The Core i7-4770K Review: Haswell Is Faster; Desktop Enthusiasts Yawn

The Core i7-4770K Review: Haswell Is Faster; Desktop Enthusiasts Yawn

Intel's Haswell architecture is finally available in the flagship Core i7-4770K processor. Designed to drop into an LGA 1150 interface, does this new quad-core CPU warrant a complete platform replacement, or is your older Sandy Bridge-E system better?

Editor’s Note: Eager to show off what it's doing with Intel’s Haswell architecture, system builder CyberPower PC is offering the Tom’s Hardware audience an opportunity to win a new system based on Intel’s Core i7-4770K processor. Read through our review, and then check out the last page for more information on the system, plus a link to enter our giveaway!
Do you know what it’s like to be at the top of your game, the nearest competitor several strides behind? Well, maybe not. But Intel sure does. When it comes to desktop CPUs, the company’s top-end parts continue to stave off AMD's best efforts. That applies to raw performance and efficiency.
We love fast, and we love efficient. But we also like to see healthy competition driving innovation. And again, on the desktop, there’s not enough of that to push Intel. Ivy Bridge-based CPUs are generally a small step up from the generation prior. And although the Sandy Bridge architecture included a number of notable improvements, unprecedented integration gave away Intel’s growing focus on mobility. Even as we got our hands on great features like Quick Sync, Intel was chiseling away at its enthusiast equity by limiting overclocking to K-series SKUs.
Expect more of the same from Haswell. You're going to see notable per-clock performance improvements, faster graphics, and additional features able to accelerate specific workloads. But you’re also going to witness a clumsy handling of overclocking (again), some strange decisions on the graphics side (again), and incremental gains that’ll have some of us upgrading our desktops, but more folks looking for Haswell-powered mobile platforms.
That's entirely by design, by the way. An emphasis on power is front and center with Haswell. And as a result, this architecture is going to span the broadest range of devices Intel has ever touched with one design. But I’ll argue that enthusiasts on the desktop take a back seat to make it all possible.

Meet Haswell, Now Known As Intel’s Fourth-Gen Core Architecture

Intel is rolling out the details of its Haswell-based processors in a staggered launch. The company plans to ship multiple variations of the architecture across a number of different interfaces, from very low-power segments to very performance-sensitive ones. However, the only arrangement emerging today is the quad-core SoC. Technically, Intel is talking desktop and mobile, though we’re deliberately focusing on the Core i7-4770K desktop CPU. I published a preview of Core i7-4770K’s performance almost three months ago, and that story has some information about Intel’s plans as well.

Haswell-based quad-core processors will ship in two configurations to cover the mobile and desktop markets. Only one is ready today, though. That chip features the HD Graphics 4600 engine, also known as GT2. The second, with Iris Pro Graphics 5200 (or GT3e) is coming later. Intel's engineers claim that Iris Pro scales incredibly well given a lofty power ceiling and enough cooling. However, CPUs endowed with the higher-end graphics engine are BGA-only, meaning they’re soldered down. So, enthusiasts buying LGA 1150-equipped motherboards will only find Core i7 and Core i5 CPUs with four cores and HD Graphics 4600 (technically, there’s also a 35 W Core i5 with fewer cores, but it’s still under wraps).
This implementation of Haswell is composed of 1.6 billion transistors, up from a comparable Ivy Bridge configuration’s 1.4 billion. Optimized expressly for Intel’s 22 nm node, the die measures 177 square millimeters, just slightly larger than quad-core Ivy Bridge at 160 mm².
Quad-Core Haswell with GT2
Quad-Core Ivy Bridge with GT2
Put Ivy Bridge and Haswell right next to each other and you might have a difficult time telling them apart. After all, there’s “only” a 200 million-transistor delta separating the two. That 14% growth in transistor count largely comes from a 25% increase in graphics resources compared to last generation.
That’s not to say the processor cores go untouched. Intel says it put specific emphasis on speeding up both today’s legacy code as well as applications we’ll see in the future. To that end, larger buffers enlarge the out-of-order window, which means instructions that would have previously waited for execution can be located and processed sooner. Haswell’s window is 192 instructions. Sandy Bridge was 168. Nehalem was 128. The Haswell branch predictor is improved, too. This is something Intel manages to do every generation—and for good reason, since it simultaneously enables better performance and prevents the wasted work of a branch getting predicted incorrectly. Previously, Intel’s architecture was able to execute six operations per clock cycle. However, Haswell gets two additional ports (one integer ALU and one store), enabling up to eight operations per cycle. And workloads with large data sets should see a benefit from a larger L2 TLB.

