AMD llano cpumark artificially inflated scores

[pman6]

Phenom II X6 1090T vs A6-3650

which do you think would be better, both having similar cpu architecture ?

The phenom x6 has 6 cores, 3.2ghz, turbo to 3.6ghz, and L3 cache.
the A6 has 4 cores, 2.6ghz, no turbo, no L3 cache.
(the A6 is priced at $115, and the X6 at $180)
The phenom is a performance cpu. The A6 is just a budget cpu.

Yet, at a glance, the similar cpumark scores would lead the unsuspecting buyer to think that cpu performance of the cheaper, weaker, clearly inferior A6-3650....... is on par with the hexacore 1090T. But it's not!

Something is wonky with the cpumark test, or AMD found a way to cheat it. ALL the AMD llano chips lately have had inflated cpumark scores that belie their actual performance.
AMD didn't just magically increase their IPC, especially using the same K10 architecture.

There needs to be a breakdown of the score. Not just simply a score.

Shoppers are getting misled into thinking that AMD cpu performance meets or exceeds intel's, when in reality, it is inferior in almost all tests, especially multithreaded ones like cinebench and x264 video encoding. amd's IPC per thread is way behind intel's.
Yet, a slower clocked llano gets roughly the same cpumark score as Intel's Core sandy bridge chip ?

A6-3650 cpumark score below
http://www.techwarelabs.com/amd-a6-3...u-processor/5/

[pman6]

http://img594.imageshack.us/img594/5145/passmarkcomp.png

I discovered something odd when I made a portable version of Passmark 7, and brought it to Best Buy electronics store to compare Sandy Bridge and Llano mobile cpu display models.

refer to picture link above.

the weird things-
1. at the top, it says version 7.66 for some reason. not 7.0
2. my Llano cpumark scores were vastly different than the average scores posted on cpubenchmark.net, but they were realistic and believable, and correspond to their actual performance compared to Intel.

Let me explain #2 further. All the laptops tested were configured exactly the same way, with exactly the same software on them. Control panel power setting set to High performance.

According to cpubenchmark.net, the cheap A4-3300m dual core gets 2412 average score.
My cpumark test, with 2 threads, shows only 1334 score.
A huge score difference.

I didn't do anything wrong either, because the dual core Intel B940 score is in line with the average score shown on cpubenchmark.net.

The dual core celeron p4600 scored 1336 in my test, similar to average score on cpubenchmark.net

The reality is that A4-3300m dual core is as slow as the P4600 in real world and dual-thread tests under 100% cpu load.
The A4 should score the same as the P4600, like my test showed.

If one were to look at the 2412 score of the A4, he would think it has similar performance to the Core i3-2310m, which scores 2500.
The fact is, the A4-3300m is much much much inferior to the Core i3.

Very misleading Llano scores posted on cpubenchmark.net

the A4-3300m vs p4600 vs B940 was an apples to apples comparison- All dual core, no hyperthreading.

[David (PassMark)]

Phenom II X6 1090T vs A6-3650
The phenom is a performance cpu. The A6 is just a budget cpu.
Yet, at a glance, the similar cpumark scores...

The Phenom II X6 1090T, is currently ranked as the 63th fastest CPU with a CPUMark of 6050.

The A6-3650, is currently ranked as the 99th fastest CPU with a CPUMark of 4844.

This is a fairly big 25% difference. Not really all that similar at all.

Shoppers are getting misled into thinking that AMD cpu performance meets or exceeds intel's

Not sure how you can compare the AMD 1090T and AMD A6-3650 then come out with a broad conclusion that they are better than all of Intel's offerings.

According to the charts Intel has 85 different CPU models already on the market that are faster than AMD A6-3650.

According to cpubenchmark.net, the cheap A4-3300m dual core gets 2412 average score.
My cpumark test, with 2 threads, shows only 1334 score.
A huge score difference.

What O/S where you running this on (32bit or 64bit)?
And what version of PerformanceTest were you running (32bit or 64bit)?

All the results in the charts for the A4-3300m are on Win7 64bit. But I would expect this CPU to do far worse if you used 32bit software, OR were running a 32bit O/S.

I think this is likely to explain the difference. The Integer maths test is very sensitive to 32/64bit. And even the Intel CPUs in your screen shot are down for this test.

So I think you are looking at what you get running 32bit software, rather than 64bit software on 64bit hardware.

at the top, it says version 7.66 for some reason. not 7.0

I am guessing you are running an old release. Maybe even an old hacked 32bit release?

Would be good if you can get the latest 64bit release and try again.

[David (PassMark)]

Also these posts might also be interesting.

Differences in 32bit and 64bit performance.
http://www.passmark.com/forum/showthread.php?t=2244
http://www.passmark.com/forum/showthread.php?t=261

PassMark Rankings of i3 and AMD A8-3850
http://www.passmark.com/forum/showthread.php?t=3209

Update Oct 2011:
See this post for a lot more details about what is going on in the Integer maths test.
http://www.passmark.com/forum/showthread.php?t=3348

Update Mar 2012:
See this post for details about how a CPU bug might be impacting the results.
http://www.passmark.com/forum/showthread.php?t=3656

[pman6]

my mistake on the A6-3650 cpumark score.

I didn't check the average score on that, and only used that website's 6000+ score.


