We’ve seen how far I can push a six-core Haswell-E in our full-sized performance build, but can the same performance level fit into a mini cube?
System Builder Marathon Q2 2015
Here are links to each of the five articles in this quarter’s System Builder Marathon (we’ll update them as each story is published). And remember, these systems are all being given away at the end of the marathon.
To enter the giveaway, please fill out this SurveyGizmo form, and be sure to read the complete rules before entering!
- $1600 Performance PC
- $1600 Mini Performance PC
- $1600 Gaming PC
- $1600 Mini Gaming PC
- System Value Compared
$1600 Mini Performance PC
Did you think I was going to say something about the packaging of explosives? Or perhaps your most burning question concerned my decision to pack a Haswell-E CPU and high-end graphics card into a mini-cube half the size of a jerry-can? The story starts with two requests from readers, first to include Haswell-E in my general performance build, and second to feature Mini ITX machines in a System Builder Marathon. My first choice was to go big with the Core i7-5820K in our traditional 3-way build-off, but then we temporarily lost a competitor to his day-job. Now down to two builders, we decided that we could follow both paths by combining traditional ATX and mini ITX builds in a single quarter. Each of the remaining builders would first build a big system, and then try to match it with compact hardware. But that doesn’t exactly explain the mini-ITX Haswell-E experiment, does it?
After building a Core i7-5820k alternative PC at the end of last-quarter’s SBM, I told readers I’d stick with that formula on this month’s big system. Replicating the performance of the big machine in a Mini ITX PC meant using the only X99 motherboard in that form factor, ASRock’s X99E-ITX/ac. I’d still need to follow budget restrictions, even though it cost $60 more than the board chosen for the big system. My zeal to meet all of my commitments while catering to the compact PC faithful left me few case/cooling/power choices, yet I’m confident in my abilities as a builder. So confident, in fact, that I chose a case roughly 2/3 the size of my competitor’s LGA-1150 machine!
- Platform Cost: $1,376
- Total Hardware Cost: $1,496
- Complete System Price: $1,596
I know that a few of our big PC builders are looking at this thing and thinking “that’ll never work”, but before you close out of this article please let me explain my theory of why it will. Am I over-confident or simply competent? Even if I fail, there’s nothing better than a car crash to keep your eyes glued to the screen, right?
I’ve fully described my selection of parts in Monday’s $1600 Performance Build, which included the previously-promised Core i7-5820K six-core CPU, externally-exhausted PNY GeForce GTX 970, value-winning Samsung 850 Evo 250GB SSD, and reader’s-choice WD Blue 1TB storage drive. There’s no better way to match the performance of these parts than to continue using them.
Now it’s time to face the strange!
Putting all of these components inside the DIYPC HTPC Cube was probably just a little trickier than rocking a rhyme that’s right on time, but I’ll leave the later to Joe, Jay and Darryl (see boss, I told you I could relate to a younger generation of builders). The first problem would be getting a CPU cooler that wasn’t designed for this motherboard to fit this motherboard.
The opening of Corsair’s bracket is nearly identical at the sides to the Cooler Master compatible bracket included with the motherboard, but the corners of the new bracket are slightly more-recessed. Corsair’s bracket is designed to push upon metal loading points at the corners of the pump, but I’m fairly certain the flatter ASRock bracket will suffice.
Unfortunately, ASRock’s bracket is 3mm too tall for the Corsair pump. Holes in ASRock’s bracket also don’t align with the loading points of the Corsair pump, so any screws used as spacer there would probably put too much pressure on the outer edge of the pump’s shell.
An S-bend in the tabs of ASRock’s bracket could have allowed it to work with Corsair’s standoffs, but that would require skills that some builders may not have. Hoping to make this repeatable by everyone, I ditched Corsair’s standoffs and bought a set of M4 x12mm coarse-thread (0.70mm pitch) screws at the local hardware store. Even at 100 times the bulk price, these four screws cost less than a dollar.
Getting rid of the Corsair standoffs because they’re too tall means losing the “stop point” for tightening the screws. I could have made 7mm-tall spacers from a piece of tubing, but instead decided to go “by feel”. The shown amount of bend should provide adequate pressure between the cooler’s water block and CPU’s heat spreader, but there is a “trick” for those who can’t feel as much: Go slowly so you don’t punch a hole in the motherboard, and if you make gentle contact with the board, back off at least one full turn.
The hard drive fits into a manufacturer-installed tray, inside the case’s top panel. Notches on the edge of the top panel provide side access for drive screws.
The SSD fits onto one of two bottom mounts. DIYPC doesn’t include shoulder screws, so builders are forced to use their best judgement concerning how far they should crush the included rubber grommets.
The power supply fits in the “wrong way”, with its intake hole facing the front panel. That’s due to a cable space issue on its other end. DIYPC provides around ¼” inch of space between the front panel and power supply to make this functional, if less than ideal.
