PDA

View Full Version : Opteron memory performance


Harnagel
04-24-03, 10:25 PM
If the Opteron has a built in dual channel DDR controller shouldn't it take 4 DIMMs to get the max bandwidth on a dual processor board, thus doubling the bandwidth over a single processor system? Why is it then that the single processor system scores higher than the dual system?

netviper13
04-24-03, 11:31 PM
It is only a single-channel DDR controller.

StealthHawk
04-25-03, 08:26 AM
Originally posted by netviper13
It is only a single-channel DDR controller.

bingo.

the "dual channel" of Opteron is actually AMD's fuzzy marketing term for doubling the size of the bus.

Harnagel
04-25-03, 02:34 PM
Oh, thanks for setting me straight on that. So it's a single channel 64bit controller?

But even with that, my real question is why doesn't having two CPUs and thus two memory controllers have more bandwidth than one? Sorry for coming across as such an ignoramous, but I thought that was one of the main points to having the on-die controller.

StealthHawk
04-25-03, 05:16 PM
Originally posted by Harnagel
Oh, thanks for setting me straight on that. So it's a single channel 64bit controller?

no, it's a 144bit controller.

But even with that, my real question is why doesn't having two CPUs and thus two memory controllers have more bandwidth than one? Sorry for coming across as such an ignoramous, but I thought that was one of the main points to having the on-die controller.

not sure on the answer to that, haven't really followed the intricacies of Opteron and multi-CPU setups.

Harnagel
04-25-03, 05:32 PM
no, it's a 144bit controller.

OK that makes sense, I forgot about ECC, and also explains why memory has to be installed in pairs. Thanks.

Anyone else happen to know anything about my second question?

SavagePaladin
04-25-03, 10:26 PM
Opteron uses a different memory setup. Each processor has its own memory, but theres a penalty involved in processor 1 using processor 2s memory, and so on. I don't know the whole deal, and my brain is operating in low power mode, so I can't tell you much more in the way of details.

Harnagel
04-25-03, 11:02 PM
NUMA (non-unified memory architecture). Does this tech refer to the on die cache memory or system memory? It seems like the OS plays a major role in making NUMA work optimally, so if this has anything to do with system memory with multi-processor setups amybe we will see an improvement with Windows2003...

I'll keep looking and post what I find.

Harnagel
04-25-03, 11:27 PM
OK grabbed this from anandtech, it pretty much answers my question.

In order for the performance benefit of this sort of memory access to be truly taken advantage of, the OS needs to be smart enough to not put all data in the first xxxMB of memory. Instead, the OS must keep data in memory in such a way as to optimize for local and non-local memory accesses. For example, in a 4-way Opteron server with each CPU having 1GB of memory, if the working dataset is only 512MB in size it shouldn't all be placed in CPU0's memory - especially if all four CPUs are using the data. It should either be copied to all four sets of memory, or it should be divided up so that all CPUs can have at least some local access to the data at their full 5.3GB/s rate. This type of memory access is known as NUMA, which stands for Non-Uniform Memory Access; Windows 2003 Server supposedly has support for NUMA.

SavagePaladin
04-26-03, 01:01 AM
Win2k3 server actually doesn't quite YET, though you can download a beta support patch, from what I know. But yeah that sums it up

Cotita
04-26-03, 03:53 PM
Opteron performance is quite good, even tough windows nor linux are optimized for it.

Once 64 bit versions of windows and linux come out, we will see even better performance.

I have my doubts about the athlon64 because of the available bandwith, but I hope its significantly faster than the barton400 at the same clockspeed

SavagePaladin
04-26-03, 07:47 PM
It has SSE2, which is an immediate benefit to gamers, anbd most of the HT links on an Opteron are used for multiprocessor stuff so wouldn't really fit on the Ath64. As for the memory stuff, it certainly could make a difference, though I don't know if it will.

Harnagel
04-27-03, 03:25 AM
It's very probable that a good portion of the performance improvements came from lower latencies. Being that the Athlon64 will also have an on-die memory controller we could very well see similar performances in games. Time will tell, but I am fairly optimistic. I expect to see substantially higher clock speeds as well, due to...

1) less l2 cache (not a big peformance issue in most apps)
2) not as many ht units (not needed as Savage explained)
3) doesn't need to be checked for multi-processor stability
4) fab process improvements between now fall

With all of this I wouldn't be surprised to see retail A64s hitting 2.5 to 2.8 Ghz by the end of the year.

SavagePaladin
04-27-03, 05:35 AM
There may yet be 1 meg l2 cache on the Ath64s. I've seen statements to this effect around, anyway

Harnagel
04-27-03, 01:52 PM
That woul be nice :D ...I hate waiting

Kain
04-27-03, 02:01 PM
Damn.

I hate waiting too. Look at my current computer, I seriously need an AMD Athlon 64. :D