Jump to content

Question on cache and memory benchmarks!


Recommended Posts

0eb30f2442a7d9335c2885ecac4bd11372f001a8

 

( The picture of testing result can also be seen here:http://imgsrc.baidu.com/forum/w%3D580/sign=5d8024603ac79f3d8fe1e4388aa0cdbc/0eb30f2442a7d9335c2885ecac4bd11372f001a8.jpg )

 

This is the testing result of my 4-way AMD Opteron 6147 machine, but some people keep telling me this result is impossible, which makes me a little doubtful now!

 

Could this be an error of Aida64, or it’s just a fact that some people refuse to believe?

 

Looking forward to your reply!

 

Best regards!

Link to comment
Share on other sites

1) Do you have 2 or 4 processor sockets populated?

 

2) Do all processors have DDR3-1333 memory installed?  So every processors use the same memory configuration?

 

In case you have 4 processor sockets populated, and every processors use the same DDR3-1333 memory configuration, that would mean you've got 16 memory channels total, each running at DDR3-1333 speed, which would mean a total memory bandwidth of 166.6 GBytes/sec.  Hence the results could well be accurate ;)

 

The confusion may come from the fact that AIDA64 (the main application itself) only supports up to 32 processor threads, so it cannot show you all processors on the Cache & Memory Benchmark Panel.  However, the multi-threaded memory benchmarks support up to 64 processor threads, so it can utilize all 48 processing threads you've got in your system (if you have all 4 processor sockets populated, each with a 12-core Magny-Cours CPU).

 


Regards,

Fiery

  • Like 1
Link to comment
Share on other sites

1) Do you have 2 or 4 processor sockets populated?

 

2) Do all processors have DDR3-1333 memory installed?  So every processors use the same memory configuration?

 

In case you have 4 processor sockets populated, and every processors use the same DDR3-1333 memory configuration, that would mean you've got 16 memory channels total, each running at DDR3-1333 speed, which would mean a total memory bandwidth of 166.6 GBytes/sec.  Hence the results could well be accurate ;)

 

The confusion may come from the fact that AIDA64 (the main application itself) only supports up to 32 processor threads, so it cannot show you all processors on the Cache & Memory Benchmark Panel.  However, the multi-threaded memory benchmarks support up to 64 processor threads, so it can utilize all 48 processing threads you've got in your system (if you have all 4 processor sockets populated, each with a 12-core Magny-Cours CPU).

 

Regards,

Fiery

 

I became not sure about my knowledge of Opteron's memory controller after I read your reply, so I checked it in AMD's official PDF about Opteron. I got this from the PDF of AMD.

 

• Package G34
• 144-bit DDR3 SDRAM controller operating at frequencies up to 1333 MT/s (667 MHz)
• Supports up to twelve (12) registered DIMMs
• Supports up to eight (8) unbuffered DIMMs
• On-line spare feature provides single-rank DRAM redundancy
 
I think from this I can make sure that we can only get two memory channels. If so, there should only be 8 memory channels in a four-way Opteron server.
Link to comment
Share on other sites

It's two memory channels per processor node.  A 12-core Magny-Cours processor has 2 nodes, so a single processor package supports 4 memory channels.  It effectively works as two 6-core processors (each supporting 2 memory channels) put into a single CPU package.

Link to comment
Share on other sites

  • 2 weeks later...

We've just extended the 32 threads benchmarking capabilities of AIDA64 to 640 threads, and added processor groups support as well.  Make sure to upgrade to the latest beta version of AIDA64 Extreme Edition available at:

 

http://www.aida64.com/downloads/aida64extremebuild2546y3mbp1kvdjzip

 

After upgrading to this new version, make sure to restart Windows to finalize the upgrade.

 

Using this release you can use all CPU, FPU and memory benchmarks of AIDA64 to measure the full 48-core performance of your configuration.

  • Like 1
Link to comment
Share on other sites

  • 1 month later...

It looks like it may be due to a BIOS issue.  We've checked both AIDA64 versions on our Intel DX79SI test system with Core i7-3960X CPU, but haven't found any difference in the scores obtained.  We did update the benchmark module between AIDA64 v3.00.2500 (stable release) and AIDA64 v3.00.2578 (latest beta release), but it only affects Ivy Bridge-E/EN/EP processors.

 

Since your memory latency scores seem unchanged, but the memory bandwidth scores seem to drop to half, it looks like if the memory controller was configured for dual-channel operation instead of quad-channel operation.  Although the memory controller (IMC) reports quad-channel mode, as AIDA64 reports it.

Link to comment
Share on other sites

  • 6 years later...
On 7/20/2013 at 3:47 AM, Fiery said:

We've just extended the 32 threads benchmarking capabilities of AIDA64 to 640 threads, and added processor groups support as well.  Make sure to upgrade to the latest beta version of AIDA64 Extreme Edition available at:

 

http://www.aida64.com/downloads/aida64extremebuild2546y3mbp1kvdjzip

 

After upgrading to this new version, make sure to restart Windows to finalize the upgrade.

 

Using this release you can use all CPU, FPU and memory benchmarks of AIDA64 to measure the full 48-core performance of your configuration.

Thanks for your reply and the update!

Link to comment
Share on other sites

Hi, Fiery!

I have purchased a new SuperServer 8049U-E1CR4T and run memory benchmarks. The read bandwidth is very close to the theoretical value, while the write bandwidth is only two-thirds of the theoretical value, could this be a problem due to hardware malfunction? The snap of the result is attached below.

PS, my SuperServer 8049U-E1CR4T has all 4 processor sockets populated, each configured with 3 memory modules.

Many thanks!

 

aida64测试结果.JPG

Link to comment
Share on other sites

On ‎1‎/‎3‎/‎2020 at 4:37 PM, Glowing1196 said:

Hi, Fiery!

I have purchased a new SuperServer 8049U-E1CR4T and run memory benchmarks. The read bandwidth is very close to the theoretical value, while the write bandwidth is only two-thirds of the theoretical value, could this be a problem due to hardware malfunction? The snap of the result is attached below.

PS, my SuperServer 8049U-E1CR4T has all 4 processor sockets populated, each configured with 3 memory modules.

Many thanks!

 

aida64测试结果.JPG

That ratio between the read and write memory bandwidth is normal on the type of system you've got.

Link to comment
Share on other sites

8 hours ago, Fiery said:

That ratio between the read and write memory bandwidth is normal on the type of system you've got.

Thanks for your reply!

The CPU itself does support 6 memory channels and now only 3 memory modules are installed, meaning a large margin between the installed memory bandwidth and the IMC bandwidth. At such case, I deem the write bandwidth should be close to the read bandwidth.

Actually, I have asked the same question to the Kingston’s guy and the Supermicro’s guy, they all suggest me to replace the memory module to the one authenticated by Supermicro due to the difference in the design of the register of the memory module.

Link to comment
Share on other sites

On 1/6/2020 at 6:33 PM, Glowing1196 said:

Thanks for your reply!

The CPU itself does support 6 memory channels and now only 3 memory modules are installed, meaning a large margin between the installed memory bandwidth and the IMC bandwidth. At such case, I deem the write bandwidth should be close to the read bandwidth.

Actually, I have asked the same question to the Kingston’s guy and the Supermicro’s guy, they all suggest me to replace the memory module to the one authenticated by Supermicro due to the difference in the design of the register of the memory module.

Well.. It's quite difficult to double-check your theory due to the lack of multi-threaded memory bandwidth benchmark software for Windows :)  But either way, it would be best to populate all available memory channels for optimal performance.

Link to comment
Share on other sites

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.

Loading...
 Share

×
×
  • Create New...