product (Core2) compared to a merger of two technology leaders (AMD/ATI). Who's the media gonna pay attention to? Who's stocks are gonna catch the interest of Wall Street?

Let's not forget about the AMD/Rambus licensing deal earlier this year as well. That gives AMD and ATI products access to XDR2 and FlexIO (think of your FSB running at 8 GHz) among other things. I don't know what they have planned, but the next 5 years should be very exciting.

Interresting.
 
XDR2 would provide enough bandwidth to feed both the CPU and the GPU.  
 
Perhaps having a GPU in a socket is not a bad idea after all.