I will install the latest passmark and see if there is any difference

[pman6]

ok. It turns out that I was using the 32bit version on windows x64 on all the laptops tested.

however, doesn't that still make each test equally crippled?
So it should still tell me the performance of the 6 cpu's relative to one another?

that link you posted in the other thread is exactly the issue I'm talking about---
My only issue is that the Llano cpumark scores are skewed, and don't tell the true story.

http://www.xtremehardware.it/eng-reviews/eng-reviews/amd-llano-a8-3850,-computational-and-graphics-performance-of-the-desktop-apu-201106305400/8/

The 1 or 2 tests the Llano excels in seem to give it a disproportionate number of cpumark points.
So, just by looking at cpumark score, the consumer might think the Llano A8 beats the Phenom x6, but as the xtremehardware website shows, the A8 cpu is actually among the weakest of the bunch.

The same thing can be said when comparing the Llano A4-3300m with the Core i3-2310m. Both cpumark scores are similar, but in real world testing, the core i3 absolutely demolishes the A4-3300, whether you're using 32 or 64bit programs.

The A4-3300m actually only has the performance level of an Athlon II p340, which scores 1300.
But to those who have never played with the A4 cpu, the A4's 2400 score makes it appear to have double the performance of the p340, when they truly have just equal performance.

We think some of the other published benchmarks show the AMD chip as being a bit slower because,
1) They aren't running on all cores
2) They are running 32bit code on 64bit hardware. Where as the AMD chips can be up to 6 times faster running 64bit native code in simple cases.

From what I can see, a number of tests fully utilize all cores, like cinebench and x264, and have 64bit code for 64bit hardware.

[David (PassMark)]

however, doesn't that still make each test equally crippled?

No. It doesn't. And that's also why I think many other public benchmarks are maybe getting it wrong. They are for the most part only looking at 32bit code. There is no mention in the benchmarks I have looked at that examines this 32bit / 64bit code behavior.

Look at this chart below. It is for the P4600 and A8-3850. But the same result can be seen for other AMD & Intel CPUs.

Intel just isn't doing as well in the move from 32bit to 64bit as AMD does. But our benchmark charts for the AMD chips are almost entirely 64bit results.

As you pointed out above our results match other public results if you only look at 32bit code.

AMD looks significantly better in pure 64bit code that fully loads the CPU.

[pman6]

ok, all that 64bit stuff is understandable.

But going back to the phenom x6 versus A8-3850,
it would appear that the bulk of the cpumark points are given to finding prime numbers, and integer math.... as the A8 only excels in these 2 cpumark tests, yet it has a higher cpumark score than phenom.

Some people say that finding prime numbers is not a good way to judge a cpu anymore, especially with multicore cpus. And how important is integer math in the real world ? what is your opinion about these 2 things?

How would you reconcile the difference between the cpumark score and actual performance of the A8 vs phenomx6, seeing that the A8 scores similar or higher, but is a weaker cpu overall ?

[David (PassMark)]

...as the A8 only excels in these 2 cpumark tests, yet it has a higher cpumark score than phenom...

Actually in our charts it doesn't. At the moment the
AMD Phenom II X6 1100T with a CPUMark score of 6281
AMD A8-3850 with a CPUMark score of 6218

But they are pretty close. So it would depend on the real life work load and number of threads as to which one performed better. I don't think you can generalize and say the 1100T is better at everything. For example in a single threaded 64bit integer work load then A8 should be significantly faster than then 1100T.

As to the question, is finding primes and integer maths relevant to real life. I would make 3 points,
1) They are only 2 of the 8 CPU tests in the suite.
2) The integer maths test makes use of subtraction, multiplication and division using integer variables. To claim that this isn't used in real life is wrong.
3) The prime number test is more artificial, in that the speed depends on how fast the CPU can complete a very tight loop, perform a modulo instruction & check some branch conditions. Real life code tends to be more complex. But it is a classic benchmark that people almost expect to see.

[nemo14]

Dears, weired thing happened to me. I run a CPU Mark test using A6-3650 and I never get over the 3400 threshold, despite of 5376 average value. Comparing different baselines I noted my system fails on Integer Maths and Find Prime Numbers (see picture below, red squared is my system)



and never reach 2.600 GHZ (it stops at 2.595.7, see picture below)



I tested teh system with W7 64 bits.
Someone can address me to solve this issue?
thx

[David (PassMark)]

My guess is that you have the same issue that was noted above.
That is to say, you are probably running the 32bit benchmark on your 64bit system.

I can also see you are using V7.0 build 1007. Which is pretty old now.

[nemo14]

My guess is that you have the same issue that was noted above.
That is to say, you are probably running the 32bit benchmark on your 64bit system.

I can also see you are using V7.0 build 1007. Which is pretty old now.

I wish you're right but unfortunately I used PerfomanceTest 64 bit build 1024
Since i have 2 samples, I test the second one getting the same results. I've also installed W7 SP1 but nothing has changed. Below the new picture to show results.

Any idea?

[David (PassMark)]

You might want to check the CPU temp to see if it is overheating.

[nemo14]

32°C, defintely low

[David (PassMark)]

32°C, defintely low

This must be at idle? Wouldn't hurt to check under high load.

Is your RAM in the right slots for dual channel?