intel-2011-lcs-01

Intel Sandy Bridge Extreme i7-3960X Processor Review

Add Your Comments

It has been about a month since our last CPU launch article. These things aren’t usually hot on each others’ heels, but we’re certainly not going to argue. AMD’s Bulldozer was not exactly what everyone – especially AMD fans – had hoped it would be. Enthusiasts were hoping for an Intel killer but got a slight upgrade instead.

Now it’s Intel’s turn. The hype surrounding Sandy Bridge – E hasn’t been the same as it was for Bulldozer: this isn’t an architecture rewrite, rather it’s an improvement on an existing core. We have already seen that AMD has nothing to compete with the Intel i7 2600K so there’s nowhere to go but up for Intel.

Sandy Bridge Extreme – It Goes to 11!

The Sandy Bridge architecture, which is just stellar in efficiency, processing power and overclocking (until you hit the multiplier wall) came out in January of this year. We went over that architecture in our Intel i7 2600K review.  Today, Intel has released the top-end hex-core version of Sandy Bridge, known as Sandy Bridge Extreme (or Sandy Bridge-E/SNB-E) and it’s a doozy. The Intel i7-3960X flagship processor is equipped with six cores and twelve threads, like the i7 980X and i7 990X from the first i7 range of chips, but these are Sandy Bridge cores. We’ve all been amazed over the last eleven months at how well these cores perform; adding two more should make the new LGA2011 platform scream.

In addition to the six full cores, the 3960X has 15 MB of L3 cache (L1 and L2 cache per core is the same as for Sandy Bridge), supports quad-channel DDR3-1600 RAM and operates at a 3.3 GHz stock clock speed. Turbo kicks that up to 3.9 GHz for lightly-threaded workloads. The CPU has 2.27 billion transistors in a rather large die coming in at 20.8 mm x 20.9 mm, versus the ca. 1 billion for quad core Sandy Bridge chips.

i7-3960x Feature Summary

i7-3960x Feature Summary

i7-3960x Die

i7-3960x Die

I’m sure that you can’t help but notice that there is space on that chip for two more cores. At this time, Intel has revealed no plans for a consumer-grade octo-core SNB-E processor. There may well be eight core Xeon chips on the way, but even that is unclear. Intel says that they had to fuse off those cores (sorry folks, no core unlocking) in order to maintain competitive clock speeds while staying within TDP. The clock speed hit from allowing all eight cores to be active wasn’t worth it. They also make the point that this is the path they’ve been on for a while and has nothing to do with AMD’s eight-core CPU being unable to compete with their hex-core CPUs.

If you haven’t heard of Intel’s Turbo feature, here’s how it works: if you have a workload that uses three to six cores, you could see a turbo boost of up to 300 MHz, depending on the CPU load. If you have a very lightly threaded load using only one or two cores, you could see a boost of up to 600 MHz. In testing, it functions fast and flawlessly.

i7-3960x Turbo Boost 2.0

i7-3960x Turbo Boost 2.0

With the new platform comes, of course, a new chipset. Intel has gone the way of the PCH for their enthusiast platform too: gone is the northbridge of the very long-in-the tooth LGA1366 platform. The CPU handles traditional northbridge functions now.

With 40 full lanes of PCIe connectivity, SNB-E has more than double the graphics throughput of the original Sandy Bridge CPUs. This gives us the ability to run two graphics cards at the full PCIe 2.0 16x for each slot. It also introduces the ability to run more than two cards without resorting to the NF200 for additional PCIe lanes. You can run four graphics cards (1 at 16x plus 3 at 8x) for example.

There has also been discussion (along with some complaints) about the X79 platform not being PCIe 3.0 compatible. The question has been put to Intel who say there just aren’t enough GPUs to test for PCIe 3.0 certification right now, so they can’t legally put the PCIe 3.0 label on their parts. It will be up to motherboard partners to claim whether their boards meet the specifications of PCIe 3.0 or not. Intel does say that the platform is fully capable of the required 8GT/s data rate, so it will run at PCIe 3.0 speeds, but they just can’t apply the label for legal reasons. There are rumors of potential LGA2011 chips coming in 2012 that will officially be considered PCIe 3.0 parts, but they’re just rumors right now.

The on-chip memory controller that forms part of the SNB-E CPU is revamped and is specified to handle quad-channel DDR3-1600 RAM. Of course, you’ve probably already seen the DDR3-2133 quad-channel kit G.Skill sent us, so it’s safe to say we can expect better things out of this memory controller.

That leaves the X79 chipset itself, which is slightly disappointing to be honest. There is no increase in SATA III (6Gb/s) connectivity, with only two ports. Just like P67 and Z68 there are also four SATA II (3Gb/s) connections. There are also no USB 3.0 ports, just 15 USB 2.0 ports. Therefore, you’re looking for at least one third party USB 3.0 controller on any decent board, plus another third party controller if the manufacturer wants to add more SATA 6Gb/s ports.

Rounding out the X79 PCH is Intel High Definition Audio (HDA), a LAN interface and eight more lanes of PCIe 2.0 connectivity. This is all connected to the CPU through a high-speed DMI interface.

