New age AGP... i mean PCI Express

[BassBomb]

Let me start off by saying, we all know AMD successfully reduced the FSB bottleneck that used to occur on standard CPU's back in the day,

With their HTT, or High speed connection from the CPU(on die memory controller) straight to Memory (Single/Dual Channel DDR) via Hyper Transport

My question for the techy's is why not create a smiliar sort of option for HTT to graphics card?
Wouldn't this help use the full potential of graphics cards and the interface 16x (since the performance gains are still not HUGE compared to AGP 4x -> AGP 8x, and AGP 8x -> PCI-E 16x)
Maybe I just dont know what im talking about.

Would a PCI-E interface have to be built on die, and thus could cause problems with motherboards which use Dual PCI-E such as SLi and CrossFire?

It would be nice to see the potential of 16x exploited greatly and on die PCI-E lanes ..

Have comments post, I'm very interested in the possibilities or opinions of such an idea!

 

[DonaldDarko]

I cant seem to start a new thread, so i'll ask here:
this is pci right?:[------------ ----]
and so is this?:[---- ------- -----]
i always thought the top one was agp? is one digital, and one analog? i'm a little confused..
i know[---------- ----][ ] is agp (4x/8x)

are you looking for a direct bus from CPU to graphics card? i think you'd need one helluva fan(or water) to cool that kind of line. so far the biggest bus iv'e seen is 400mhz. to have one from CPU to graphics, it would have to be big, like 1G or bigger to prevent not only lag, but heat to. HTT to DDR AND to graphics would be a an overload. (maybe 2 separate processors? lol, each having dual core?) i dunno, thats my guess..

 

[BassBomb]

check out tomshardware latest article on pc interfaces.. it deals with that

the differnece between those are the older and newer pci slots dealing with differnet voltages.. so you wont be able to stick a 5V card into a 3.3V slot and risk damage

 

[Peter]

First get your facts right. The RAM doesn't connect through HyperTransport. The RAM controller is entirely direct-on-chip, on the CPU internal crossbar.

The HyperTransport link(s) go out to the chipset components (and more CPUs). And guess what a "chipset northbridge" does in such a system? Tunnel HTT through to the southbridge, and fork PCIE off the HTT uplink to the CPU.

Performance gains would come from further parallelization, like putting PCIE root ports directly onto the CPU die.

 

[soccerballtux]

they're going to be doing this in the future, they said they are recently.

 

[icarus4586]

I think your understanding is a little off.

Hypertransport is for 1 CPU to another, or for the CPU to the Northbridge. The memory bus on an K8 system is not Hypertransport. The memory controller is built into the CPU, and this helps to reduce latency, but that's it.

As far as having an AGP or PCI-Express controller on-die, like the memory controller, it is a possibility. They're looking into it.

 

[coomar]

Originally posted by: soccerballtux
they're going to be doing this in the future, they said they are recently.

but it would be 4 years from now

 

[Matthias99]

Originally posted by: BassBomb
Let me start off by saying, we all know AMD successfully reduced the FSB bottleneck that used to occur on standard CPU's back in the day,

With their HTT, or High speed connection from the CPU(on die memory controller) straight to Memory (Single/Dual Channel DDR) via Hyper Transport

My question for the techy's is why not create a smiliar sort of option for HTT to graphics card?
Wouldn't this help use the full potential of graphics cards and the interface 16x (since the performance gains are still not HUGE compared to AGP 4x -> AGP 8x, and AGP 8x -> PCI-E 16x)
Maybe I just dont know what im talking about.

Would a PCI-E interface have to be built on die, and thus could cause problems with motherboards which use Dual PCI-E such as SLi and CrossFire?

It would be nice to see the potential of 16x exploited greatly and on die PCI-E lanes ..

Have comments post, I'm very interested in the possibilities or opinions of such an idea!

I'm a little confused by your question. Video cards aren't really constrained right now by the bandwidth to and from the CPU -- relatively little data actually has to be sent back and forth to do the rendering. And PCI-Ex16 is pretty dang fast; video cards aren't really bottlenecked at anything about AGP 4X. Improving how fast you can shuttle data between the CPU and GPU wouldn't really help that much -- unless you could move data back and forth *so* fast that you could share RAM between them freely. Then you wouldn't need to put so much memory on the graphics card.

I'm not sure that on-die PCIe would really improve performance, although it could simpify motherboard layouts (since then just the PCI/SATA/ATA controllers and any onboard sound/graphics would be in the north/southbridge).

 

[Peter]

It'd simplify standard K8 desktop mainboard design because the northbridge would be obsolete altogether - as it is now, northbridges of two-piece K8 chipsets do nothing but bridge to PCIE and hand everything else over to the south. ULi, ATI and NVidia are going to like this (because they already use HTT for the north-south interconnect), all others are going to hate it because they'll have to ditch their proprietary north-south interfaces.

 

[Googer]

As part of the Hypertransport specification, PCI like slots are allowed and there are a few rare devices that can connect to an external hypertransport link. Technicly it is possible for something like a video card to use hypertransport, but we would most likely see that in something designed from the ground up like a TiVO, Videogame Console, or some other specialized custom built computational box or proprietary computer system.

I have seen a rare motherboard with a hypertransport (HTX slot) slot on it. I'll be tough but I will do my best to try to find a photograph of one.

http://www.amdboard.com/htt_110904.html

http://www.pathscale.com/ppr_110904.html

http://www.home.agilent.com/USeng/nav/-536898278.0/pc.html


Last time that I read the specifications (a year or two ago) an external link was also allowed. This could possibly act in a similar way that FcAL does to conect a PC to an external drive array or highspeed RAM Drive. Similar to SCSI such a device would requre termination.

 

[Googer]

Originally posted by: DonaldDarko
I cant seem to start a new thread, so i'll ask here:
this is pci right?:[------------ ----]
and so is this?:[---- ------- -----]
i always thought the top one was agp? is one digital, and one analog? i'm a little confused..
i know[---------- ----][ ] is agp (4x/8x)

are you looking for a direct bus from CPU to graphics card? i think you'd need one helluva fan(or water) to cool that kind of line. so far the biggest bus iv'e seen is 400mhz. to have one from CPU to graphics, it would have to be big, like 1G or bigger to prevent not only lag, but heat to. HTT to DDR AND to graphics would be a an overload. (maybe 2 separate processors? lol, each having dual core?) i dunno, thats my guess..

PCI and AGP are very similar. AGP is an enhanced form of the PCI SIG protocol. PCI is a way for perpherials to commicate with the rest of the computer, there is no such thing as analogue PCI. All internal PC interconnects are digital, meaning they transmit 0's and 1's and not an infinately variable (analogue) waveform.

AGP is just a turbocharged version of PCI that is intended to enhance graphics cards by giving them more bandwith to communicate with the CPU and RAM. Both PCI and AGP are parallel buses'. PCI-express is the next generation of PCI, it is much faster with more bandwith and runs in a serial form.