Intel Woodcrest, AMD's Opteron and Sun's UltraSparc T1: Server CPU Shoot-out

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.

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.