HOT: In depth memory performance analysis

Using Cachemem 2.65MMX (latest version I believe). Booted into dos with no mscdex or other programs:

250FSB 5:4 ratio, 200Mhz 2-2-2-5 timings, GAT: SR-A-A-D-D, 800PSB Strap (Partial PAT)

Cache size/Memory speed info tool 2.65MMX - (c) 1999-2001, LRMS - DJGPP compiled
CPUID support detected... 'GenuineIntel' with FPU TSC MMX
Family=15 Model=2 Step=9 Type=0 Chipset (Vendor/Device ID(Rev)): Intel/2570(02)
CPU clock: 3007.7 MHz
Using 32MB physical memory block (alignment = 32)
Bandwidth - MMX linear access test... Read/Write/Copy (MB/s)
Block of 1KB: 23684.3 / 9989.2 / 19988.5
Block of 2KB: 23308.8 / 10092.5 / 19522.2
Block of 4KB: 22422.1 / 9983.7 / 19780.7
Block of 8KB: 22097.1 / 10089.6 / 18677.1
Block of 16KB: 12226.4 / 10144.7 / 18855.0
Block of 32KB: 12230.9 / 10170.9 / 19097.5
Block of 64KB: 12234.1 / 10167.8 / 19271.5
Block of 128KB: 12235.4 / 10182.6 / 17915.4
Block of 256KB: 12177.7 / 10095.4 / 18486.0
Block of 512KB: 11757.9 / 9955.8 / 3050.9
Block of 1024KB: 3400.0 / 2399.4 / 2394.1
Block of 2048KB: 3418.7 / 2398.8 / 2394.6
Block of 4096KB: 3418.8 / 2399.0 / 2394.6
Block of 8192KB: 3418.9 / 2399.1 / 2394.7
Block of 16384KB: 3419.1 / 2399.2 / 2394.7
Block of 32768KB: 3419.2 / 2397.8
Latency - Memory walk tests... ("pointer chasing")
Null size: 2 cycles 0 cycles (overhead 376 cycles)
steps: 4 8 16 32 64 128 256 512 1k 2k 4k (bytes)
Block of 1KB: 1 1 1 1 1 1 1 1 - - - cycles
Block of 2KB: 1 1 1 1 1 1 1 1 1 - - cycles
Block of 4KB: 1 1 1 1 1 1 1 1 1 1 - cycles
Block of 8KB: 1 2 2 4 2 2 1 1 1 1 1 cycles
Block of 16KB: 2 3 8 18 18 18 18 18 18 19 2 cycles
Block of 32KB: 2 3 8 18 18 18 18 18 18 18 19 cycles
Block of 64KB: 2 2 4 10 17 17 18 20 20 20 24 cycles
Block of 128KB: 2 2 4 10 18 18 18 19 21 20 24 cycles
Block of 256KB: 2 2 4 10 17 18 18 22 24 23 24 cycles
Block of 512KB: 5 3 4 12 18 21 21 24 25 32 43 cycles
Block of 1024KB: 10 8 17 33 65 242 242 245 248 250 253 cycles
Block of 2048KB: 11 8 16 32 62 236 238 243 244 250 256 cycles
Block of 4096KB: 11 9 16 31 62 236 237 244 249 252 257 cycles
Block of 8192KB: 10 9 17 33 57 202 203 212 216 225 236 cycles
Block of 16384KB: 11 8 15 28 56 203 204 211 218 227 243 cycles
Block of 32768KB: 10 8 15 28 56 202 202 212 218 228 243 cycles
This system appears to have 2 cache levels (enabled).
L1 cache (8KB) speed (MB/s): Read=22422.1, Write=9983.7
L2 cache (512KB) speed (MB/s): Read=12177.7, Write=10095.4
Main memory speed (MB/s): Read=3419.1, Write=2399.2





250FSB 5:4 ratio, 200Mhz 2-2-2-5 timings, GAT: SR-A-A-D-D, 667PSB Strap (FULL PAT)

