1. Home >
  2. Computing

Haswell review: Intel's Core i7-4770K takes over the pole position

Intel's new Core i7-4770K launches today, into a market fraught with challenge. Is this the architecture Intel needs to win back space in mobile computing?
By Joel Hruska
Haswell die
Today, Intel is launching its much-discussed line of Haswell processors. By the end of the day, you should be able to pick up a Core i7-4770K for $339 from your favorite component stockist. Unlike Sandy Bridge, which launched in Q1 2011 into a relatively calm market, Haswell is debuting at a time when Intel is pivoting to challenge ARM for the future of handheld computing. There's a great deal riding on this microarchitecture -- does the desktop hardware measure up?

Core, rearchitected

Intel may call every "tock" of the tick-tock model a new architecture, but not all architectural iterations are created equal. With Haswell, Intel did something rather dangerous. Specifically, Chipzilla decided to dramatically increase L1/L2 cache bandwidth, add execution resources, beef up the chip's out-of-order execution capabilities, and integrate a new voltage regulator directly into the die.

Building a low-power, high-performance x86 core that uses less power than its predecessor is challenging. Continuing to tweak a highly evolved design to improve execution efficiency and performance per watt is challenging. Doing them both at the same time is an order of magnitude more difficult, but that's the task Intel set for itself. Has it succeeded? We're going to find out.

The CPU we're reviewing today is the Intel Core i7-4770K, a quad-core, Hyper-Threaded processor with a 3.5GHz base clock and a 3.9GHz Turbo Mode. The 4770K has an 84W TDP (up 10% from the Ivy Bridge-based Core i7-3770K's 77W). At first glance, that's the biggest difference between the two processors. But the processors' specs, in this case, don't even start to tell the story. At a high level, Ivy Bridge and Haswell look almost identical. They aren't.

Cache bandwidth comparison

This chart captures the cache bandwidth gap between Ivy Bridge and Haswell. If you compare across all three generations, you'll see that Sandy Bridge's internal bandwidth was also higher than the first-generation Core i7's in a number of places. Haswell builds on Sandy Bridge, but adds bandwidth in more areas. L2 cache bandwidth (shown below) also doubles, up to 64 bytes per clock, from 32.

Microarchitecture enhancements

More bandwidth is useless without more execution resources, which is why Intel has added two new execution ports to Haswell. The simplest way to think about these new ports is that they add execution flexibility. Haswell has larger reorder buffers for fine-grained control over the out-of-order execution process, and it has 168 integer registers and 168 AVX registers, up from 160 and 144 respectively on Ivy Bridge. The AVX integer registers have also been bulked up, and are capable of handling 256-bit data structures rather than being limited to 128-bit.

The first iteration of AVX was introduced in 2011, with the launch of Sandy Bridge. AVX has several advantages over Intel's older SIMD extensions, but couldn't handle integer code. Given that the vast majority of computing workloads use integer math, the value of AVX has been somewhat limited. AVX2 lifts those restrictions by adding support for integer instructions and expanding the integer registers to 256 bits. Haswell is Intel's first chip capable of using the FMA3 instruction set. Chip and Chimpzilla have been batting two different standards back and forth (AMD was first to market with both). Now that both companies have settled on FMA3, adoption should start to improve.

Focusing in on AVX2, IPC, low-level differences

We tested the Core i7-4770K using Intel's DZ87KLT-75K motherboard and 8GB of DDR3-2133 RAM in two DIMM slots. The Core i7-3770K was tested using the Intel DZ77GA-70K motherboard with the same RAM, at the same clock speed. An OCZ Vector 256GB SSD was used in both cases, as was an Nvidia GTX 680 graphics card. All testing was done using a fully patched version of Windows 7 64-bit. Power was provided by a 1275W Thermaltake Toughpower 80 Plus Platinum PSU. We used the same stock Intel cooler on both chips to compare thermals in an apples-to-apples environment.

Different motherboards were necessary because Haswell uses Intel's new LGA1150 design while Ivy Bridge and Sandy Bridge used the same LGA1156 design. LGA1156 coolers will fit LGA1150 sockets -- if you upgraded to Sandy Bridge several years ago and want to move to Haswell, you won't need a new CPU cooler to do it.

If you're primarily curious about high-level application performance in common programs, the PC Magazine review likely has more of the data you're looking for(Opens in a new window). Here, we're going to focus on the low-level impact of AVX2, FMA3, and the general gain in CPU efficiency between Ivy Bridge and Haswell in current workloads. Unfortunately, problems with the motherboard Intel sent us for review prevented us from testing Haswell's new integrated graphics solution. This examination, therefore, is CPU-only.

Next page: Haswell benchmarked

Sisoft Sandra 2013

Adrian Silasi recently updated SiSoft Sandra to take advantage of AVX2 and FMA3 on Haswell. The standard arithmetic tests show no gap between the two chips, but the multimedia tests are rather interesting. We tested Haswell with AVX2 and FMA3 disabled, with FMA3 enabled, and finally with both features.

Sandra multimedia

Haswell rockets forward in integer code, outstripping Ivy Bridge by a full 78%. Gains in the the other three tests are more modest, but the core still manages a full 20% improvement over last year's model. Cache bandwidth benchmarks are a useful way for checking whether or not the chip is living up to its bandwidth claims.

Cache bandwidth

We're surprised that the L2 cache is just 7% faster on Haswell, but the massive L1 bandwidth is a huge jump. Interestingly, Sandra didn't show us much of a benefit in standard AVX code, without FMA3 or AVX2 enabled. Let's see if that continues to be the case.

