2 configs: Same mem MHZ, different CPU GHZ, yet slower CPU hashigher mem bandwidth

[lorkp]

I've noticed something. Running 210 HTT 1:1 gives higher bandwidth according to memtest and sisoft than 228HTT 200:183 divider. Both with 11 multipliers. Both equate to about 210MHZ memory (confirmed by CPU Z), and the later configuration has a faster processor speed (2.5GHZ vs 2.3GHZ). Yet in memtest, the former configuration scores roughly 1700MB/s vs. 1580MB/s in memtest and 2700 vs. 2400 in sisoft. Why is this?

 

[Lord Evermore]

Technically the hypertransport bus isn't even involved in memory transfers, since the CPU has the memory controller and a direct path to memory. However if any of the data has to be measured including the HTT path, to be determined by the benchmark apps, that could result in differences.

What you may be running into is the result of the differences in transfer cycles. If you run two busses out of sync, not 1:1, then you end up with one or the other often having to "wait" a clock tick or two or 3 for the other one to be ready for a data transfer. Even though one might be ready "more often" because of a higher frequency, it still isn't ready at exactly the same time. Running things at a ratio of like 2:1 can eliminate some of that, since it's always exactly the same number of cycles difference, but at that point you have lost a lot of the pure bandwidth on one side so it's still slower.

Running the busses at different speeds can still result in higher performance, if it allows one of them to reach massively higher speeds. For instance P4's got to much higher bus speeds than the memory speeds could keep up, and it was still faster to have them at that bus speed than to have them synced with the memory. But if you've only got a difference of a few megahertz, it's better to maintain sync.

 

[lorkp]

Thanks for your reponse Lord Evermore, it makes sense.

Here's something else I noticed. at gogar's athlon 64 OC optimizer a sample result suggests a 11 multiplier with 227 HTT (i'm not sure what to lable that number) and a 183 ram divider. One would think, that this would yield a MHZ of roughly 208, but the calculator says 192.

However, when using a multiplier of 10, the expected ratio is acurate. For example, a 241 HTT with a 166 divider is 200.

But the figure of 192MHZ with the 183 ram divider seems accurate. WHen I purposely set the RAM to 192MHZ with a 10x, 231 HTT and 166 divider, the bandwidth is roughly the same as the figure I got with 227 HTT and 183 ram divider and 11 multiplier.

So I guess it has something to do with the multiplier. Maybe this has something to do with the clock syncing thing seeing that 10 is easier to work with?

 

[Lord Evermore]

I don't get so intense with overclocking and sucking the last drop of performance out of things, my head hurts just looking at a table of results from that optimizer page. I'm not going to try to figure out what formulas he's using to come up with results. And I am happy with low digit dividers, like 5/4. So I can't really help anymore.

 

[lorkp]

haha, okay