First of all let’s check out the practical bandwidth and latency of the new DDR3 SDRAM. For our tests we chose Everest Ultimate Edition 4.00 utility.
The first the very first tests of the read speed from the memory reveals very interesting results. Namely, we see that DDR2 SDRAM shouldn’t be retired just yet. Thanks to lower latency, it shows better results not only when it is running at the same frequency as DDR3 SDRAM, but can also compete with DDR3 SDRAM working at “one step higher” frequency. Thus, DDR2-800 with 4-4-4-12 timings shows comparable read speed as DDR3-1066 with 8-8-8-22 timings, and DDR2-1066 with 5-5-5-15 timings come neck and neck with DDR3-1333 with 9-9-9-24 timings. So, we can conclude that high latency of DDR3 SDRAM has certain negative effect on streaming operations. That is why only the fastest DDR3-1333 in the undocumented work mode with 6-7-6-18 timings can actually demonstrate an indisputable advantage over DDR2 SDRAM.
The write speed to the memory is limited by the practical bandwidth of the processor bus that is why the results of the second Everest test cannot give us any new food for thought.
During copy speed tests, the previous-generation DDR2 shows even more impressive results than during read speed tests. DDR2-1066 with 5-5-5-15 timings is not the only one that can compete successfully with DDR3-1333 with 7-7-7-18 timings. Even DDR2-800 with 3-3-3-10 is fast enough to catch up. So, it looks like high latencies have greater effect on the copy speed than on the read speed.
The last synthetic memory subsystem benchmark aimed at revealing practical latency gives us no reason to be optimistic about the new DDR3 SDRAM. It is evident that from performance prospective DDR3-800 SDRAM doesn’t make any sense at all, as it yields to any DDR2-800 and DDR2-1066 memory. As for DDR3-1066, the only memory that may be of interest to computer enthusiasts is the one with the lowest timings. Otherwise, DDR3 SDRAM platforms will stand no chance in the performance challenge.
However, these are the conclusions we made judging only by the results of the memory subsystem performance. Now we still have to check out complex benchmarks and real applications to get a complete picture.