Are AMD going HTX for Desktops?

[Techno+]

http://www.theinquirer.net/default.aspx?article=36589

it says that the chipset supports HTX slots, so does this mean AMD is making GPUs with HTX slots?

 

[3dilettante]

A 16-bit HT link would supply ~6 GB/s each way, while a PCI-E 16x gives a bit more than 4.

Bandwidth wise, it would be better, and it would remove the latency of the chipset bridge.

On the other hand, there are few HT links to spare on desktop A64s. I think current desktop chips disable most of the links. Maybe one could be used, I forget how many are disabled, but that would rule out Crossfire setups.

It would be a server or workstation CPU with more links and a board that could support them that could use this.

Perhaps 4x4 could be made to do this, or even two HTX cards, since that platform is more workstation than desktop with the extra sockets.

It sounds doable, but it also sounds like a niche product.

I'm not sure how much there is to gain, since the long-latency link between video card and the rest of the system has been designed around for so long.

I suppose benchmarks would be needed, if AMD ever decided to make such a product available for general consumption. Some HPC clients might use it, since they have more money to burn than a gamer.

 

[DudeMiester]

You don't need the CPU to have the required number of HT links, since you can make the NB an arbiter, just like it currently is with PCIe. Although, this would increase the latency somewhat, it would still be an improvement. Also, HT 3.0 may likely enable much faster HTX connectors.

 

[3dilettante]

You don't need the CPU to have the required number of HT links, since you can make the NB an arbiter, just like it currently is with PCIe. Although, this would increase the latency somewhat, it would still be an improvement. Also, HT 3.0 may likely enable much faster HTX connectors.

From the ambiguous wording in the INQ story, it said they were connecting the card to CPU and memory, which sounded like a direct connection.

That wouldn't really be a chipset issue, so your idea makes sense, the INQ's grammar notwithstanding.

 

[fehu]

You don't need the CPU to have the required number of HT links, since you can make the NB an arbiter, just like it currently is with PCIe. Although, this would increase the latency somewhat, it would still be an improvement. Also, HT 3.0 may likely enable much faster HTX connectors.

why go on a shared bus (nb to cpu) when already exist PCIe 4x and the advantage of HTX is latency and direct connection to the cpu?

who want to relise an htx board that can only work on amd 4x4?
even ageia is not so suicide

 

[3dilettante]

An arbiter wouldn't need to go through the process of converting from PCI-E signalling to HT.

It would be lower latency than a bridge chip, but longer latency than a direct connection.
I'm not sure the difference would be enough to care either way.

I'm not clear on whether this HTX slot would be kept coherent, though probably not.
Something like that would eat into the use of Torrenza, and it would would increase memory latencies to that of a 2-socket setup without having 2 sockets.

 

[fehu]

...

even if it's technically doable, is not commercially

whit everything integrated on the mboard, pci, unused pcie 4x, and PCIe2 on the orizon who need another niche not coherent connection?

 

[3dilettante]

I've already said it doesn't sound like much of a general product.

However, there may be some niche customers that wouldn't mind it.