X79 Chipset Block Diagram

X79 Chipset Block Diagram

Now we move on to the numbers from the Intel presentation. These were provided and we’ll share them because they show the differences between their previous generation and the current one. We’ll show our own numbers below, but this is how Intel sees the differences. First up is a  i7-990X vs. i7-3960X comparison.

i7-3960X vs. i7 990X Content Creation

i7-3960X vs. i7 990X Content Creation

i7-3960X vs. i7 990X Larger / Faster Cache

i7-3960X vs. i7 990X Larger / Faster Cache

i7-3960X vs. i7 990X Sandra

i7-3960X vs. i7 990X Sandra

i7-3960X vs. i7 990X SPECint

i7-3960X vs. i7 990X SPECint

There is nothing at all to sneeze at there, with significant gains over the previous generation across the board. Intel actually uses several benchmarks that we use too, so we get to test these out. Up next is an i7-2600K vs. i7-3960X comparison:

i7-3960X vs. i7-2600K Content Creation

i7-3960X vs. i7-2600K Content Creation

i7-3960X vs. i7-2600K Larger / Faster Cache

i7-3960X vs. i7-2600K Larger / Faster Cache

i7-3960X vs. i7-2600K Sandra

i7-3960X vs. i7-2600K Sandra

i7-3960X vs. i7-2600K SPECint

i7-3960X vs. i7-2600K SPECint

While these numbers are impressive, it’s very important to note these tests do focus heavily on multi-threaded applications. The single-threaded performance should remain pretty similar between the i7-2600K and the i7-3960X, with the exception of memory-intensive loads that should take advantage of the SNB-E quad-channel memory.

Intel has changed one other important thing with SNB-E: as rumored, SNB-E will not ship with any cooling solution. The vast majority of people purchasing this platform will view this  as a complete non-issue, especially when Intel is still offering their full 3-year warranty with these CPUs sans heatsink. However, that brings us to one potential issue – this is a new socket. Thankfully it’s not as big of an issue as you may think. The difference between LGA 2011 and LGA 1366 is threading. The hole spacing is the same, but the LGA2011 backplate has threads in the holes. ASUS addresses this on their Rampage IV Extreme by giving you a star-drive key and 1366 backplate, which just holds down the socket bracket and opens up the holes for your existing LGA 1366 cooler.

If your board doesn’t come with a backplate change kit, you may have to get inventive. If you happen to be using an EK Supreme HF water block with the old mounting kit (like I am), you’ll be happy to know that the bolts fit perfectly into the backplate; just screw them in and you’re good to go. Another solution would be to drill out the screw holes to fit your heatsink which may lead to warranty problems, so the last solution would be a new mounting kit from the manufacturer of your cooling solution.

Thankfully, some manufacturers are ahead of the game. Noctua, Thermaltake, and Cooler Master (Hyper 212 only) are already offering free  LGA 2011 brackets. Hopefully others will follow.

Intel is still offering their little air cooler for less than twenty bucks but, at any rate, I’d buy a better solution for $30 before even considering those dinky things. Intel is also offering their own closed-loop liquid cooling solution. Built by Asetek, these are very close to the kit AMD is bringing to market for Bulldozer. There are significant differences though: the Intel kit only comes with one fan and doesn’t have a user-selectable LED on the coldplate/pump assembly. These kits are expected to be priced somewhere around $85-$100.

Socket 2011 Air Cooler

Socket 2011 Air Cooler

Intel Closed Loop Water Cooler

Intel Closed Loop Water Cooler

Tick tock goes the Intel clock, but there is no tick nor tock this time around. These chips are manufactured on the same 32 nm process and have the same architecture as Sandy Bridge, it’s just optimized with a couple cores and features added. In recent years there hasn’t been a major Intel release that wasn’t either a tick or a tock, so this is a slight departure from the norm. Never fear though, Ivy Bridge’s tick is supposed to be coming soon in 2012. Plus, if you have one major competitor and they can’t compete with your current mainstream chips, why force a new architecture or process shrink when you don’t need one? Just give it a couple more cores, optimize it and equip it with a quad-channel memory controller and call it a day.

Sandy Bridge-E - Tock Continued

Sandy Bridge-E - Tock Continued

Now we tackle an important issue to those considering this platform – price. There is not a large departure here from previous generations. Intel’s flagship i7-3960X will be priced (per 1000 units) at just below a kilo-buck – $990.

The big news here is that the 3960X is not the only unlocked CPU. Unlike the i7-970 CPU, which had a locked multiplier, the i7-3930K has an unlocked multiplier. The only difference on paper between the 3930K and 3960X is the stock speed (100 MHz at stock and turbo) and the loss of 3 MB of L3 cache. Priced at $555, the 3930K will most likely be the perfect CPU for overclockers wanting a strong Intel hex-core set-up but without spending a thousand dollars on the CPU alone.

How that cache difference impacts upon performance is yet to be seen. Plus, we don’t know how the binning process is different. Will the K-series average lower maximum clock speeds? Time will tell. We’ll know for at least a couple of CPUs soon enough though, as Overclockers Editors MattNo5ss and EarthDog as well as Benching Team Leader Brolloks all plan on purchasing the 3930K to see what it’s made of.

