Core i7: Memory Speed

Page 1
Intro:

When Intel released their Core i7 CPU and the platform that enabled its use, it brought about a few new technologies that haven't been seen on the Intel side of things. One of them is the QPI or Quick Path Interconnect. Intel, being the king of "not invented here" syndrome, couldn't accept AMD's Hypertransport point to point high speed data bus. QPI is basically the same thing as Hypertransport and Intel has finally moved the memory controller onto the CPU die. No longer is there a FSB (sort of) that communicates with different components and the CPU.

The real interesting thing about the new Core i7 is the amount of memory bandwidth available to the CPU. Just how important is the speed of memory to overall system performance? While the synthetic tests prove that memory speed is faster, the real world tests show interesting results.

The DX58SO:



The motherboard being used is an Intel DX58SO known as the "Smackover." It doesn't have the same ring to it as the Bad Axe or Bonetrail/Skulltrail, but it is a new generation of product and this is the Intel extreme board. Other manufactures have taken the X58 chipset and made boards for all ends of the spectrum. The DX58SO represents a stable reference platform for use with the new Core i7 platform.

Since the northbridge no longer handles the CPU-Memory communication on Intel boards, each motherboard married to the same CPU should have similar performance for the memory subsystem unless the board manufacture added tweaks. Regardless, any board with the same CPU should give roughly the same performance.

The Core i7 920:



The CPU being used is the Core i7 920. It is a quad core processor running at 2.66GHz. Making a debut from the extreme series of CPUs is hyper-threading. Hyper-threading was very useful in the old P4 days when the CPU had resources not being used and dual core was not even on the cusp of coming out. These days, dual core CPUs are more than able to handle loads given by the majority of systems. No longer do systems freeze when doing a virus scan or other intensive process (if only they could do something about I/O transfers). On a quad core setup, hyper-threading seems to be overkill.

Kingston's Triple Channel Memory:



The memory being used will be the Kingston 3GB triple channel memory kit specifically designed for use with the Core i7 platform. The memory is rated for 1600MHz, but we opted to run it at the normal JEDEC speed of 1067MHz for this specific article. We are concerned with how the memory channel interleaving affects performance regardless of the speed and running it faster would hide the benefits. That is, if there are benefits at all. Triple channel memory kits are coming out in 3GB and 6GB kits. The 6GB kit will push you over the edge into 64-bit OS's. Linux and others like it have no problem and Windows seems to be puttering along with a few 64-bit versions as well. You'll find caveats when running 64-bit OS's. Some programs won't work and some drivers won't work. This goes for all OS's.

The Rest:

The OS being used will be Ubuntu 8.10 freshly installed with no updates on an Intel 80GB SSD (X25-M). A Radeon 4850 was used with the fglrx binary driver from the repositories. ASE Labs is proud to say it will start using the Phoronix Test Suite in aiding to write articles and reviews. Michael at Phoronix has always been a proponent of open source software and it is good to use software that is well written (and open source as well!). We will be doing our own graphing, of course. If you use a POSIX style OS, you should give the Phoronix Test Suite a shot. I'm sure there will be a GUI built for it some day. Right now it is command line only, but it more than gets the job done.

Page 2
Results:

We will be focusing on real world testing. Encoding, gaming, and other test suites were used for the real world benchmarking. RAMspeed was used for a highly synthetic test. You will see the big difference in how synthetic tests can really skew results. The new Version 4 graphing system is in use and it shows everything you need to determine the results.



RAMspeed shows that a higher memory throughput helps immensely. This test is truly synthetic and is meaningless to compare with real world results. It is very interesting to see and that is why synthetic tests are used on many sites.



The audio encoding tests show that the extra memory bandwidth is helpful in most tests, but really isn't a deal breaker if you can't have the memory running at the higher speed. Most of the remaining tests will show more of the same.



The database workloads show much of the same. The faster memory speeds help overall, but not as much as you would expect with a huge increase in memory speed.



The gaming tests show that overall, more memory bandwidth is better. If the resolution was lowered, the testing would probably show more of an increase. That is not the point when people are running games at as high a resolution as possible. What is being shown is that the memory speed isn't that big of a deal if you can't utilize it over a lower speed.

Conclusion:

There is no question about the benefit of using a higher speed memory in your system. What is interesting is when the cost benefit really goes away. Think about using a slower speed memory for half the price of the high speed counterparts. System performance may not be as slow as what was once thought. This doesn't mean if you crave speed you shouldn't buy the best of the best.

Both of the memory articles really show that overall system performance doesn't really increase unless you take a part that is really slow to begin with. The switch over to SSDs really took performance to a new level. Memory speed might not be an important enough factor when running at stock speed, but it does have an impact on how much you can overclock a system. Higher speed memory will be able to handle itself better when overclocking and that's the real benefit of using the good stuff.