Intel Woodcrest, AMD's Opteron and Sun's UltraSparc T1: Server CPU Shoot-out
by Johan De Gelas on June 7, 2006 12:00 PM EST- Posted in
- IT Computing
The Official SPEC Numbers
SPEC FP and Int 2000 are the standard benchmarks to evaluate CPU performance. However, the benchmark numbers are highly dependant on the compiler. SPEC fp and Integer show the best case performance as the CPU runs on the aggressively compiled and highly optimized code. In the real world, code is compiled in a more conservative/less optimized way.
In practice this means that Intel's SPEC numbers - thanks to it's highly capable compiler team - are (slightly) higher than in real applications. Nevertheless, SPEC CPU 2000 is a good starting point to understand what a CPU is capable off. As mentioned earlier, the Xeon 5100 is the Xeon Woodcrest, based on the new core architecture.
The new Woodcrest is about 20-25% faster than the fastest dual-core Opteron. The 7% clockspeed advantage is most likely a result of the fact that the Woodcrest was baked with a newer 65nm process. If AMD manages to keep up with Intel when it comes to clockspeed, the advantage of their newest CPU might shrink to 15% or less. However, Intel's Woodcrest will have a much bigger advantage in all applications that make heavy use of 64 and 128-bit SSE.
When it comes to integer performance, the Woodcrest numbers are simply stunning and vastly superior to any other architecture. Let us find out if this vastly superior integer performance in SPEC Int 2000 pays off in server applications.
Latencies...
LMBench is a set of micro-benchmarks which can be helpful for determining memory latency and instruction latencies. We tested with LMBench 3.0a-5. It must be said that LMBench is usually right, but not always. If the benchmark is not aware of some of the particularities of a certain architecture, it can measure wrong values. So we have to double check if the values measured make sense.
The massive 4 MB L2 cache has an amazingly low latency of 14 cycles. This seems to be the worst case, as we have measured 12 cycles with other benchmarking tools such as ScienceMark. Nevertheless, even 14 cycles at 3 GHz is pretty amazing. The Core Duo, a.k.a. Yonah, accesses a shared cache that's half as large in 14 cycles at a substantially lower 2.33 GHz.
On the other hand, the memory latency very high; luckily the 4 MB L2 cache will minimize that effect. The problem seems to be the FB-DIMMs. The Advanced Memory Buffer introduces extra latency, and of course the registered DDR-2 533 chips with a CAS latency of 4 have a higher latency by themselves. This results in a memory subsystem with pretty high 115 ns latency, while the Opteron has access to the RAM in only 73 ns
ScienceMark didn't agree completely and reported about 65-70 ns latency on the Opteron system and 70-76 ns (230 cycles) on the Woodcrest system. We have reason to believe that Woodcrest's latency is closer to what LMBench reports: the excellent prefetchers are hiding the true latency numbers from Sciencemark. It must also be said that the measurements for the Opteron on the Opteron are only for the local memory, not the remote memory.
SPEC FP and Int 2000 are the standard benchmarks to evaluate CPU performance. However, the benchmark numbers are highly dependant on the compiler. SPEC fp and Integer show the best case performance as the CPU runs on the aggressively compiled and highly optimized code. In the real world, code is compiled in a more conservative/less optimized way.
In practice this means that Intel's SPEC numbers - thanks to it's highly capable compiler team - are (slightly) higher than in real applications. Nevertheless, SPEC CPU 2000 is a good starting point to understand what a CPU is capable off. As mentioned earlier, the Xeon 5100 is the Xeon Woodcrest, based on the new core architecture.
| SPECfp | ||
| Clockspeed | SPEC fp 2000 | |
| POWER5+ | 2200 | 3271 |
| Itanium 2 | 1666 | 2851 |
| Xeon 5160 | 3000 | 2783 |
| Opteron | 2800 | 2256 |
| Pentium 4 E | 3733 | 2232 |
The new Woodcrest is about 20-25% faster than the fastest dual-core Opteron. The 7% clockspeed advantage is most likely a result of the fact that the Woodcrest was baked with a newer 65nm process. If AMD manages to keep up with Intel when it comes to clockspeed, the advantage of their newest CPU might shrink to 15% or less. However, Intel's Woodcrest will have a much bigger advantage in all applications that make heavy use of 64 and 128-bit SSE.
| SPECint | ||
| Clockspeed | SPEC Int 2000 | |
| Xeon 5160 | 3000 | 3057 |
| Pentium 4 E | 3733 | 1870 |
| Opteron | 2800 | 1837 |
| Pentium 4 Xeon | 3733 | 1813 |
| POWER5+ | 2200 | 1705 |
| Itanium 2 | 1666 | 1502 |
When it comes to integer performance, the Woodcrest numbers are simply stunning and vastly superior to any other architecture. Let us find out if this vastly superior integer performance in SPEC Int 2000 pays off in server applications.
