Bandwidth & Latency
First, let's look at latency. One thing we saw with our last article is that generally latency gets higher with higher clock - thankfully bandwidth also goes up. We ran tests under several conditions, using different FSBs - 266 MHz (default on the E6300), 333 MHz (default on the E6550) and 400 MHz. We also change the processor multipliers - from experience, using a higher multiplier help reduce latency (somewhat). For this test, we've decided to let the motherboard apply SPD timing values, but unlike our last article we set the memory to 800 MHz (or as close as possible to 800 MHz)- which means its running asynchronous at 266 and 333 MHz.Core 2 Duo E6300 & P965
| 7x266 MHz | 7x333 MHz | 6x400 MHz | 7x400 MHz | |
| Bandwidth | 5151.21 MB/s | 5490.95 MB/s | 5893.33 MB/s | 6355.65 MB/s |
| Latency | ||||
| 4 byte stride | 3 cycles | 3 cycles | 3 cycles | 3 cycles |
| 16 byte stride | 6 | 8 | 9 | 9 |
| 64 byte stride | 26 | 32 | 35 | 35 |
| 256 byte stride | 87 | 111 | 128 | 124 |
| 512 byte stride | 99 | 126 | 147 | 141 |
Core 2 Duo E6300 & P35 - DDR2
| 7x266 MHz | 7x333 MHz | 6x400 MHz | 7x400 MHz | |
| Bandwidth | 5043.08 MB/s | 5663.68 MB/s | 5947.79 MB/s | 6285.54 MB/s |
| Latency | ||||
| 4 byte stride | 3 cycles | 3 cycles | 3 cycles | 3 cycles |
| 16 byte stride | 6 | 8 | 9 | 9 |
| 64 byte stride | 26 | 31 | 36 | 35 |
| 256 byte stride | 88 | 110 | 128 | 127 |
| 512 byte stride | 99 | 128 | 148 | 147 |
As expected, increased FSBs means higher latencies and higher bandwidth. Latency increases with increased FSBs are pretty different between the P965 and P35. Look at latencies on 256 byte stride on the table above - with the same multiplier, the P965 jumped from 87 to 111 to 124 (24 and 13 cycles, respectively). The P35 has different 'jumps' - 88 to 110 to 127 (22 and 17 cycles respectively). The same holds true to 512 byte stride as well.
At 7x400, the P35 has slightly higher latencies but surprisingly less bandwidth than the P965. Notice the increase in bandwidth is slightly different on both chipsets. On the P965, we saw an increase of around 6 percent (266 to 333 MHz) to 15 percent (333 to 400 MHz), but on the P35 its more like 12 percent (266 to 333 MHz) and 10 percent (333 MHz to 400 MHz). Now, you might say - isn't that bad? Well, not necessarily. Intel very likely tweaked the P35 memory controller so it behaves more 'predictably' with FSBs higher than 400 MHz and also works better with asynchronous memory settings - important for DDR3 modules.
Core 2 Duo E6300 & P35 - DDR3
| 7x266 MHz | 7x333 MHz | 6x400 MHz | |
| Bandwidth | 4893.08 MB/s | 5457.03 MB/s | 5849.65MB/s |
| Latency | |||
| 4 byte stride | 3 cycles | 3 cycles | 3 cycles |
| 16 byte stride | 6 | 8 | 9 |
| 64 byte stride | 25 | 32 | 36 |
| 256 byte stride | 85 | 111 | 129 |
| 512 byte stride | 96 | 130 | 153 |
We didn't ran 7x400 since from the results, it's already apparent that strapping a pair of DDR3 modules isn't going to offer much improvement. Although the DDR3 we're using ran at slightly higher clocks (480 MHz) than our DDR2 PC6400 (400 MHz), the higher latencies of the modules kept it from offering more bandwidth. Keep in mind that that official DDR3 spec memory runs at 533 MHz (1066 MHz effective). It's possible at that speed the situation is different - the additional clock should mean latencies are comparable to DDR2 800 with lower latencies.