2nd Generation i7 Family Comparison

2nd Generation i7 Family Comparison

Now that we’ve gone through the slide presentation, let’s look at the chip itself. As you can see, it’s huge. Massive even. That little CPU to the left? That’s a 2600K. Yowza!

i7 2600K & i7-3960X

i7 2600K and i7-3960X

i7 2600K and i7-3960x

i7 2600K and i7-3960x

As you know I have a penchant for taking plenty of photos, so here are a few more.

i7-3960x

i7-3960x

i7-3960x

i7-3960x

i7-3960x

i7-3960x

i7-3960x

i7-3960x

There you have it folks; you’ve been introduced to the new Sandy Bridge Extreme. Impressed? Not yet? Well, keep reading, we’ll see if benchmarks will impress you instead.

Intel DX79SI Motherboard

Intel sent their DX79SI “Siler” motherboard with the i7-3960X to evaluate SNB-E’s performance. We’ll do a relatively quick overview of it here, but to be honest, people looking at a thousand dollar CPU will probably want more than this board has to offer. It’s solid, stable and will run the CPU with no problems; it’s a no-nonsense, get-the-job-done-reliably motherboard that does well in that regard.

Packing & First Look

Intel has gone the windowed route too and offers a preview of the board through a window in their box. The flap and the box have specifications and features all over it. They’re not extensive like some other boards, but the feature set is solid for what it is.

Intel DX79SI Box

Intel DX79SI Box

Intel DX79SI First Peek

Intel DX79SI First Peek

Intel DX79SI Box Rear

Intel DX79SI Box Rear

Intel DX79SI in Hard Plastic

Intel DX79SI in Hard Plastic

The board itself isn’t bad looking at all. They did away with the blinking skull and kept this one LED-free. The PCH is under the skull and that heatsink attaches to the one mid-board to offer additional cooling. The VRM heatsink is at the top and does a good job of dissipating the heat from the power section.

Intel DX79SI In The Flesh

Intel DX79SI In The Flesh

Intel DX79SI

Intel DX79SI

Intel DX79SI

Intel DX79SI

Intel DX79SI

Intel DX79SI

Intel DX79SI

Intel DX79SI

Intel DX79SI Back

Intel DX79SI Back

Board Close-Up

Here we have our first look at the socket area on Intel’s board. This ‘backplate’ is strange compared to other offerings in that the threads are on top of the board. On the ASUS Rampage IV Extreme, the threads for the socket plate are actually on a backplate, which would offer more stout heatsink pressure.

As you can see, the Intel board can take up to eight DIMMs, which can – if you have the money – handle up to 64 GB of memory.

LGA2011 Socket Area

LGA2011 Socket Area

There isn’t a ton of connectivity on this board. Intel have chosen not to expand SATA-based storage capability with a third party controller, so you have the standard six SATA ports available from the X79 PCH: two SATA 6 Gb/s and four SATA 3 Gb/s.

The backplate is similarly sparse. If you have a PS/2 keyboard, you’ll need to upgrade or get an adapter for this board. Intel did include a third party controller for USB 3.0, which gives you two ports. There are six USB 2.0 ports, dual Intel LAN ports, a FireWire port, audio output, and a button that forces booting into BIOS when an overclock fails. Note, however, that it is not a clear CMOS button.

SATA Ports

SATA Ports

Rear I/O

Rear I/O

Here are a couple quick pictures of controllers. The board uses a Realtek audio codec (on the left) and on the right you can see the LAN controller.

Intel Ethernet Controller & Realtek Sound

Realtek Sound

Intel Ethernet Controller

Intel Ethernet Controller

There are some enthusiast features to be had, though. Onboard power and reset switches and a POST code indicator are supplied. In the far bottom left are LEDs that correspond to different facets of the system. When booting up, these will blink through to let you know where a fault is if booting fails. This is similar to ASUS’ QLED and is definitely a nice feature to have for diagnosing trouble spots. There is also a USB 3.0 front panel header which is handy. In a nod to benchmarking enthusiasts there are even voltage read points for VCore and VDIMM. These are a nice touch, but I doubt that anybody is going to be pushing voltages on this board so much that you need a multimeter.

The rest is standard fare – front panel audio, power/reset/USB headers, and so on.

On-board Power & Reset, POST Code Indicator

On-board Power & Reset, POST Code Indicator

Additional Front Panel Connectivity

Additional Front Panel Connectivity

Like the DP67BG supplied for the Sandy Bridge review, the BIOS on this board doesn’t have extensive tweaking options. Most annoying is the fact that you cannot increase the base multiplier past the stock setting of 33x. To overclock you must use the turbo multipliers. You can set it so you’re only adjusting one multiplier instead of six, but there is no way to turn off speedstep and overclock. Frankly, that’s just a pain.