Cache size/Memory speed info tool 2.65MMX - (c) 1999-2001, LRMS - DJGPP compiled
CPUID support detected... 'GenuineIntel' with FPU TSC MMX
Family=15 Model=2 Step=9 Type=0 Chipset (Vendor/Device ID(Rev)): Intel/2570(02)
CPU clock: 3007.7 MHz
Using 32MB physical memory block (alignment = 32)
Bandwidth - MMX linear access test... Read/Write/Copy (MB/s)
Block of 1KB: 23684.3 / 9989.8 / 19988.9
Block of 2KB: 23308.8 / 10092.6 / 19518.6
Block of 4KB: 22418.9 / 9983.8 / 19780.6
Block of 8KB: 22100.1 / 10089.5 / 18685.0
Block of 16KB: 12226.5 / 10144.6 / 18928.2
Block of 32KB: 12231.2 / 10171.1 / 19099.4
Block of 64KB: 12234.1 / 10167.8 / 19270.9
Block of 128KB: 12222.3 / 10182.5 / 18180.7
Block of 256KB: 12184.4 / 10095.5 / 18420.5
Block of 512KB: 11780.3 / 9958.9 / 3089.8
Block of 1024KB: 3798.9 / 2406.4 / 2416.5
Block of 2048KB: 3814.0 / 2406.0 / 2416.2
Block of 4096KB: 3814.1 / 2406.5 / 2416.5
Block of 8192KB: 3814.2 / 2405.5 / 2416.5
Block of 16384KB: 3814.3 / 2407.0 / 2416.4
Block of 32768KB: 3814.4 / 2407.1
Latency - Memory walk tests... ("pointer chasing")
Null size: 2 cycles 0 cycles (overhead 376 cycles)
steps: 4 8 16 32 64 128 256 512 1k 2k 4k (bytes)
Block of 1KB: 1 1 1 1 1 1 1 1 - - - cycles
Block of 2KB: 1 1 1 1 1 1 1 1 1 - - cycles
Block of 4KB: 1 1 1 1 1 1 1 1 1 1 - cycles
Block of 8KB: 1 2 2 4 2 2 1 1 1 1 1 cycles
Block of 16KB: 2 3 8 18 18 18 18 18 18 20 2 cycles
Block of 32KB: 2 3 8 18 18 18 18 18 18 18 19 cycles
Block of 64KB: 2 2 4 10 17 17 17 20 20 20 24 cycles
Block of 128KB: 2 2 4 10 18 18 18 19 20 20 24 cycles
Block of 256KB: 2 2 4 10 17 18 18 21 22 23 24 cycles
Block of 512KB: 4 3 4 12 18 20 21 24 25 32 42 cycles
Block of 1024KB: 9 7 14 28 56 196 196 204 208 216 219 cycles
Block of 2048KB: 9 8 19 28 56 196 196 204 208 215 222 cycles
Block of 4096KB: 8 8 15 27 53 210 212 239 231 237 239 cycles
Block of 8192KB: 9 8 14 25 51 166 167 176 182 191 200 cycles
Block of 16384KB: 8 7 13 25 49 166 167 176 181 192 207 cycles
Block of 32768KB: 11 7 13 25 49 166 168 176 181 193 208 cycles
This system appears to have 2 cache levels (enabled).
L1 cache (8KB) speed (MB/s): Read=22418.9, Write=9983.8
L2 cache (512KB) speed (MB/s): Read=12184.4, Write=10095.5
Main memory speed (MB/s): Read=3814.3, Write=2407.0




300FSB 3:2 ratio, 200Mhz 2-2-2-5 timings, GAT: SR-A-A-D-D, 800PSB Strap (Partial PAT)