Latencies...
LMBench is a set of micro-benchmarks which can be helpful for determining memory latency and instruction latencies. We tested with LMBench 3.0a-5. It must be said that LMBench is usually right, but not always. If the benchmark is not aware of some of the particularities of a certain architecture, it can measure wrong values. So we have to double check if the values measured make sense.
| LMBench | |||||||
| Clockspeed | L1 (ns) | L1 (cycles) | L2 (ns) | L2 (cycles) | RAM (ns) | RAM (cycles) | |
| Xeon 5160 3 GHz | 3000 | 1.01 | 3 | 4.7 | 14 | 117.3 | 345 |
| Pentium- M 1.6 GHz | 1593 | 2 | 3 | 6 | 10 | 92.1 | 147 |
| Sun T1 1 GHz | 980 | 3 | 3 | 22.1 | 22 | 107.5 | 105 |
| Opteron 275 | 2209 | 1 | 3 | 5.5 | 12 | 73 | 161 |
| Xeon Irwindale 3.6 GHz | 3594 | 1 | 4 | 8 | 28 | 48.8 | 175 |
The massive 4 MB L2 cache has an amazingly low latency of 14 cycles. This seems to be the worst case, as we have measured 12 cycles with other benchmarking tools such as ScienceMark. Nevertheless, even 14 cycles at 3 GHz is pretty amazing. The Core Duo, a.k.a. Yonah, accesses a shared cache that's half as large in 14 cycles at a substantially lower 2.33 GHz.
On the other hand, the memory latency very high; luckily the 4 MB L2 cache will minimize that effect. The problem seems to be the FB-DIMMs. The Advanced Memory Buffer introduces extra latency, and of course the registered DDR-2 533 chips with a CAS latency of 4 have a higher latency by themselves. This results in a memory subsystem with pretty high 115 ns latency, while the Opteron has access to the RAM in only 73 ns
ScienceMark didn't agree completely and reported about 65-70 ns latency on the Opteron system and 70-76 ns (230 cycles) on the Woodcrest system. We have reason to believe that Woodcrest's latency is closer to what LMBench reports: the excellent prefetchers are hiding the true latency numbers from Sciencemark. It must also be said that the measurements for the Opteron on the Opteron are only for the local memory, not the remote memory.
Facebook
Twitter
RSS
91 Comments
View All Comments
blackbrrd - Wednesday, June 7, 2006 - link
Finally Intel can give AMD some real competition in the two socket server market. This shows why Dell only wanted to go with AMD for 4S and not 2S server systems...245w vs 374w and a huge performance lead over the previous Intel generation is a huge leap for Intel.
It will be interesting to see how much these systems are going to cost:
1) is the fb-dimm's gonna be expensive?
2) is the cpu's gonna be expensive?
3) is the motherboards gonna be expensive?
For AMD neither the ram nor the motherboards are expensive, so I am curious how this goes..
If anybody thinks I am an Intel fanboy, I have bought in this sequence: intel amd intel intel, and I would have gotten and amd instead of an intel for the last computer, except I wanted a laptop ;)
JarredWalton - Wednesday, June 7, 2006 - link
For enterprise servers, price isn't usually a critical concern. You often buy what runs your company best, though of course there are plenty of corporations that basically say "Buy the fastest Dell" and leave it at that.FB-DIMMs should cost slightly more than registered DDR2, but not a huge difference. The CPUs should actually be pretty reasonably priced, at least for the standard models. (There will certainly be models with lots of L3 cache that will cost an arm and a leg, but that's a different target market.)
Motherboards for 2S/4S are always pretty expensive - especially 4S. I would guess Intel's boards will be a bit more expensive than equivalent AMD boards on average, but nothing critical. (Note the "equivalent" - comparing boards with integrated SCSI and 16 DIMM slots to boards that have 6 DIMM slots is not fair, right?)