The power cable’s relief section was also trimmed to stop it from pushing against the power supply’s external switch. This can be done with a knife, unless you’re clumsy, wherein you should hand the knife to your dexterous friend.
If you think the cables appear to be shoved into the case, that’s because they are. If you think the power supply is facing the wrong way, that’s because it’s the only way it would fit with the standard-length GTX 970 in place (shorter cards weren’t available yet at or near its price). If on the other hand you think this means the CPU, GPU or power supply will overheat, you’re mistaken. The space between the front panel and power supply inlet did keep the unit from getting hot enough to stink under extended full load, and the radiator fan (attached as a rear-panel intake) is powerful enough to feed both the graphics card and power supply. You’ll see those temperatures in the test!
ASRocks’ cable kit presented another small problem, in that it contained one straight cable and one right-angle cable. I found a place to connect the right-angle end two photos above, right next to the USB 3.0 header. Meanwhile, the case leaves enough space between the motherboard tray and power supply to install a straight cable in the two-port connector’s bottom row.
After spending a full day on this build, I was relieved when I finally turned it on and found that the power LED was connected at the right polarity.
I know all the big system builders were expecting a huge overclocking failure from such a tiny machine, but the proof of its capabilities are in the numbers: Configured as an intake fan, the Corsair H60 flowed enough air to support this CPU at 1.200V and full load.
Unlike my big machine, the CPU in this one is friendly-enough to run a fixed 4.30 GHz at 1.20V, compared to the ATX system’s thermally-restricted variable frequency of 4.0 to 4.3 GHz at 1.22 volts. The smaller systems’ lower temperature is partly due to its lower voltage, and partly due to the larger machine’s underperforming CPU cooler.
Stuffed into the bottom of the case and twice the density of the model used in the big system, the G.Skill DDR4-2400 used in this build didn’t overclock very easily. Perhaps low airflow at the bottom was part of the problem, but the memory didn’t overclock any better at 1.35V than it did at 1.25V. Eventually I settled for DDR4-2666 CAS 15, which was the starting point for the big machine. Worse still was that the smaller motherboard supports only two DIMMs for dual-channel mode, where my ATX PC used a quad-channel kit appropriately.
One might have expected the graphics card to overclock worse in the small machine, due to the tight confines surrounding its top-mounted orientation. Yet we were surprised once again as the GPU and graphics RAM each clocked 50 MHz higher on the smaller machine, which in turn caused it to finally reach the point of thermal restriction, which in turn forced me to go into advanced fan settings and lower the 100% fan level from 90° to 80 °Celsius.
Better overclocking due to a better CPU cooler, better CPU sample, better graphics card sample and, almost-unbelievably, adequate airflow.
Here’s how the Q2 mini PC compares to its full-ATX sibling and my previous-quarter’s graphics-heavy build. Notice that both the Q2 ATX system and the Q1 SLI machine reached the same GPU overclock, which hardly seems like a coincidence, and that the Q1 is only handicapped by two factors: A 4-core CPU more appropriately picked for gaming, and a higher price that hurts it in the value comparison.
|Q2 $1600 Mini Performance PC||Q2 $1600 Performance PC||Q1 $1750 Performance PC|
|Intel Core i7-5820K: 3.30 GHz -|
3.60 GHz, Four Physical Cores
O/C to 4.3GHz, 1.20V
|Intel Core i7-5820K: 3.30 GHz -|
3.60 GHz, Four Physical Cores
O/C to 4.0-4.3GHz, 1.22V
|Intel Core i7-4790K: 4.00 GHz -|
4.40 GHz, Four Physical Cores
O/C to 4.60-4.80 GHz, +20mV
|PNY GTX 970: <1178 MHz GPU, GDDR5-7012 O/C to <1378 MHz, GDDR5-7512||PNY GTX 970: <1178 MHz GPU, GDDR5-7012 O/C to <1328 MHz, GDDR5-7412||2x PNY GTX 970: <1178 MHz GPU, GDDR5-7012 O/C to <1328 MHz, GDDR5-7312|
|16GB G.Skill DDR4-2400 CAS 15-15-15-35, O/C to DDR4-2666 CL 15-15-15-35, 1.325V||16GB G.Skill DDR4-2666 CAS 15-15-15-35, O/C to DDR4-3200 CL 16-18-18-36, 1.30V||16GB G.Skill DDR3-1866 CAS 10-11-10-28, O/C to DDR3-2133 CL 11-12-11-24, 1.60V|
LGA 2011-v3, Intel X99
Stock 100 MHz BCLK
|MSI X99 SLI Plus:|
LGA 2011-v3, Intel X99
Stock 100 MHz BCLK
|Gigabyte Z97X-Gaming 5:|
LGA 1150, Intel Z97 Express
Stock 100 MHz BCLK
|Case||DIYPC HTPC-Cube-BK||ZALMAN Z11 Neo||Corsair Graphite 230T|
|CPU Cooler||Corsair H60 Closed-Loop||Cooler Master Hyper 612 Ver.2||Corsair H100i Closed-Loop|
|Hard Drive||Samsung 850 Evo 250GB SATA 6Gb/s SSD||Samsung 850 Evo 250GB SATA 6Gb/s SSD||Crucial MX100 256GB SATA 6Gb/s SSD|
|Power||Rosewill RG630-S12: 630W 80 PLUS Bronze||SeaSonic SSR-650RM: 650W, 80 PLUS Gold||Rosewill CAPSTONE-750: 750W, 80 PLUS Gold|
|OS||Microsoft Windows 8 Pro x64||Microsoft Windows 8 Pro x64||Microsoft Windows 8 Pro x64|
|Graphics||Nvidia GeForce 352.86||Nvidia GeForce 352.86||Nvidia GeForce 347.25|
|Chipset||Intel INF 22.214.171.1249||Intel INF 126.96.36.1999||Intel INF 188.8.131.526|
Futuremark 3DMark & PCMark
One of the performance secrets of the full ATX build is that it uses “Enhanced” turbo ratios, whereby the CPU operates at its maximum turbo ratio regardless of the number of active cores, by default. I didn’t expect to see that affect 3DMark scores, but I did notice the stark difference between CPU-based Physics tests. Conversely, the Mini ITX machine overclocked to a fixed 4.3 GHz where the full-sized system varied from 4.0 to 4.3 GHz, and the opposite ratio of Physics performance appears.