So while this board won’t tickle the fancy of die-hard overclockers, it does run the SNB-E platform reliably and overclocks decently. Massman recently said that “no enthusiast in his right mind would opt for the Intel board.” While I think it’s a little harsh, with better options from partners available for roughly the same price (ASUS’ P9X79 should retail for $299, this board will come in between $280 and $300), I’m inclined to agree with him.

It’s not that this is a bad board by any stretch, it’s just not as good as its competition. If it were priced at, say, $250, it could be a good option for people on a budget: maybe for someone running distributed computing applications that keeps the system at stock or uses moderate overclocks for ultimate 24/7  stability.

Power Consumption, With a Little Folding Thrown In

We never focus on this too heavily, but power consumption is definitely an important measurement for some – especially those on our distributed computing teams (and I guess those other teams too). So we’ll have a couple quick numbers and move on with the review. Load numbers were taken after applying a load on the CPU with Prime95 Small FFTs.

Test Setup Idle (Watts) CPU Loaded (Watts)
i7 2600K 97 W 158 W
FX-8150 121 W 246 W
i7-3960X 104 W 244 W

It definitely requires more power than the 2600K and is pretty much on par with the FX-8150 under load. People complained about that power consumption, but I think the performance-per-watt was the killer there. For instance, Folding@Home on an SMP work unit, the FX-8150 could only churn out 13,700 ppd (points per day). By comparison, look at what an i7-3960X can pump out.

Folding@Home - SMP Work Unit

Folding@Home - SMP Work Unit

Yes, you read that right. Over 43,500 ppd. For the same power draw, this CPU puts out well over three times as much work. While it may not be the most power-sipping processor out there, it sure does a whole lot of work per watt.

Overclocking for Stability

Overclocking Sandy Bridge-E is as easy or as complicated as you want it to be. Since the K and X chips have unlocked multipliers it really can be as easy as setting the VCore and multiplier. That’s not the end of the road like Sandy Bridge though. SNB-E gives you a new option – FSB strap. While everything is still linked to the FSB, like Sandy Bridge, SNB-E brings in a FSB multiplier to give you more flexibility with overclocking. Depending on the board, there can be three or four straps available to use. This Intel board has 1.0x, 1.25x and 1.66x. The ASUS Rampage IV Extreme we’re reviewing adds a fourth strap at 2.5x. Those are multiplied by the stock 100 MHz FSB to get 125 MHz, 166 MHz or 250 MHz FSB.

Above that, you have about the same range the original Sandy Bridge had with regard to FSB frequency. The maximum that you’re likely to see is +/-  about 10 MHz (probably less). It’s not truly a resurgence of complete FSB control, because of all the traditional northbridge functions located on the CPU die. What these multipliers do give you is lots of flexibility relative to the original Sandy Bridge CPU range. You can have FSB set to up to 250 MHz +/- ~10 MHz and there are a lot of available RAM multipliers. Tweaking is back with SNB-E; it’s not as finely-grained as some would like, but it’s plenty to keep you busy for a long time.

Author’s Note – For clarity, I’m using FSB interchangeably with base clock (or BCLK) here.  FSB overclocking like on socket 775 is dead. Thanks to Brolloks for pointing this out.

For a 24/7 overclock, I set the CPU voltage in BIOS to 1.4 V and saw what I could come up with. The limited Vdroop control was set to its performance setting, so there was likely still some Vdroop. I didn’t measure with a multimeter and monitoring software failed on Intel’s board. Anyway, with a 1.4 V Vcore set in BIOS, the 3960X achieved a very respectable frequency of 4.5 GHz, which is a 1.2 GHz overclock.  As this board wouldn’t cooperate with DDR3-2133 memory, I used the 1.25x FSB strap combined with the memory divider that gave DDR3-2000. Obviously, CPU-Z doesn’t read memory properly on this board (or Vcore), but the memory was running at DDR3-2000 with timings of 9-11-9-28.

4.5 GHz LinX Stable

4.5 GHz LinX Stable

As you can see, temperatures (in degrees Celsius) were only in the upper 50’s to low 60’s. This was on a decent water loop consisting of a Swiftech MCP-35X pump and reservoir, Swiftech MCR-320 radiator with three 120 x 38 mm ultra high speed Panaflo fans (running only about 2/3 power) and an EK Supreme HF Cu CPU block. This is the same loop that we ran our sample chip with for the Bulldozer review previously.

Test System, Opponents and Methodology

Firstly, a huge THANK YOU to benching team member thobel for generously volunteering his time by running 2D benchmarks on his i7-980X. Without numbers from that CPU, this review would have been missing a key component. Thobel stepped up to the plate and took care of business. Thank you so much! While we’re thanking, the 2500K results are the same as were used for our Bulldozer review and were run by our editor EarthDog.

We have a lot of competition today. Not all CPUs are featured in every benchmark. The ones that are will be clearly labeled in the graphs. First-generation Sandy Bridge chips are lighter blue hues, AMD chips are different shades of red, the previous generation hex core i7 980X is green and the i7-3960X is two shades of darker blue, depending on whether it’s stock or overclocked.

