AMD Dual 6272 performance is 60% lower than benchmarks

[reiddar]

Hi,

I've ran many of your tests on various machines and this is the first time I've had results so far off your benchmarks.

System has:

Dual AMD 6272
MB is H8DDGI-F
64 GB of Kingston 1333mhz RAM
Dual OCZ SSD drives

I've read your causes for poor performance and I don't meet any of those.
This is a brand new custom build and it is not overheating.

The CPU score running 32 processes was between 9500 - 10,200
removing one chip and running on 16, it was 8500.
Changed the chip and still around 8500.

Is is possible the passmark software is not compatible with my setup?
More likely is it a bad MB?

The RAM scores were off also.

For Single thread, the dual chips scored an 850.

The 64gb of ram scored 806 overall

Thank you

[reiddar]

MB is H8DGI-F not H8DDGI-F

[David (PassMark)]

Multi-CPU systems are rare compared to single CPU systems. They tend to use rare and expensive CPUs as well. So we don't get as many results come in.

So to keep the charts comprehensive as possible during the change over from PerformanceTest V7 to V8 we averaged all available Multi-CPU results to make up the Multi-CPU chart.

The benchmark results returned by V7 and V8 can be a bit different for however. For some CPUs, there is a small difference, for other CPUs the difference is larger. Some AMD chips had a significant drop. Especially those CPUs with large numbers of relatively slow cores.

You can find a discussion here,
http://www.passmark.com/forum/showth...-go-live-today
and here
http://www.passmark.com/forum/showth...Please-Explain

But a couple of months have now passed & more PT8 results have come in.
So we decided today that we'll remove the PT7 results from the chart and archive them off to a separate static graph for posterity.

This will reduce the number of different types in the chart, but all results will be from PT8. This will be done in the next couple of days.

In PT7 a dual CPU Opteron 6272 could expect to get around 18,000 on the CPUMark.
In PT8 a dual CPU Opteron 6272 could expect to get around 10,000 on the CPUMark.

So in fact I think your system in getting the expected results, or fairly near expected results.

[reiddar]

Multi-CPU systems are rare compared to single CPU systems. They tend to use rare and expensive CPUs as well. So we don't get as many results come in.

So to keep the charts comprehensive as possible during the change over from PerformanceTest V7 to V8 we averaged all available Multi-CPU results to make up the Multi-CPU chart.

The benchmark results returned by V7 and V8 can be a bit different for however. For some CPUs, there is a small difference, for other CPUs the difference is larger. Some AMD chips had a significant drop. Especially those CPUs with large numbers of relatively slow cores.

You can find a discussion here,
http://www.passmark.com/forum/showth...-go-live-today
and here
http://www.passmark.com/forum/showth...Please-Explain

But a couple of months have now passed & more PT8 results have come in.
So we decided today that we'll remove the PT7 results from the chart and archive them off to a separate static graph for posterity.

This will reduce the number of different types in the chart, but all results will be from PT8. This will be done in the next couple of days.

In PT7 a dual CPU Opteron 6272 could expect to get around 18,000 on the CPUMark.
In PT8 a dual CPU Opteron 6272 could expect to get around 10,000 on the CPUMark.

So in fact I think your system in getting the expected results, or fairly near expected results.

That is terrible news. I decided on a dual 6272 based on your benchmarks and compare to other chips.

On the new version, one 6272 was scoring 8500-9000 and a dual chip, 10,000.

Does that even make any sense?

What about my low memory score? Is that also normal?

It scored an 806 with 64 gb of 1333MHZ DDR3 RAM, latency was 128ms, DB operations of 23.8, read uncached of 3155, etc

Can you post a link for PT7 to make sure the CPU is at least benchmarking how it used to.

Thank you

[David (PassMark)]

You can find old versions of PT at the bottom of this page,
http://www.passmark.com/download/pt_download.htm

PT7 and PT8 are testing different things. For example PT8 includes a single threaded test. So the change in the score doesn't mean the CPU is now slower than it was before. It would still be a reasonable CPU for situations where you have dozens of concurrent independent tasks running. (e.g. a web server). It would be a rather poor CPU for running one single threaded application however.

