Intel's Core 2 Extreme & Core 2 Duo: The Empire Strikes Back
by Anand Lal Shimpi on July 14, 2006 12:00 AM EST- Posted in
- CPUs
Memory Latency: No Integrated Memory Controller Necessary
One thing we pointed out in our earlier preview coverage of Intel's Core 2 Extreme is that the new processors have extremely low latency memory access, despite relying on the same memory controller as the previous generation of Intel CPUs.
Without an on-die memory controller Intel's Core 2 processor must use the memory controller in its chipset, which currently means the 128-bit DDR2 memory controller in either Intel's 965 or 975X chipset. The confusing thing is that although the Core 2 processors use the same memory controller as the old NetBurst processors, memory latency has been improved tremendously:
Intel's Core 2 processors now offer even quicker memory access than AMD's Athlon 64 X2, without resorting to an on-die memory controller. While Intel will eventually add one, the fact of the matter is that it's simply not necessary for competitive memory performance today thanks to Intel's revamped architecture. Update:As many astute readers have pointed out, Core 2's prefetchers are able to work their magic with ScienceMark 2.0, which results in the significant memory latency advantage over AMD's Athlon 64 FX-62. This advantage will not always exist; where it doesn't, AMD will continue to have lower latency memory access and where it does, Intel can gain performance advantages similar to what ScienceMark 2.0 shows.
| CPU | Everest |
CPU-Z 1.35 (8192KB, 128-byte stride) | CPU-Z 1.35 (8192KB, 64-byte stride) |
| AMD Athlon 64 FX-62 (2.8GHz) | 45.9 ns | 43.2 ns | 19.3 ns |
| Intel Core 2 Extreme X6800 (2.93GHz) | 59.8 ns | 52.8 ns | 10.9 ns |
Because of the Core 2's intelligent prefetchers doing too good of a job with ScienceMark memory latency test, we wanted to also showcase situations where Core 2 would feature slower memory accesses than the AMD platform with its integrated memory controller. Everest's results are more in line with what we'd expect to see, with the FX-62 offering over 23% faster memory accesses than the X6800. CPU-Z's latency tool also reported somewhat similar findings, with an 18% performance advantage due to AMD's integrated memory controller. CPU-Z also provided us with numbers that showcase how well Core 2 can perform if its prefetchers are able to "guess" correctly; at lower strides the Core 2 Extreme manages faster memory access than the FX-62. The 128-byte stride numbers are indicative of what will happen if the pre-fetchers are not able to get the Core 2 the data it needs, when it needs it, while the 64-byte numbers show you what can happen when things go well.
Facebook
Twitter
RSS
202 Comments
View All Comments
code255 - Friday, July 14, 2006 - link
Thanks a lot for the Rise of Legends benchmark! I play the game, and I was really interested in seeing how different CPUs perform in it.And GAWD DAMN the Core 2 totally owns in RoL, and that's only in a timedemo playback environment. Imagine how much better it'll be over AMD in single-player games where lots of AI calculations need to be done, and when the settings are at max; the high-quality physics settings are very CPU intensive...
I've so gotta get a Core 2 when they come out!
Locutus465 - Friday, July 14, 2006 - link
It's good to see intel is back. Now hopefully we'll be seeing some real innovation in the CPU market again. I wonder what the picture is going to look like in a couple years when I'm ready to upgrade again!Spoonbender - Friday, July 14, 2006 - link
First, isn't it misleading to say "memory latency" is better than on AMD systems?What happens is that the actual latency for *memory* access is still (more or less) the same. But the huge cache + misc. clever tricks means you don't have to go all the way to memory as often.
Next up, what about 64-bit? Wouldn't it be relevant to see if Conroe's lead is as impressive in 64-bit? Or is it the same horrible implementation that Netburst used?
JarredWalton - Friday, July 14, 2006 - link
Actually, it's the "clever tricks" that are reducing latency. (Latency is generally calculated with very large data sets, so even if you have 8 or 16 MB of cache the program can still determine how fast the system memory is.) If the CPU can analyze RAM access requests in advance and queue up the request earlier, main memory has more time to get ready, thus reducing perceived latency from the CPU. It's a matter of using transistors to accomplish this vs. using them elsewhere.It may also be that current latency applications will need to be adjusted to properly compute latency on Core 2, but if their results are representative of how real world applications will perceive latency, it doesn't really matter. Right now, it appears that Core 2 is properly architected to deal with latency, bandwidth, etc. very well.
Spoonbender - Friday, July 14, 2006 - link
Well, when I think of latency, I think worst-case latency, when, for some reason, you need to access something that is still in memory, and haven't already been queued.Now, if their prefetching tricks can *always* start memory loads before they're needed, I'll agree, their effective latency is lower. But if it only works, say, 95% of the time, I'd still say their latency is however long it takes for me to issue a memory load request, and wait for it to get back, without a cache hit, and without the prefetch mechanism being able to kick in.
Just technical nitpicking, I suppose. I agree, the latency that applications will typcially perceive is what the graph shows. I just think it's misleading to call that "memory latency"
As you say, it's architected to hide the latency very well. Which is a damn good idea. But that's still not quite the same as reducing the latency, imo.
Calin - Friday, July 14, 2006 - link
You could find the real latency (or most of it) by reading random locations in the main memory. Even the 4MB cache on the Conroe won't be able to prefetch all the main memory.Anyway, the most interesting is what memory latency the application that run feels. This latency might be lower on high-load, high-memory server processors (not that current benchmarks hint at this for Opteron against server-level Core2)
JarredWalton - Friday, July 14, 2006 - link
"You could find the real latency (or most of it) by reading random locations in the main memory."I'm pretty sure that's how ScienceMark 2.0 calculates latency. You have to remember, even with the memory latency of approximately 35 ns, that delay means the CPU now has approximately 100 cycles to go and find other stuff to do. At an instruction fetch rate of 4 instructions per cycle, that's a lot of untapped power. So, while it waits on main memory access one, it can be scanning the next accesses that are likely to take place and start queuing them up. And the net result is that you may never actually be able to measure latency higher than about 35 ns or whatever.
The way I think of it is this: pipeline issues aside, a large portion of what allowed Athlon 64 to outperform at first was reduced memory latency. Remember, Pentium 4 was easily able to outperform Athlon XP in the majority of benchmarks -- it just did so at higher clock speeds. (Don't *even* try to tell me that the Athlon XP 3200+ was as fast as a Pentium 4 3.2 GHz... LOL.) AMD boosted performance by about 25% by adding an integrated memory controller. Now Intel is faster at similar clock speeds, and although the 4-wide architectural design helps, they almost certainly wouldn't be able to improve performance without improving memory latency -- not just, but in actual practice. With us, I have to think that our memory latency scores are generally representative of what applications see. All I can say is, nice design Intel!
JarredWalton - Friday, July 14, 2006 - link
"...allowed Athlon 64 to outperform at first was...."Should be:
"...allowed Athlon 64 to outperform NetBurst was..."
Bad Dragon NaturallySpeaking!
yacoub - Friday, July 14, 2006 - link
""Another way of looking at it is that Intel's Core 2 Duo E6600 is effectively a $316 FX-62".Then the only question that matters at all for those of us with AMD systems is: Can I get an FX-62 for $316 or less (and run it on my socket-939 board)? If so, I would pick one up. If not, I would go Intel.
End of story.
Gary Key - Friday, July 14, 2006 - link
A very good statement. :)