CPU Intel i5 2500K Intel i7 2600K Phenom II x6 1100T
Stock / Turbo 3.3 / 3.7 3.4 / 3.8 3.3
Motherboard Gigabyte G1 Sniper2 ASUS P8P67 WS Revolution ASUS Crosshair V Formula
RAM DDR3-1600 9-11-9-24 G.Skill RipjawsX DDR3-2133 9-11-9-24 G.Skill Flare DDR3-2000 7-9-7-24
GPU n/a AMD HD6970 AMD HD6970
CPU AMD FX-8150 Intel i7 980X Intel i7-3960X
Stock / Turbo 3.6 / 4.2 3.3 3.3 / 3.9
Motherboard ASUS Crosshair V Formula ASUS Rampage II Extreme Intel DX79SI
RAM G.Skill Flare DDR3-2000 7-9-7-24 Super Talent DDR3-1866 8-10-8-24 G.Skill RipjawsZ DDR3-1600 9-11-9-28
GPU AMD HD6970 n/a AMD HD6970

The BIOS was the latest available from Intel at the time of testing. They did not update prior to launch, so it’s the best we had at our disposal. All benchmarks were run on Windows 7 Professional x64. The stock benchmarks were run three times each and the results you see are averaged. The only exceptions were 3D benchmarks, game tests, and overclocked benchmarks, which were run once each.

The results you see below are graphed relative to the Intel i7-3960X’s stock performance. This means that results by the i7-3960X at stock all equal 100% and the other results are graphed as a percentage relative to the its performance. So, for instance, if the i7-3960X scored 200 points on a benchmark and the i7 980X scored 180 points, on the graph the  i7-3960X would be 100% and the i7 980X would be 90%, meaning in that benchmark, the i7 980X is 90% as good as the  i7-3960X.

Benchmark Results

We’ll start out as usual with the only benchmarks run solely at stock – the AIDA 64 benchmark suite.

AIDA 64 Benchmarks

First up in the AIDA 64 suite are the CPU benchmarks.

AIDA CPU Benchmarks

AIDA CPU Benchmarks

AIDA CPU Benchmarks II

AIDA CPU Benchmarks II

Well then, I certainly hope the rest of the review looks like this. If so, Intel has mopped the floor with…well, with everything. Only in CPU Hash does AMD even come close. In everything else, the Gulftown chip beats it with the new SNB-E chip opening a can of whoop-arse even on the Gulftown chip. How does it perform with floating point operations?

AIDA FPU Benchmarks

AIDA FPU Benchmarks

AIDA FPU Benchmarks II

AIDA FPU Benchmarks II

For the most part another great performance. In the SinJulia test, the 980X get close, but everything else is completely trounced by the 3960X. What about that quad channel memory?

AIDA  Memory Read & Write

AIDA Memory Read & Write

AIDA Memory Copy and Latency

AIDA Memory Copy and Latency

It’s not looking so great for quad channel memory; the 2600K seems to out perform its bigger sibling. However, in a conference call with Intel I mentioned the AIDA memory results. They conducted their own testing and found out that AIDA tests are single-threaded. That’s not nearly enough to saturate all of the memory channels available, and I actually remember similar comments when tri-channel memory first came out.

There is an application with that kind of ability though – SiSoft Sandra. I ran their cache and memory bandwidth and latency tests against the 2600K. Unfortunately, this information came too late to test against other chips, but considering the 2600K beat everything else already, if the 3960X comes out ahead, it is reasonable to consider it the strongest memory performer

SiSoft Sandra Cache & Memory

SiSoft Sandra Cache & Memory

Wow. That’s more along the lines of what you should see from quad channel memory. Bandwidth for memory and cache simply trounces the dual-channel Sandy Bridge.

Reviewer’s note: In the interests of full disclosure, because of when this info came to light, the Sandra tests were run on the Rampage IV Extreme in testing now. This meant that the memory speed was greater than that above, at the full DDR3-2133 / 9-11-9-28. This still demonstrates that the quad channel memory set-up is superior to the dual channel configuration used for the i7-2 range of CPUs.

3D Benchmarking

When Sandy Bridge came out, many folks that didn’t have an i7 980X already scrambled to grab a 2600K because of its stellar performance in 3D benchmarks. Even those with 980X’s considered it because to score better than a 2600K on air / water, you had to take the i7-980X cryogenic. With this launch I had a feeling people into heavy 3D benching would need to bench those cards all over again.

3DMark06

3DMark06

Starting with 3DMark06, my instinct was dead on, with the 3960X beating up on the competition not only in the CPU test, but making a dent in graphics scores as well.  Overclocking shows tremendous increases across the board for both graphics and CPU tests.

3DMark Vantage

3DMark Vantage

Vantage tells a very similar story. The graphics gains are there, but are not nearly as tied to the CPU as the 06 results. The CPU score at stock beats the 2600K by almost 35%! Overclocked, that increases to a whopping 59%.

When it came out, 3DMark 11 was a GPU-bound monster. Can the 3960X change that?

3DMark 11

3DMark 11

It’s not a night-and-day difference in overall score but, with this CPU, 3DMark 11 gives up a little of its GPU-bounded ground. There is a noticeable difference moving to this CPU, even more so when overclocked. There is a greater than 200 point increase from the stock 2600K to the overclocked 3960X. Looking at the overclocked Physics score, it’s just insane: 14780 points on water alone is crazy.

