Project SandyBridge-E, Part III – Test Results


Well the build is complete, the system is stable and the initial test results are in!  Waiting for test results is always nerve wracking and this was no exception!  Ok sure these aren’t exactly life altering results, but how worthwhile a new build turns out to be is always a “moment of truth” situation since it represents the value proposition of a multi-thousand dollar investment.

Benchmark Results: SSD – Crystal DiskMark 3.0.1 x64

First up, is disk performance.  Interesting thing here.  Results were actually different before I got the system truly stable and fully functional (higher actually), but since everything wasn’t working quite yet, I’m going to disregard those results and go with these:

At their highest, during the testing and tweaking phase, these drives managed to hit 1GB/s sequential performance (this was with the Intel 3.5.x drivers).  I may go to that rev again at some point, especially since I don’t think any of the issues were really related directly to the Intel C600 RAID drivers, but at this point I don’t want to mess with a configuration that’s working.  For comparison, here are the results from my Alienware M18x R1 which is running 2 x Crucial M4 128GB drives in RAID 0 (so pretty similar setup really – SandyBridge HM67 with i2670QM@2.2Ghz, 16GB RAM, ATI 6990M in X-Fire):

With the exception of large sequential (the least important metric really), clearly a slamdunk for the enthusiast desktop chipset and the larger M4s (as well it should be!).  For comparison the single 512GB M4 on the desktop was pulling about 540MB per second sequential and my old Vertex 2 SATA 2 RAID 0 was pulling about 500MB sequential.  No photos unfortunately and I only have the sequential numbers on hand.

Of course the real story here is that 4K small random result.  25MB/s read and over 100MB/s write is just staggering performance on small random.  Insane really.  Consider that the best 15k SAS drives in RAID 0 struggle to even break into  double digits here (generally 1-2MB/s read and possibly as high as 10MB/s or so on write)  Given how important small random transfers are, this really explains just how powerful of a “seat of the pants” impact SSDs have as a system upgrade.  In these days of diminishing returns on GPU and CPU upgrades,  spinning platter to SSD remains a shining beacon of huge value proposition.

Benchmark Results: 3DMark Vantage

3DMark Vantage, in my opinion, is still a decent way to compare platform deltas.  I had run the 980x tri-sli build through and re-ran the testing again with the 3960x.  980x was at 4.0Ghz and the 3960 was at 4.4Ghz.  I consider this fair since those are the core frequencies both CPUs were able to hit with no additonal tweaks (straight multiplier) on Corsair cooling (H50 for 980, H80 for 3960) and at 100% stability.  I think it is a reasonable component of the CPU vs CPU aspect of the test to factor in how far they can be pushed with as close to no effort as possible without losing stability (even though this can often come down to the luck of the die), but it is worth noting the asymetry here in the interest of fairness.  This is how the scores landed in High Quality mode (1680×1050):

The difference here is largely in the CPU score.  Here are the details:

Graphics Score
3960x result:     51543.98
980x result:       47944.98
CPU Score
3960x result:    45366.26
980x result:      37429.96
We can see there is a roughly 21% difference in CPU score.  Figuring about a 10-15% clock for clock IPC difference between Westmere and SandyBridge, coupled with a 5% difference in actual clock speed, and it puts us right about there.
The difference in GPU score is interesting because it does imply that, at least under 3DMark Vantage conditions, the GPUs can benefit from more CPU.   Worth noting though, that we are only talking about a few percentage point gain in GPU from a greater than 20% gain in CPU performance.  Definitely diminishing returns.  The difference in VRAM (4GB cards for the 3960 testing vs 2GB cards for the 980x testing) is immaterial here since VRAM usage under 3DMark never went over 1100MB.
As a curiosity, it is interesting to compare the mobile SandyBridge platform, I think.  Here is what the Alienware M18x was able to pull off (once again,  HM67 with i2670QM@2.2Ghz, 16GB RAM, ATI 6990M in X-Fire):
Of course no surprise that the desktop platform actually triples the result here (even with the 980x), but 16,000 is actually a mighty impressive score for a mobile platform on High settings.  Looking at these scores it becomes clear how, for single monitor PC gaming, a high end laptop like the Alienware M18x really could be one’s only “gaming PC”.  For 1080P gaming, mobile SLI or XFire and a SandyBridge mobile CPU really provides all of the power you need for 90% of the games out there!  That said, lets see how the systems stacked up with the less demanding “Performance” setting of 3DMark Vantage.
The 10% or so overall advantage the 3960 platform enjoyed with high settings actually widens to a full 15% on performance.  We can see a 7000 point gap now.  Taking a deeper look at the details reveals why:
Graphics Score
3960x result:   61174.95
980x result:     55291.38
CPU Score
3960x result:   45302.56
980x result:     37014.3
CPU score, interestingly, remains identical even as resolution drops.  This is likely a result of the CPU tests being independent and entirely CPU dependent.  We can see that under the performance preset, the 3960 is able to drive disproportionate performance gains from the 680s in comparison to the higher settings.  So the difference is entirely in the 3960s ability to push the 680s in lower resolutions, rather than the pure processing capability of the CPU itself.  And how does the Alienware fare?  Pretty well actually; a Vantage score of nearly 22,000:
Benchmark Results: Metro 2033 Surround/3D Surround
There is no doubt about it, Metro 2033 continues to be a beast.  Whether that is the result of advanced engine technology, or simply sloppy unoptimized code, is open for debate, but either way it can still bring even the mightiest systems to their knees.  Armed now with 3960X at 4.4Ghz, 3 x GTX680 in SLI at 1.2Ghz, 16GB of 1866Mhz quad channel DDR3 and SATA 3 RAID 0 SSD, it was once again time to do battle!  And the results are in… First, the worst case: 3D Surround, 5040×1050, Very High Quality, DirectX 11, PhysX on (DOF still off since it’s pointless in 3D).  Here are the gories:
The settings… Brutal stuff and in 3D no less! 120Hz won’t be pretty!

The outcome.  Not horrible actually, but nowhere near what I would consider truly playable.  Just look at the lows.  Not pretty:

sub 10fps mins and just over 30fps average… yikes!

So mega fidelity, DirectX 11, 3D surround Metro 2033 still isn’t quite here and possibly never will be (I do believe the problem is an inefficient engine).  That said, it did break 30fps average and frontline is a pretty brutal run through.  Also, for comparison, the 980x had managed a 28.5 so the 3960x has given us a few extra fps, but at a clock speed that is 10% higher, so any IPC gains are really lost here and it is clearly diminishing returns.  It actually is possible to play this way and this was absolute worst case as can be seen from the settings.  Of course this isn’t 3 x 1080P.  For that you’re going to absolutely need 4 x SLI to even maintain these levels and some serious OC’s to get closer to truly smooth framerates.  OK so the max of the max is out of reach.  Lets see how things fare when we relax the settings a bit!

Settings on high, PhysX enabled

Much better results.  Lows are still dismal and can’t break 10, but the average has now hit 40 and “high” still looks quite good.  Interesting to note that disabling PhysX didn’t do much at all:

Gain of only 1fps… PhysX is “free” at this level… Leave it on

OK, so 3D is a killer and 120Hz (60fps in 3D) can’t be sustained clearly.  Just how high can the average get then?  Let’s take a look at surround only (no 3D), setttings on high, PhysX enabled:

This tells the tale.  With 3D off average is, not surprisingly, only up to 71.  So we can see with the latest drivers you actually don’t even take a 50% framerate hit any more (good work to NVidia on this!), but the tri-SLI is just well short of the 100fps + you would need to see really playable framerates in 3D.  The lows are still dismal at sub 10fps which, in my opinion, again points to either some hitching condition somewhere or horrible inefficiency in this engine.  A spread of 8-200fps in a canned benchmark provided by the developer isn’t good with this level of hardware.

Next lets see how things fare on “Very High” settings.  DOF still off:

Unfortunately “very high” is enough to take is below 60fps with the average coming in at 54fps and the low sticking to sub 10 thanks to whatever game engine condition causes that total dropout (it isn’t VRAM, that never gets above 2GB and these are now 4GB cards).  Interesting to note that turning on DOF kills the framerate, dropping the average to 35fps without 3D and actually causes visual artifacting, not to mention seems to add nothing in terms of enhancing visuals.  Really strange that DOF has a steeper penalty than 3D Vision and another indictment of this engine, if you ask me.

For fun and reference, here is what the M18x 6990M XFire managed at 1080p (no 3D obviously):

Pushing right up against 40fps; pretty good for a laptop!  Of course it does weigh like 10+ lbs 🙂

Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s