Opteron, socket 940 headroom?

[CB]

Any idea what the future of the current configuration (socket 940) for
Opterons is? I note that AMD is moving their desktop line to a socket
939 design. The implication I see is that 940 pins is sufficient for
current and future speed increases. Does this suggest a reasonable
processor upgrade path for someone considering investing in a current
generation dual opteron board such as the Tyan 2885 or the iWill DK8x
(not cheap).

I have also read that the single processor implementations using
Nvidia's nForce 3 chipset (Asus SK8N for example) don't take full
advantage of the memory bandwidth available. Is this so?

 

[Tony Hill]

Any idea what the future of the current configuration (socket 940) for
Opterons is? I note that AMD is moving their desktop line to a socket
939 design. The implication I see is that 940 pins is sufficient for
current and future speed increases. Does this suggest a reasonable
processor upgrade path for someone considering investing in a current
generation dual opteron board such as the Tyan 2885 or the iWill DK8x
(not cheap).

My understanding is that AMD will stick with Socket 940 for the
Opteron through 2005, so you've probably got at least a few speed
bumps in the pipeline. That being said, just because the socket is
the same doesn't necessarily mean that the chip will work. Changes in
power supply requirements are likely to be more of a limiting factor
than any socket change. I have no clue how far today's motherboards
will be able to go in this regard, but I would guess that it would be
at least 3 or 4 speed grades beyond the fastest available today (x48
Opteron chips).

I have also read that the single processor implementations using
Nvidia's nForce 3 chipset (Asus SK8N for example) don't take full
advantage of the memory bandwidth available. Is this so?

For memory bandwidth, no. What you've likely being hearing about is
the lower hypertransport speed of the nForce3 150 chipset. The AMD
Opteron (and Athlon64) supports HT speeds of up to 1600MT/s and
16-bits in either direction, for a total bandwidth of 3.2GB/s each
way. However the nVidia chipset does not support this full bandwidth,
it only runs it's HT links at 1200MT/s and 16-bits in the
CPU-to-chipset direction and 1600MT/s and 8-bits in the chipset-to-CPU
direction (I *think* I've got those numbers right, though they might
be off by a little bit). This reduces the I/O bandwidth available for
AGP and DMA devices hanging off the chipset/PCI bus.

Mostly this only affects high-end graphics work (including the odd
game) and doesn't have a huge effect on performance, but it's enough
that the Athlon64/Opteron chipsets from VIA and SiS are usually a bit
faster. Fortunately nVidia has fixed this problem with their new
nForce3 250 chipset that should start appearing soon.

 

[CB]

My understanding is that AMD will stick with Socket 940 for the
Opteron through 2005, so you've probably got at least a few speed
bumps in the pipeline. That being said, just because the socket is
the same doesn't necessarily mean that the chip will work. Changes in
power supply requirements are likely to be more of a limiting factor
than any socket change. I have no clue how far today's motherboards
will be able to go in this regard, but I would guess that it would be
at least 3 or 4 speed grades beyond the fastest available today (x48
Opteron chips).

Thanks Tony. I set up a workstation about 18 months ago with dual FC
1Ghz PIIIs expecting to be able to bump up the performance down the
road. But the Tulatin chips that would have offered a nice boost were
pin incompatible and the PIIIs have been plagued with overheating
problems despite considerable fussing with fans and heatsinks.

I need to upgrade the system and wish to avoid similar problems. I'd
like to invest in a good stable base platform and save funds by starting
off with the mid-range processor speed stepping (should still provide a
considerable boost in performance) and perhaps upgrade the processors in
6 months when the prices have dropped a bit.

The system runs spatial analysis and GIS image processing tasks that
often run for minutes to hours or overnight on the current box. The
user would like to be able to run such a task in the background while
working on other aspects of the project - thus the dual cpu platform is
attractive.