HWBot Heaven is also somewhat of a GPU bound bench, but not as much so as 3DMark 11.

Heaven Benchmark

Heaven Benchmark

This result was quite surprising. The DirectX 9 “basic” test shows very little increase when overclocking, but the “Xtreme” DirectX 11 test got a massive boost. Also surprising is the fact that it doesn’t seem to take advantage of all the cores available to it, with the 2600K beating the stock 3960X by a couple percent.

Gaming

While we don’t do a ton of game testing here, we do understand there is a large portion of our audience that games and overclocks. In fact, that’s how I got into overclocking in the first place: trying to squeeze a few more FPS out of a 7600GT/Core 2 Duo E4400 combo in Day of Defeat 1.3b.

All games that we benchmark are run at a resolution of 1920 x 1080 with all settings turned up (except one iteration of Aliens vs. Predator). When we do game testing, we find the most value in tests that will show real-world conditions, not some artificial testing at 640 x 480 to attempt a CPU bottleneck. First is Stalker: Call of Pripyat.

Stalker: Call of Pripyat

Stalker: Call of Pripyat

We already knew the 2600K was a stellar gaming CPU and that holds true here. No doubt the scores would improve equally well if you overclocked it to the same level.

The next graph has a trio of game tests. Aliens vs. Predator was run in two configurations – the default settings and with everything turned to maximum. HAWX 2 and Dirt 2 were both run at the usual maximum settings.

AvP DX11 Bench, HAWX 2 and Dirt 2

AvP DX11 Bench, HAWX 2 and Dirt 2

Aliens vs. Predator doesn’t seem to care much about frequency or thread count; it was pretty close to equal across the board, with the i7-3960X actually bringing up the rear. HAWX 2 definitely shows more CPU dependency, but the 2600K still wins out. Dirt 2 doesn’t look so hot either.

The biggest conclusion here is that these games don’t really scale with thread count and in games that don’t, the 2600K does just as well as the 3960X. If you can find a game that uses all the threads available, no doubt this one would come out ahead. Time didn’t allow to get Battlefield 3 installed for this review, but we’ll see if we can’t add it in the comments.

Rendering, Video Conversion and Compression

Now to where this CPU should really shine – real world, CPU intensive applications like rendering, encoding and compression.  First up, the former – rendering.

Cinebench R10

Cinebench R10

Cinebench R11.5

Cinebench R11.5

That’s what I’m talking about. Trouncing the competition, then grinding its defeated silicon into the ground to finish it off overclocked. With a couple rendering tests in the bag, let’s see one more.

PoV Ray 3.7 RC3

PoV Ray 3.7 RC3

Very similar scaling here. Now a different task – encoding. x264 runs two passes. The first is a scanning pass and the second performs the actual encoding.

x264 Benchmark

x264 Benchmark

More stellar scaling here. This is shaping up to be quite the work horse. Last in our real world testing is compression with 7zip.

7zip

7zip

Again, this CPU comes out far and away on the top of the pack. SNB-E is most definitely a great choice for anyone that uses their CPU for tasks that can use all of its horsepower.

2D Benchmarking

Now the fun part. This is what many benchmarkers have been waiting for. Is it better than Sandy Bridge for single-threaded applications? How about multi-threaded FPU-based WPrime? The time has come to find out.

SuperPi 1M

SuperPi 1M

SuperPi 32M

SuperPi 32M

SuperPi definitely doesn’t disappoint. As expected, it’s very similar to the 2600K’s performance. With the slight turbo advantage up to 3.9 GHz it has a small edge. Quad channel also comes into play in SuperPi 32M, showing that it can make up for the difference between DDR3-2133 (dual channel) and DDR3-1866 (quad channel).

Last, but certainly not least, one of my favorites – WPrime.

WPrime 32M

WPrime 32M

WPrime 1024M

WPrime 1024M

Wow. It doesn’t put a the i7 980X to shame but without a doubt it’s a very strong improvement. Overclocked, the 3960X really flexes its muscle. This is going to be quite the competitor for six-core WPrime on HWBot! With the FSB strap flexibility it may even turn out to be better competition than just seeing who can bin the most chips to find one that can reach the highest multiplier.

Pushing the Envelope

Now the fun part – seeing how high we can go! Remember though, this is the Intel board. It suffers from Vdroop, so with another board (like the ASUS Rampage IV Extreme I’m also working on), there is the potential for more at this same voltage.

The self-imposed limit in this case was 1.52 V set in BIOS. High clocks in this case ended up at 5.0 GHz. The good part is that it would even run WPrime 1024M there; so as long as the chip is up for it, there is room to go higher, even on ambient cooling. SuperPi first.