Most companies will just get complete systems anyway, so the individual component costs are only a factor for small businesses that want to take the time to build and support their own hardware.
blackbrrd - Wednesday, June 7, 2006 - link
Registered DDR2 is dirt cheap, so if FB-DIMMs are only slightly more expensive thats really good.About compareing 6 DIMM slot and 16 DIMM slot motherboards, I agree, you can't do it. The number of banks is also important, we have a motherboard at work with 8 ranks and 6 DIMM slots, so only two of the slots can be filled with the cheapest 2gb dual rank memory. Currently 2gb single ranks modules are 50% more expensive than dual rank modules.
Which brings another question.. Does FB-DIMM have the same "problem" with rank limit in addition to slot limit? Or does the FB part take care of that?
BaronMatrix - Wednesday, June 7, 2006 - link
why are we running servers with only 4GB RAM. I have that in my desktop. Not ot nitpick but I think you shuld load up 16GB and rerun the tests. If not this is a low end test, not HPC. I saw the last Apache comparison and it seems like the benchmark is different. Opteron was winning by 200-400% in those tests. What happened?JohanAnandtech - Wednesday, June 7, 2006 - link
Feel free to send me 12 GB of FBDIMMs. And it sure isn't a HPC test, it is a server test."I saw the last Apache comparison and it seems like the benchmark is different. Opteron was winning by 200-400% in those tests. What happened? "
A new Intel architecture called "Core" was introduced :-)
BaronMatrix - Wednesday, June 7, 2006 - link
I didn't say the scores, I said the units in the benchmark. I'm not attacking you. It just stuck out in my head that the units didn't seem to be the same as the last test with Paxville. By saying HPC, I mean apps that use 16GB RAM, like Apache/Linux/Solaris. I'm not saying you purposely couldn't get 12 more GB of RAM but all things being equal 16GB would be a better config for both systems.I've been looking for that article but couldn't find it.
JohanAnandtech - Wednesday, June 7, 2006 - link
No problem. Point is your feedback is rather unclear. AFAIK, I haven't tested with Paxville. Maybe you are referring to my T2000 review, where we used a different LAMP test, as I explained in this article. In this article the LAMP server has a lot more PHP and MySQL work.http://www.anandtech.com/IT/showdoc.aspx?i=2772&am...">http://www.anandtech.com/IT/showdoc.aspx?i=2772&am...
See the first paragraph
And the 4 GB was simply a matter of the fact that Woodcrest had 4 GB of FB DIMM.
JarredWalton - Wednesday, June 7, 2006 - link
Most HPC usage models don't depend on massive amounts of RAM, but rather on data that can be broken down into massively parallel chunks. IBM's BlueGene for example only has 256MB (maybe 512MB?) of RAM per node. When I think of HPC, that's what comes to mind, not 4-way to 16-way servers.The amount of memory used in these benchmarks is reasonable, since more RAM only really matters if you have data sets that are too large to fit with the memory. Since our server data sets are (I believe) around 1-2GB, having more than 4GB of RAM won't help matters. Database servers are generally designed to having enough RAM to fit the vast majority of the database into memory, at least where possible.
If we had 10-14GB databases, we would likely get lower results (more RAM = higher latency among other things), but the fundamental differences between platforms shouldn't change by more than 10%, and probably closer to 5%. Running larger databases with less memory would alter the benchmarks to the point where they would largely be stressing the I/O of the system - meaning the HDD array. Since HDDs are so much slower than RAM (even 15K SCSI models), enterprise servers try to keep as much of the regularly accessed data in memory as possible.
As for the Paxville article, click on the "IT Computing" link at the top of the website. Paxville is the second article in that list (and it was also linked once or twice within this article). Or http://www.anandtech.com/IT/showdoc.aspx?i=2745">here's the direct link.
BaronMatrix - Wednesday, June 7, 2006 - link
Thx for the link, but the test I was looking at was Apache and showed concurrency tests. At any rate, just don't think I was attacking you. I was curious as to the change in units I noticed.MrKaz - Wednesday, June 7, 2006 - link
How much will it cost?If Conroe XE 2.9Ghz is 1000$.
Then I assume that this will cost more.
I think looks good, but it will depends a lot of the final price.
Also does that FBdimm have a premium price over the regular ones?