HTX Coprocessor for AMD

[The little lost angel]

Just wondering, since HT can be used as a bus the way PCI-Express can,
is there any reason why a coprocessor on a HT connection shoud not/
cannot be done in the way of a slot in card? Wouldn't this take up
less board real estate and allow much more flexible designs since the
card can be much bigger than a predefined socket size?

 

[Del Cecchi]

Just wondering, since HT can be used as a bus the way PCI-Express can,
is there any reason why a coprocessor on a HT connection shoud not/
cannot be done in the way of a slot in card? Wouldn't this take up
less board real estate and allow much more flexible designs since the
card can be much bigger than a predefined socket size?

Already someone sells InfiniBand connection that plugs into HTX. And the
new HT3 has the stuff to go more than a few inches.

 

[Yousuf Khan]

Just wondering, since HT can be used as a bus the way PCI-Express can,
is there any reason why a coprocessor on a HT connection shoud not/
cannot be done in the way of a slot in card? Wouldn't this take up
less board real estate and allow much more flexible designs since the
card can be much bigger than a predefined socket size?

I think they're considering that too.

Yousuf Khan

 

[nobody]

I think they're considering that too.

Yousuf Khan

Maybe that's a stretch but...Could that coprocessor be a GPU? HTX
video card, anyone? With truckloads of bandwidth to CPU and main
memory, and no extra latency that it is otherwise inevitably
introduced by any intermediate components (PCIe bus, PCIe controller,
and the chipset as a whole).

NNN

 

[The little lost angel]

Maybe that's a stretch but...Could that coprocessor be a GPU? HTX
video card, anyone? With truckloads of bandwidth to CPU and main
memory, and no extra latency that it is otherwise inevitably
introduced by any intermediate components (PCIe bus, PCIe controller,
and the chipset as a whole).

That would make quite an ugly hack with some cables running off to one
side in order to place the video connector no?

 

[Yousuf Khan]

Maybe that's a stretch but...Could that coprocessor be a GPU? HTX
video card, anyone? With truckloads of bandwidth to CPU and main
memory, and no extra latency that it is otherwise inevitably
introduced by any intermediate components (PCIe bus, PCIe controller,
and the chipset as a whole).

I wouldn't call a GPU a coprocessor. Although, I don't know why, guess
I'm just used to it. It would probably fit the definition of a coprocessor.

Yousuf Khan

 

[Tony Hill]

Maybe that's a stretch but...Could that coprocessor be a GPU? HTX
video card, anyone? With truckloads of bandwidth to CPU and main
memory, and no extra latency that it is otherwise inevitably
introduced by any intermediate components (PCIe bus, PCIe controller,
and the chipset as a whole).

It's possible, yes. However, as with most things, there probably
isn't much point in doing so. Video cards are quite bandwidth
intensive, particularly when compared to most other devices in PCs,
but not all that dependant on low latency. With PCI-E 16x they have
the same sort of bandwidth that Hypertransport, in it's current
incarnation, provides already and future higher-bandwidth versions
aren't likely to produce much of a performance boost (really we
haven't seen too much of an improvement since the bandwidth figures of
AGP 2x).

The problem is simply that if a video card has to go off-card for
memory access, it's going to be VERY slow in terms of both bandwidth
and latency, regardless of what sort of interface you use. An extra
couple of gigabytes/second or a few nanoseconds less latency isn't
going to help matters much, it's still REAL slow as compared to the
25GB/s+ bandwidth and extremely low latency of on-card memory.

So basically while an HTX video card is certainly possible, it would
probably only buy you about 1% over PCI-E 16x, it would cost a lot
more to develop and it would only be usable by a VERY small number of
users. Even if all AMD-based systems came with HTX slots you would
still only be looking at ~20% of the market, probably not worthwhile.

 

[Keith]

It's possible, yes. However, as with most things, there probably
isn't much point in doing so. Video cards are quite bandwidth
intensive, particularly when compared to most other devices in PCs,
but not all that dependant on low latency. With PCI-E 16x they have
the same sort of bandwidth that Hypertransport, in it's current
incarnation, provides already and future higher-bandwidth versions
aren't likely to produce much of a performance boost (really we
haven't seen too much of an improvement since the bandwidth figures of
AGP 2x).

The problem is simply that if a video card has to go off-card for
memory access, it's going to be VERY slow in terms of both bandwidth
and latency, regardless of what sort of interface you use. An extra
couple of gigabytes/second or a few nanoseconds less latency isn't
going to help matters much, it's still REAL slow as compared to the
25GB/s+ bandwidth and extremely low latency of on-card memory.

Why do you assume there would be no local memory on this widget?

So basically while an HTX video card is certainly possible, it would
probably only buy you about 1% over PCI-E 16x, it would cost a lot more
to develop and it would only be usable by a VERY small number of users.
Even if all AMD-based systems came with HTX slots you would still only
be looking at ~20% of the market, probably not worthwhile.

Ok, lets play with this a while. Instead of a plug-in card, suppose we
use the (S-939, or whatever) socket that's there and use one of the HT
links to send data to a sorta normal (or perhaps stripped) graphics card.
The GPU co-processor would then have its own memory and access to the
CPU's over ccHT. The CPU could use the cc-link to expand memory too.

 

[daytripper]

Why do you assume there would be no local memory on this widget?


Ok, lets play with this a while. Instead of a plug-in card, suppose we
use the (S-939, or whatever) socket that's there and use one of the HT
links to send data to a sorta normal (or perhaps stripped) graphics card.
The GPU co-processor would then have its own memory and access to the
CPU's over ccHT. The CPU could use the cc-link to expand memory too.

Put the bong down, and slowly back away from the keyboard...

/daytripper (hth ;-)

 

[Keith]

Put the bong down, and slowly back away from the keyboard...

Bong? I got no steenkin' bong, man! ...just poured a second cuppa
coffee (first evidently didn't work).

 

[Tony Hill]

Why do you assume there would be no local memory on this widget?

I'm not. I'm just saying that the off-chip memory bandwidth is going
to be slow regardless, so everything is going to be designed to avoid
going off-chip as much as possible, thereby minimizing the penalty for
such memory access.

Ok, lets play with this a while. Instead of a plug-in card, suppose we
use the (S-939, or whatever) socket that's there and use one of the HT
links to send data to a sorta normal (or perhaps stripped) graphics card.
The GPU co-processor would then have its own memory and access to the
CPU's over ccHT. The CPU could use the cc-link to expand memory too.

I suppose it's possible, but I just don't see it happening in the PC
world. Too much complexity (aka cost) for too little gain.