打印

[硬件] To qc: Tualatin vs CuMine@same speed

To qc: Tualatin vs CuMine@same speed

o個日講完之後唔覺意俾cutter虼鴗漇, 所以拖左幾日至test.

一共試左3粒u:
1. Tualatin (256KB L2) tB1 933MHz ES, multiplier 4.5-7x
2. Coppermine cB0 733MHz
3. Coppermine cD0 800MHz

Results:

Coppermine cB0 733MHz
CPUMark99: 66.8
SuperPi 1M: 2min 45s
SuperPi 2M: 6min 23s

Tualatin@733Mz tB1:
CPUMark99: 69
SuperPi 1M: 2min 42s
SuperPi 2M: 6min 17s

Coppermine cD0 800MHz:
Tualatin@733Mz tB1:
CPUMark99: 73.7
SuperPi 1M: 2min 37s
SuperPi 2M: 6min 04s

Tualatin@800Mz tB1:
CPUMark99: 74.4
SuperPi 1M: 2 min 35s
SuperPi 2M: 6min 03s


L2 Latency (using CPU-Z Latency test)

Coppermine cB0:
L1 = 3 clocks
L2 = 10 clocks (!!!)

Coppermine cD0
L1 = 3 clocks
L2 = 4 clocks

Tualatin tB1:
L1 = 3 clocks
L2 = 5 clocks

Test多粒Coppermine cC0 (Celeron 800):
L1 = 3 clocks
L2 = 4 clocks


分析:
做latency test嘅時候俾Coppermine cB0嚇左一跳, 以為latency test有事但reboot (using WinPE)左幾次再試都係同一結果. 再睇返同時脈之下嘅性能差距Coppermine cB0 vs Tualatin比Coppermine cD0落後得多 (尤其CPUMark99), 唯一嘅結論就係....

*Coppermine cB0入面有足以明顯影響性能嘅bug*

而Tualatin o係L2 latency高過coppermine cD0嘅情況下都可以取勝, 我比較過兩者L1/L2 buswidth, way associative發覺完全相同. 相信係Tualatin多左HW data prefetch logic嘅功勞.

[ Last edited by trentcys on 2006-8-2 at 23:06 ]

TOP