SuperPi 1M @ 5000 MHz

SuperPi 1M @ 5000 MHz

SuperPi 32M @ 5000 MHz

SuperPi 32M @ 5000 MHz

Not bad, not bad at all. Coincidentally, this is the same speed with an Intel board the 2600K got to in its initial review. The RAM speed was different but the CPU was the same. SuperPi 1M lost 0.093 seconds, which is basically just a matter of running the bench again, so I’d say that’s parity at the same clocks. SuperPi 32M is a different story. As expected, the quad channel RAM showed its true power, shaving 5.289 seconds off. That’s no rounding error.

Now to where this CPU will shine – WPrime.

WPrime 32M @ 5000 MHz

WPrime 32M @ 5000 MHz

WPrime 1024M @ 5000 MHz

WPrime 1024M @ 5000 MHz

WPrime 1024M isn’t quite as impressive as expected actually, scoring about the same as 980X’s around that same clock speed, but this is with Windows 7; times would definitely improve in XP. WPrime 32M on the other hand is killer and scores around where 980X and 990X CPUs on HWBot are at 5.6-5.8 GHz.  I’d say a roughly 700 MHz gain is a significant jump for one generation to the next.

Clocks will improve too. I’ve already seen 5.1 GHz+ (still on water) with the Rampage IV Extreme and have high hopes once temperatures go below zero. Speaking of sub-zero, on that conference call I asked about that. We had already expected as much, but it’s for certain that we’ll see the same behavior as Sandy Bridge. I’ve been told by reliable sources that they stop scaling around -50 degrees C, so a strong single stage (or dry ice) should be enough for these too.

It’s important to note that Intel has said repeatedly that their focus is on optimizing these CPUs for air and water cooling. Their “…more integrated PCIe architecture is optimized for more real world conditions vs. maximum frequency under extreme cooling.” They feel that’s where they should focus their efforts and indeed on a decent water loop these things perform wonderfully with low temperatures. Like Sandy Bridge, chances are going cold might allow you higher clocks but it will definitely allow the same clocks you get on ambient cooling to be more stable and run with less fear of killing your chip.

Final Thoughts and Conclusion

Sandy Bridge-E is just plain awesome. The i7-3960X out-performs every other CPU on the market, bar none. AMD has nothing to get close to competing with SNB-E. Bulldozer disappointed many and this just makes it look even worse. They also have nothing on this price bracket.

Price is going to be a huge factor in the adoption of SNB-E, especially in this economy. But if you’re going to get the best, you’re going to have to pay for it. Just like the extreme editions before it, the i7-3960X will sell for a cool grand. That’s a lot of scratch and isn’t really something the majority of people are willing to drop on a CPU. Intel has a savior though – the i7-3930K.

Unless we find compelling evidence that the 3930K is binned to have a significantly lower top end than the 3960X, that’s the area where many overclockers will focus their energies. We don’t yet know how the loss of 3 MB of L3 cache will affect the performance, but we’ll find out soon when our other editors get their hands on one. SuperPi 32M immediately comes to mind, but the other benchmarks may not show that significant of a difference. More is always better, but if that’s the only difference between the chips, it’s hard to think it will be worth an extra $450.

Regardless, Intel hasn’t changed. Their best CPU – and the best CPU you can purchase on the market – will cost you $1,000. We’ll leave it to you to decide if it’s worth that to you, but it is by far better than anything else that competes with it right now. If all you do is lightly threaded applications, the 2600K is still going to be your go-to CPU. If you do anything at all that takes advantage of the available threads on this beast, or if you multitask and take advantage of them manually, this is the processor for you. It clocks like a dream, runs cool and processes everything you can throw at it with the stellar Sandy Bridge efficiency we’ve come to love.

 

– Jeremy Vaughan (hokiealumnus)

If this leaves you wanting more on this new platform, never fear; we have lots of SNB-E coverage coming for you! Motherboards from ASUS, MSI and Gigabyte as well as RAM from G.Skill, Kingston and Patriot. For now, we’ll leave you with what’s currently on the test bed…

ASUS Rampage IV Extreme Testbed

ASUS Rampage IV Extreme Test Bed

Leave a Reply

Your email address will not be published. Required fields are marked *

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>

Discussion
  1. Great but... that price.... deal breaker for me.

    Sure the performance out does the BD on pretty much everything here but... the trends has been the same after AMD's Anthlon days. The thing that keeps me from ever buying the chip is it's price. That is still a bit ridiculous for me.

    Power consumption is never a concern since it only adds two or three cents to my electric bill for a high end part, but I am not willing to shell out that kind of cabbage for something I don't even know a program will ever run. I will stick with my phenom and intel celeron for now :P.

    Also, a quick hack for those who program. Turn off the freaking Intel Genuine on intel's instruction compilers. Not everyone is rich for a Intel part, and must suffice with AMD.
    Let me save you some time, Ive been running 2600k vs. 3960X head to head for a few weeks now. Clock for clock, SB and SB-E perf is the same for single thread and single GPU but SB-E dominates in things like rendering, video/audio editing, or anything multi-thread. SB-E also outperforms SB in multi-GPU by 2-5% depending on the test and the GPU (better GPU = more of an advantage). Add more video cards and the perf advantage of SB-E grows more. That is it in a nutshell.