Cache size/Memory speed info tool 2.65MMX - (c) 1999-2001, LRMS - DJGPP compiled
CPUID support detected... 'GenuineIntel' with FPU TSC MMX
Family=15 Model=2 Step=9 Type=0 Chipset (Vendor/Device ID(Rev)): Intel/2570(02)
CPU clock: 3602.0 MHz
Using 32MB physical memory block (alignment = 32)
Bandwidth - MMX linear access test... Read/Write/Copy (MB/s)
Block of 1KB: 28364.6 / 11963.2 / 23938.9
Block of 2KB: 27914.9 / 12086.6 / 23380.0
Block of 4KB: 26849.1 / 11956.7 / 23689.2
Block of 8KB: 26481.3 / 12083.3 / 22323.0
Block of 16KB: 14642.4 / 12149.3 / 22670.9
Block of 32KB: 14648.1 / 12181.0 / 22871.9
Block of 64KB: 14620.4 / 12177.0 / 23078.5
Block of 128KB: 14653.3 / 12194.8 / 20770.6
Block of 256KB: 14588.9 / 12090.4 / 21982.3
Block of 512KB: 14068.2 / 11920.8 / 3095.8
Block of 1024KB: 3982.5 / 2394.7 / 2414.9
Block of 2048KB: 3995.0 / 2391.8 / 2414.6
Block of 4096KB: 3999.4 / 2392.1 / 2415.0
Block of 8192KB: 3997.0 / 2392.3 / 2415.0
Block of 16384KB: 3998.0 / 2392.3 / 2414.8
Block of 32768KB: 3998.0 / 2392.4
Latency - Memory walk tests... ("pointer chasing")
Null size: 2 cycles 0 cycles (overhead 376 cycles)
steps: 4 8 16 32 64 128 256 512 1k 2k 4k (bytes)
Block of 1KB: 1 1 1 1 1 1 1 1 - - - cycles
Block of 2KB: 1 1 1 1 1 1 1 1 1 - - cycles
Block of 4KB: 1 1 1 1 1 1 1 1 1 1 - cycles
Block of 8KB: 1 2 2 4 2 2 1 1 1 1 1 cycles
Block of 16KB: 2 3 8 18 18 18 18 18 18 20 2 cycles
Block of 32KB: 2 3 8 18 18 18 18 18 18 18 19 cycles
Block of 64KB: 2 2 4 10 17 17 17 20 20 20 24 cycles
Block of 128KB: 2 2 4 10 18 18 18 18 20 21 24 cycles
Block of 256KB: 2 2 4 10 17 18 18 21 22 23 24 cycles
Block of 512KB: 4 3 5 11 18 21 21 22 25 33 42 cycles
Block of 1024KB: 11 8 15 31 60 288 253 287 286 284 282 cycles
Block of 2048KB: 10 9 16 31 60 238 235 243 250 255 259 cycles
Block of 4096KB: 10 9 17 33 64 237 236 245 248 253 264 cycles
Block of 8192KB: 11 8 15 31 59 199 201 211 217 223 235 cycles
Block of 16384KB: 10 8 15 29 57 202 201 210 216 231 246 cycles
Block of 32768KB: 9 8 15 30 57 202 202 210 218 230 244 cycles
This system appears to have 2 cache levels (enabled).
L1 cache (8KB) speed (MB/s): Read=26849.1, Write=11956.7
L2 cache (512KB) speed (MB/s): Read=14588.9, Write=12090.4
Main memory speed (MB/s): Read=3998.0, Write=2392.3

I cannot boot with 3:2 divider with FULL pat enabled, no matter what FSB or voltage, would of been interesting to see it break 4000MB/s read

Please post your results using the same test method. Booting into dos is the purest form since it skips any windows program influences.

Discuss.

I noticed that too, but I get better performance in 1:1 than I do at 5:4, with both the memory and cpu speeds being equal. (ES chips are so great for testing )

I get about 1000 more MBs in 5:4, but I get about 1000 more 3dmarks in 1:1. My games also play better in 1:1.

Bump, anymore input?

So a possible conclusion could be that memory bandwidth is no longer an issue, at least in highly overclocked Intels? Intel has been known for having huge memory bandwidth.

It shows that there is quite a big of difference Full PAT and Partial PAT... For all you springdale owners, try to get Full PAT working!

What kind of memory do you have?

JDizzle said:

What kind of memory do you have?

Click to expand...

In my sig: 2x 512MB Corsair PC3500C2 Rev 1.1 BH5s