All of those changes add up to significant improvement in Haswell’s IPC compared to Ivy Bridge. That’s where we expect most of the speed-up in general-purpose apps to come from this generation, since the top-end Core i7-4770K runs at the same 3.5 GHz as -3770K.

Sure enough, when we set five different processors (employing four different architectures) to the same constant 4 GHz, we see, first, how much more work Intel gets done compared to AMD and, second, a steady progression forward in Intel’s performance.
In addition to the two execution ports Intel adds to Haswell, ports one and two now feature 256-bit Fused Multiply-Add units, doubling the number of peak theoretical floating-point operations per cycle. Integer math gets a big boost as well from AVX2 instruction support.

Of course, multiplying the architecture’s compute potential means little if you can’t get data into the core fast enough. So, Intel also made a number of changes to its caches. Haswell’s L1 and L2 caches are the same size as they were in Ivy Bridge (there’s a 32 KB L1 data, 32 KB L1 instruction, and 256 KB L2 cache per core). Bandwidth to the caches is up to doubled, though, and we’ll see in our synthetic testing that the L1D is indeed quite a bit faster. Intel claims that it can do one read every cycle from the L2 (versus one read every other cycle in Ivy Bridge), but we aren’t able to replicate those figures in our own testing.

	Cores / Threads	Base Freq.	Max. Turbo	L3	HD Graphics	Graphics Max Freq.	TDP	Price
Fourth-Gen Core i7 Family
4770T	4/8	2.5 GHz	3.7 GHz	8 MB	4600	1,200 MHz	45 W	$303
4770S	4/8	3.1 GHz	3.9 GHz	8 MB	4600	1,200 MHz	65 W	$303
4770	4/8	3.4 GHz	3.9 GHz	8 MB	4600	1,200 MHz	84 W	$303
4770K	4/8	3.5 GHz	3.9 GHz	8 MB	4600	1,250 MHz	84 W	$339
4770R	4/8	3.2 GHz	3.9 GHz	6 MB	Iris Pro 5200	1,300 MHz	65 W	N/A
4765T	4/8	2.0 GHz	3.0 GHz	8 MB	4600	1,200 MHz	35 W	$303
Fourth-Gen Core i5 Family
4670T	4/4	2.3 GHz	3.3 GHz	6 MB	4600	1,200 MHz	45 W	$213
4670S	4/4	3.1 GHz	3.8 GHz	6 MB	4600	1,200 MHz	65 W	$213
4670K	4/4	3.4 GHz	3.8 GHz	6 MB	4600	1,200 MHz	84 W	$242
4670	4/4	3.4 GHz	3.8 GHz	6 MB	4600	1,200 MHz	84 W	$213
4570	4/4	3.2 GHz	3.6 GHz	6 MB	4600	1,150 MHz	84 W	$192
4570S	4/4	2.9 GHz	3.6 GHz	6 MB	4600	1,150 MHz	65 W	$192

The Core i7-4770K gives us an 8 MB shared L3 cache, similar to Core i7s before it. Although the Sandy and Ivy Bridge designs employed a single clock domain that kept the cores and L3 running at the same speed, Haswell decouples them. Our cache bandwidth benchmark reveals a slight hit to L3 throughput, though improvements elsewhere in the System Agent keep the results fairly even.
Haswell offers the same 16 lanes of PCI Express 3.0 connectivity as Ivy Bridge, and validated memory data rates up to 1,600 MT/s. The desktop line-up’s thermal targets are quite a bit different as a result of Intel’s fully-integrated voltage regulator, but an upper bound of 84 W isn’t extreme by any stretch and a floor of 35 W is pretty familiar.
All of Intel’s upgradable processors now drop into an LGA 1150 interface, meaning any decision to adopt Haswell is also going to require a motherboard purchase, at least. So, before you drop several hundred dollars on a brand new platform, let’s figure out if Core i7-4770K is worth the investment.