    Brolloks
    First thing I want to do this weekend is do clock for clock compatisons between a 2600k and 3930k to see if there is any gain with the new chipset, I would expect so with quad channel RAM and more cache but I don't expect more than 5-10%.
    hokiealumnus
    Thanks, I'll do that. I haven't folded in a long time and had no idea the difference in WUs was so much, my apologies.


    There's a FAH benchmarking program running around the net. http://www.avid-edge.com/fah/FAHbench_v12.zip

    It has a couple standard WU's that were "captured" (older) and the program allows you to select the WU and number of ticks (% of the WU) to fold. It'll allow you to bench a3's and a5's. The WU's are constant with the program so will allow everyone do directly compare performance using the program.
    David
    These machines only really suit massive megamultitaskers or people who crunch lots and lots of data.

    I'm sure Xeon-based datacrunching rigs based on these SNB-E chips will be very popular.


    Of course I agree completely, I was referring only to enthusiasts. More power is always useful for enterprise data crunching.
    Great review. I'm in the same boat as everyone else thinking that unless you are a cruncher, bencher, or video encoder then you aren't getting much benefit over SB. I'm considering building a new pc, but with IB looming I'll keep being patient.
    ChasR
    Nice review.

    A suggestion for future folding Benchmarks, resize the columns in HFM so we can at least see what WU the benchmark was run on. Performance on 2600Ks (4.8 GHz) on normal SMP WUs in Linux vary from 25K to 45K ppd depending on the WU. 43,500ppd on p7500 isn't really Impressive, while that performance on p6058 would be fantastic.


    Thanks, I'll do that. I haven't folded in a long time and had no idea the difference in WUs was so much, my apologies.
    Nice review.

    A suggestion for future folding Benchmarks, resize the columns in HFM so we can at least see what WU the benchmark was run on. Performance on 2600Ks (4.8 GHz) on normal SMP WUs in Linux vary from 25K to 45K ppd depending on the WU. 43,500ppd on p7500 isn't really Impressive, while that performance on p6058 would be fantastic.
    No doubt Intel's R&D is top class, I'm really impressed how they can punch out such step up technology in less than 12 months from each other, exciting for us who like to play with new stuff :)
    I told myself I was probably going to upgrade from my 920 when 2011 rolled around. Now that it is here, I find myself thinking that the $700 for the K CPU and a motherboard would be better spent on a video card, SSD, and raid storage combo for my existing rig.

    It isn't that this socket is bad, but it seems like anyone sitting on any generation of i7 quad core won't see much improvement in day to day tasks by upgrading to the newest release.

    Intel is kicking so much ass that developers seem to be 3 or so years behind the hardware curve at this point. Insane...
    First thing I want to do this weekend is do clock for clock compatisons between a 2600k and 3930k to see if there is any gain with the new chipset, I would expect so with quad channel RAM and more cache but I don't expect more than 5-10%.
    Angry
    Great review, Hokie.

    As for the processor...."meh."

    $1k is just to much to ask in my opinion, esp since the price hasnt dropped on a 990x.....


    Intels flagship processors never get a noticeable price drop. They just get bought out of stock then disappear.
    Brolloks
    My recommendation is unless you bench or do heavy encoding where time is of the essence this platform is not worth upgrading for, it is after all targeted as an Enthusiast platform, nice to play with but way to rich to aquire as a everyday rig or for gaming.

    The 2500k/2600k remains imo still the best setup, very efficient and affordable.


    Completely agreed with this. The SB-E chips are not made for gaming of this generation and maybe not even the next until they start using heavily threaded applications. They are for workstation/benchmarking and not too much else at this point in my opinion.
    Brolloks
    My recommendation is unless you bench or do heavy encoding where time is of the essence this platform is not worth upgrading for, it is after all targeted as an Enthusiast platform, nice to play with but way to rich to aquire as a everyday rig or for gaming.

    The 2500k/2600k remains imo still the best setup, very efficient and affordable.


    How's the (EVGA?) board working out for you?
    My recommendation is unless you bench or do heavy encoding where time is of the essence this platform is not worth upgrading for, it is after all targeted as an Enthusiast platform, nice to play with but way to rich to aquire as a everyday rig or for gaming.

    The 2500k/2600k remains imo still the best setup, very efficient and affordable.
    Great review, Hokie.

    As for the processor...."meh."

    $1k is just to much to ask in my opinion, esp since the price hasnt dropped on a 990x.....
    3930K/3960K is currently the best cpu on the market. Even though, as Hokie's review clearly pointed out, core for core isnt better than 2600K which was not unexpected, it was designed to be and is 2 more cores of the best cpu (2500K, 2600K) on the market, albeit at a premium price.

    If most software could utilize 6 cores, this cpu would be seen in a very different light (though maybe not the shortcomings of platform).

    It is just the fact that most software that people routinely use doesn't tax 4 cores let alone 6, hence most people are going to stick with 2500K, 2600K.

    And in that regard, only benchers/folders/crunchers are likely to pay for increased cost since only they are likely to benefit.

    But really not a cpu issue, as it is the best on market, it is a failure of software to keep up with hardware.

    Edit: god Im a slow typist, hokie beat me by 9 mins.