Core 2 Duo E6550 & P965
| 7x333 MHz | 6x400 MHz | 7x400 MHz | |
| Bandwidth | 5829.77 MB/s | 5849.68 MB/s | 6299.01 MB/s |
| Latency | |||
| 4 byte stride | 3 cycles | 3 cycles | 3 cycles |
| 16 byte stride | 8 | 9 | 9 |
| 64 byte stride | 30 | 36 | 35 |
| 256 byte stride | 108 | 132 | 127 |
| 512 byte stride | 124 | 150 | 146 |
Core 2 Duo E6550 & P35 - DDR2
| 7x333 MHz | 6x400 MHz | 7x400 MHz | |
| Bandwidth | 5747.82 MB/s | 5986.77 MB/s | 6324.57 MB/s |
| Latency | |||
| 4 byte stride | 3 cycles | 3 cycles | 3 cycles |
| 16 byte stride | 8 | 10 | 9 |
| 64 byte stride | 31 | 42 | 35 |
| 256 byte stride | 106 | 129 | 124 |
| 512 byte stride | 129 | 144 | 139 |
The primary difference between the E6300 and E6550 is cache size. With higher FSBs and memory running synchronously, cache becomes less important because the penalty of fetching data from memory is smaller. However, the opposite is also true - a large cache (with the same clock and timing) should be more important when there is a difference between memory and FSB bandwidth - like in an asynchronous memory environment. At least, that's the theory. Do remember that we had to use synchronous memory settings for the E6550 with the P965. On the P35, the DDR2 memory is running at 400 MHz (actually 415 MHz at 333 MHz FSB) but timings should not change much.
We ran the E6550 from 333 MHz upwards since the processor's default FSB is 333 MHz - no sense in running the processor slower than default. Now, lets look at latencies on different FSBs without changing multipliers. At default FSB of 333 MHz, the P965 generally has higher latencies - we expected that after looking at the E6300 results. On a side note, compare the bandwidth of a synchronous bus - memory on the P965 and asynchronous bus - memory on the P35. Synchronous is still the best way to go. But at 400 MHz, running synchronously, the P35 is the one with lower latencies. Comparing bandwidth increases, the P965 offers about 8 percent increase (from 333 MHz to 400 MHz) while the P35 is about 10 percent. Latencies 'jumps' are 108 to 127 (19 cycles) for 256 byte stride and 124 to 146 (22 cycles) for 512 byte stride on the P965, while on the P35 its 106 to 124 (18 cycles) for 256 byte stride and 129 to 139 (10 cycles) for 512 byte stride.
This is where it's get interesting. You'll no doubt notice bandwidth increases (in percent) are actually smaller than what we saw with E6300 with the P35 (10 - 12 percent) and P965 (6 - 15 percent). Also, with the E6300 latencies jumped 13 cycles with P965 and 17 cycles with P35 for 256 byte stride. That's, well, that's a 6 cycle difference on the P965 and 1 cycle on the P35 with 256 byte stride and 7 and 9 cycles with the bigger, 512 byte stride. The kicker? P35 512 byte stride latency is LOWER by 9 cycles.Asynchronous / synchronous settings are not the factor here, since we're comparing results on the same chipset but different processors - both were run asynchronously at 333 MHz. With the FSB effectively the same, we wondered if cache is the only difference between these two processors. If cache is the differentiating factor, we wondered how the same P35 will handle Wolfsdale's and Yorkfield's bigger yet faster (lower latency) cache .
Core 2 Duo E6550 & P35 - DDR3
| 7x333 MHz | 6x400 MHz | 7x400 MHz | |
| Bandwidth | 5541.27 MB/s | 5685.75 MB/s | 6217.81 MB/s |
| Latency | |||
| 4 byte stride | 3 cycles | 3 cycles | 3 cycles |
| 16 byte stride | 8 | 9 | 9 |
| 64 byte stride | 32 | 37 | 35 |
| 256 byte stride | 111 | 130 | 125 |
| 512 byte stride | 127 | 150 | 145 |
Well, no surprises here. DDR3's higher latency means we're getting slightly less bandwidth than if we were to pair the P35 chipset with lower latency DDR2 modules (at relatively the same clock).
[Previous Page]
[Go to top]
[Next Page]