 

 

Kribibench 3.0

When Sandy Bridge launched, the web-based Kribibench player made the rounds as one of a handful of benchmarks that could be run with and without AVX enabled. We've returned to it for this round of testing. Remember, AVX code should theoretically be faster on Haswell thanks to cache and dispatch improvements, even if the software hasn't been recompiled to take advantage of AVX2 or FMA3. Kribibench 3.0 has three tests -- Skyline_City, Robots, and Koteks.

 

Kribibench 3.0Kribibench 3.0 Even in non-optimized code, Haswell pulls forward by 10-12% in all three tests.

x264 Encoding

x264 is an easy place where AVX2 could make a difference in performance, but the TechARP benchmark was last updated a year ago. We've opted to show two sets of results here. First, the benchmark in its current form, and second,  the relative performance between the two chips if we swap to the newest x264 binary, released several days ago.

x264-Original

In the default test, Haswell is faster on both counts. Pass 1 completes 13.5% faster, while Pass 2 picks up 7%. So what happens when we update to a version of the program that incorporates additional AVX support?

 

x264 Encoding (5.01)

 

Performance skyrockets,  at least in Pass 1. The Core i7-3770K is 34% faster, while Haswell's performance jumps 45%. The drop-off in Pass 2 is a bit puzzling, but given the newness of the executable, we're betting it's a minor issue that'll resolve with a tweaked version. Regardless, the R2334 binary delivers far more performance than the older program.

Cinebench 11.5

Finally, I want to talk about Cinebench performance. This 3D rendering application is three years old, and it's not been optimized for AVX, much less AVX2. It's an example of what we can expect from Haswell in comparisons against Ivy Bridge when both chips are handed older software.

Cinebench 11.5

Again, Haswell comes through with a solid 7% gain in single-thread, 10% in multi-thread while at the same clock speed. For unoptimized code that doesn't play to any of the chip's strengths, those are reasonable gains.

Next page: Power consumption and overclocking

The other shoe: Power consumption and overclocking

A year ago, I wrote an article discussing Ivy Bridge's weak overclocking potential when compared to previous Core i7 processors. At the time, I was openly dubious of the idea that Haswell would bring back the golden age of overclocking. To make a long story short -- it hasn't. The consumer Haswell processors, including the 4770K, continue to use thermal paste rather than solder, which has a significant impact on temperatures and maximum overclocking headroom.

We've spoken to several boutique manufacturers and high-end overclockers on the topic, all of whom have tested multiple chips. Group consensus is that while Ivy Bridge could hit 4.6 - 4.7GHz on air, Haswell struggles to reach 4.5GHz. Memory clock speeds are dicey at this speed, with very few chips capable of hitting DDR3-1866 or higher. As for power consumption, it looks like this:

Power Consumption

Idle power is measured at the desktop, 15 minutes post-boot. Load power was measured using Cinebench 11.5, after three consecutive runs. Peak power consumption was measured after 10 minutes of Prime95's Torture Test, using the latest AVX-capable executable and the second Torture option for max heat and power consumption. We retested power consumption multiple times after adjusting various BIOS settings. Even after allowing for a margin of error, the evidence points to Haswell drawing slightly more power than Ivy Bridge, not less.

The 4770K is also more conservative when calculating available headroom. After 10 minutes of Prime95, using the same CPU cooler, the i7-3770K's temperature was 90C. The chip's clock speed, as measured by CoreTemp, remained at 3.7GHz. Haswell, in contrast, throttled back much more quickly. While it started off at 3.7GHz in Prime95, the core dropped back to 3.5GHz within seconds. This may be a mechanism for ensuring the core doesn't overheat -- when we took manual control of Turbo Mode and locked in the 3.7GHz Turbo Boost, the 4770K's temperature shot up to 100C -- a full 10C hotter than the 3770K.

The culprit here is almost certainly the fully integrated voltage regulator, or FIVR. Integrating the VRM into the die gives Intel fine-grained control, but there's no avoiding the substantial thermal dissipation of the VRM under heavy load. Intel has claimed that Haswell will deliver 50% longer battery life than Ivy Bridge. We're not going to say that's impossible, but based on what we've seen of the 4770K, it's obvious that the power savings didn't translate over to the desktop line.

Next page: Conclusion

Putting it altogether: Haswell or Hasbeen?

Haswell is 6-10% faster than Ivy Bridge, almost without exception. In AVX-optimized code, that gap can grow to 15-20%. In a few cases, FMA3 and AVX2 kick that delta to over 50%. This design has legs for the long haul, and there are going to be people in certain fields who are downright excited to see it ship. The new core, with its integrated VRM and lower-power targets in mobile, was unlike anything Intel has built before, and the company managed to increase execution efficiency while keeping power consumption steady on the desktop. That's notable.

Is it good enough to justify an upgrade? That's not a simple question. If you have an old Core 2 Duo system, then yes, absolutely. If you're still rocking a Core i7-920, I'd say the same -- 2.66Ghz to 3.5GHz is a major jump in its own right. Sandy Bridge and Ivy Bridge owners, however, probably don't have much reason to buy into Haswell at this point in time. Waiting another year or two won't meaningfully degrade your current system's performance, and the amount of AVX-ready software is only likely to increase.

Our low-level benchmarks made one thing clear -- long term, this chip has wings. It's a solid product, even if the realities of semiconductor manufacturing have pushed upgrade cycles out to lengths that were unimaginable 10 years ago. Even if you don't buy a Haswell this cycle, you'll use the technologies it pioneers in future hardware, and for Intel, that's arguably more important.

Tagged In

Haswell Core I7 Cpu Components Core I7-4770k

More from Computing

Subscribe Today to get the latest ExtremeTech news delivered right to your inbox.
This newsletter may contain advertising, deals, or affiliate links. Subscribing to a newsletter indicates your consent to our Terms of use(Opens in a new window) and Privacy Policy. You may unsubscribe from the newsletter at any time.
Thanks for Signing Up