 Overclocking Haswell: You’ll Pay For That

If you purchase one of the two K-series parts in Intel’s Haswell-based portfolio (and you have a Z87-based motherboard), you’ll have access to the same knobs and dials that were available from the Core i7-3770K and i5-3570K, with one notable change: the BCLK ratios popularized by X79 Express are back, facilitating some manipulation of the number that gets multiplied against the clock ratio. Intel exposes 1.0x, 1.25x, 1.67x, and 2.5x straps.
There’s a colossal difference between what Haswell-based LGA 1150 processors can do in theory, and what they can do after Intel’s product guys get done disabling the bits and pieces to create a differentiated stack, though.

Overclockers maintain control over core frequency using ratios up to 80x in 100 MHz increments. The on-die graphics engine is also adjustable in 50 MHz increments via ratios as high as 60x. Further, the memory controller is technically unlocked, allowing options for 200 and 266 MHz steps with a logical ceiling at 2,933 MT/s. Finally, the platform controller hub’s clock generator is unlocked too, supporting frequencies as high as 200 MHz.
But as with generations past, there remains a relationship between the DMI clock and PCI Express clocks—and you want PCI Express to stay as close to 100 MHz as possible. So, Intel implements 5:5, 5:4, 5:3, and 5:2 ratios to maintain constant PCIe clocks with a variable BCLK. Unlike the LGA 2011 platform, which used the CPU to adjust those ratios, Haswell requests this from the PCH. The outcome is similar, though. Expect a useable range between 5-7% around the ratio you choose for fine-tuning your overclock.
That’d all be well and good if Intel was enabling this new level of flexibility for the folks who don’t spend extra on K-series parts, giving them the ability to pick higher BCLK settings without access to the ratio multiplier. However, the company instead chooses to restrict the ratios to the Core i7-4770K and i5-4670K—the same ones you can already overclock in 100 MHz increments. Anyone buying one of the 11 other SKUs in Intel’s new Core i7 and i5 line-up is out of luck.

Overclocking Core i7-4770K

Moving on, what can you expect from a Core i7-4770K, in terms of overclocking headroom? We have a couple in the lab, and are getting 4.7 GHz, at most, across all cores using Prime95 to test for stability. However, those samples come from Intel. We were much more interested in feedback from someone with many, many retail parts at their disposal.

Our first-hand information involves a high double-digit number of processors, including samples and final shipping boxed CPUs. Sort testing was limited to 1.2 V to keep heat manageable. Ring/cache ratios are pegged at 3.9 GHz, with the memory controller operating at 1,333 MT/s. Of the chips available for sorting, only one is stable at 4.6 GHz under full load. A few are capable of operating at 4.5 GHz. More run stably at 4.4 GHz. Most are solid at 4.3 GHz and down. As you stretch above a 1,600 MT/s memory data rate or a ring ratio to match your highest single-core Turbo Boost ratio (which helps maximize performance), your top stable core frequency tends to drop.

Load-line calibration has very little, if any, effect with Haswell. There’s simply less flexibility to coax a killer overclock out of one of these CPUs using levers that were popular previously. Although the most enthusiast-friendly motherboards still include pages of options, a number of them simply don’t do anything beneficial. Flatly, it sounds like engineers and enthusiasts alike are still trying to figure out what the more obscure settings actually do.

We’re going to have to accept that Haswell-based parts get hotter, faster, it sounds like, and that they might fall a few hundred megahertz short of comparable Ivy Bridge-based parts with conventional cooling methods. It’s a good thing, then, that the architecture is inherently faster to help compensate.