You might also be able to tweak / configure the machine for better performance.

128ns is pretty awful latency for DDR3 RAM. Some desktop machines are 6 times faster. Is this ECC / registered RAM? In the worst case ECC can nearly half the performance. This MB supports Quad channel RAM, are you running in quad channel in BIOS?

[David (PassMark)]

New multi-cpu chart is now up.

The old V7 results have been moved to a separate static chart.

All the results (both Intel and AMD) have come down a bit.

The dual CPU Opteron 6272 went from 19th position to 21st position. So not as dramatic a change as the changed CPUMark value might lead you to think.

[reiddar]

New multi-cpu chart is now up.

The old V7 results have been moved to a separate static chart.

All the results (both Intel and AMD) have come down a bit.

The dual CPU Opteron 6272 went from 19th position to 21st position. So not as dramatic a change as the changed CPUMark value might lead you to think.

The RAM is Kingston KVR1333D#D8R9S/4G 4gb sticks DDR3-1333 CL9 DIMM ECC

So the benchmarks for dual CPU's are benchmarked against each other and not against a single CPU?

I sent this server back thinking that if the benchmarks were 16,000 and I was getting 10,000, that something must be off.

Although it does seem like the RAM is scoring quite poorly?

At any rate, I think I will to overclock the CPU to 3.0ghz as it runs terribly at present.

So to be clear, the CPU scores on a single CPU do not relate to dual CPU?

ie if an i7 scores a 10,000, a dual version will not score ~20,000. It will be benchmarked against other dual cpus and receive a score for dual cpus?

[David (PassMark)]

The single and dual (and quad) CPU scores are comparable. Assuming all the results come from the same version of PerformanceTest.

But doubling the number of CPUs (or cores) doesn't double performance. In just about all real world applications and in our benchmark, there is a non linear relationship between number or cores and performance. There is always some other bottle neck that comes into play as you increase the core count. It might be RAM, bus speed, shared cache, network capacity, disk speed that prevent linear scaling. Meaning that all those CPU cores might be spending most of their time waiting for your high latency RAM to respond.

In some case, with single threaded applications, there can be zero scaling. So adding a second or third core (or CPU) does nothing to improve performance. In this case you are better off without the extra CPU.

This type of expensive high end server typically gets dedicated to one task. So the best benchmark is to use the actual application you intend to run on the machine. e.g. if this is going to be a mySQL database server, then run the actual mySQL application on the machine and measure the speed.

[hawks]

Amd ram scores dont come close to intel. My ram is one the faster one for amd (on passmark test) and still only get around 1500 almost any intel blows it away G. Skill F3-2133c9q-32GXH thats 4 8GB chips

[rvborgh]

reiddar,

not sure if you'll receive an email notification but here is the fix:

http://www.passmark.com/forum/showth...Opteron-6180SE

Both of our boards are from the same family of SuperMicro Opteron server boards... so this should take care of your problem as well.

The memory mark being low is probably due to the the Passmark memory benches not being NUMA aware. So you have to run the system in with node interleaving enabled to present memory in a uniform fashion to the benchmark.

If you want a benchmark that i've used for years for measuring memory throughput. Try Stream... make sure you run it with thread affinity... you should see on the order of 46 GB/s.

Hope this helps

Hi,

I've ran many of your tests on various machines and this is the first time I've had results so far off your benchmarks.

System has:

Dual AMD 6272
MB is H8DDGI-F
64 GB of Kingston 1333mhz RAM
Dual OCZ SSD drives

I've read your causes for poor performance and I don't meet any of those.
This is a brand new custom build and it is not overheating.

The CPU score running 32 processes was between 9500 - 10,200
removing one chip and running on 16, it was 8500.
Changed the chip and still around 8500.

Is is possible the passmark software is not compatible with my setup?
More likely is it a bad MB?

The RAM scores were off also.

For Single thread, the dual chips scored an 850.

The 64gb of ram scored 806 overall

Thank you