PCMark shows that the Samsung 850 Evo in both Q2 machines outpaces the Crucial MX100 of the Q1 system, but not much else.
Sandra Arithmetic loves the fixed overclock of the little system, but otherwise favor’s the big machines quad-channel memory configuration. Cryptography for example gets a big boost in bandwidth-intense Encoding/Decoding turnaround.
Fortunately for the small system, the bandwidth of quad-channel mode isn’t nearly double that of dual-channel mode. Not even when the big machine is overclocked to DDR4-3200 while the small system struggles to reach DDR4-2666.
Arma 3 sees the weakest benefit from the Q1 system’s SLI configuration, while Battlefield 4 and Far Cry 3 get the most benefit from the extra card. The 200FPS cap in Battlefield 4 is a primary reason why SLI systems don’t do better in our SBM gaming tests, and a good reason why we have a separate value chart for gaming at high resolution. Unlike most other games, Far Cry 3 actually needed the second card to play Ultra Quality smoothly at 5760x1080.
Grid 2’s “High Quality” is actually a fairly low setting for testing high-end graphics cards, and is typically limited first by memory performance, then by CPU performance. We indeed see the benefits of DDR4 at our lowest Grid 2 test settings, but surprisingly see little difference between dual-channel (Mini ITX) and quad-channel (Full ATX) configurations.
Media, Productivity And Compression
For the most part, single-threaded applications such as audio encoding favored the higher clock frequencies of the Q1 system, while multi-threaded benchmarks such as video encoding favor the six-core processors of both Q2 machines. The tiny system outperforms the big one when overclocked, because its processor was able to run at a fixed 4.3 GHz frequency.
Remembering that less time is more performance in timed benchmarks, the Q2 ATX machine found the same “fixed frequency” –type advantage at stock speed due to its use of “Enhanced” turbo ratios.
Power, Heat, Efficiency And Value
The Q2 Mini ITX build has fewer onboard components than its full ATX rival, so it consumes less idle power. It also overclocks better, so it consumed more power when overclocked. Yet even the compact system can’t compare to the wild power swings of Q1, where a lower CPU core count produced even smaller power numbers while a pair of GPUs consumed enormous energy when loaded.
I don’t want to start a format war, but the CPU of today’s Mini ITX build did run a little cooler than its Full ATX counterpart, and though that was mostly due to its better CPU cooler, the numbers reveal that it was indeed flowing a sufficient volume of air. Its GPU ran a little warmer, but that same GPU was shoved into the top of the case. Moreover, it had 3% better efficiency than the big system at stock clocks, and its higher overclock only gave it a 2% efficiency disadvantage compared to the overclocked big system.
Oh, hey! Looking at performance-per-dollar, I’d call the overclocked compact machine a success. Of course the big machine was also a success before we overclocked it. Then again, the big machine was overclocked before I overclocked it due to enhanced Turbo ratios. Hmm. Maybe I’ll just let you pick your favorite machine!
And now for the big let-down: The old Q1 system was far heartier at running high resolutions. I kind of expected that. High-resolution performance might have been better still had that machine contained a pair of R9 290X cards rather than a pair of GTX 270s. I’m willing to discuss that, along with other aspects of future builds, in the response thread below.
At least I was able to build a tiny Haswell-E system that, in spite of its “crippling” dual-channel DDR4 configuration and wasting of PCIe 3.0 lanes, was still able to match the average performance and overclocking capability of its full-sized rival. I’d call that a huge success in a machine that’s smaller than a size-13 shoebox, especially since the machine slightly exceeded my own expectations.