Update: Leaked Intel Roadmap Reveals New Batch of Haswell Chips
Faster hardware shouldn’t be this somber. Yet we can’t help but furrow our brow in concern over Intel’s fourth-generation Core i7 CPU, Haswell. Yes, in typical Intel fashion, it’s a tour de force of technical achievement and features that’s the envy of the free world. It’s also, by the way, quite fast.
How fast? *Spoiler alert* Let’s just say that the new Core i7-4770K easily unseats the previous midrange sweetheart, the Core i7-3770K, as the best all-around performer, and even gives the high-end hexa-core part a hard time.
So, why are we so sad? Maybe it’s the continual whispers of the PC’s impending doom—that despite the pure joy a powerful PC can bring the world, its days are numbered.
Or maybe it’s because it’s clear that, while Haswell is fast, it’s a part that is obviously designed primarily to benefit laptops, tablets, and other small-computing needs rather than desktops. Let’s just say, as happy as we are about where Haswell lands in performance, we’re still concerned about Intel’s commitment to performance desktop computing, and that doesn’t make us feel good.
The Haswell Lineup
One new Haswell includes the south bridge inside the package.
Haswell’s Mixed Bag
A new platform is just one of the features that might irk enthusiasts. The most noticeable change for any enthusiast is the introduction of a new socket. LGA1155 has carried us from Sandy Bridge through Ivy Bridge, but as Intel doesn’t like you to ever get too comfortable with a motherboard, it’s shedding that old LGA1155 for a new LGA1150 socket. The two are, of course, incompatible. Why? It’s not just to piss you off, but more likely due to the fact that Intel can’t integrate the new Haswell features in LGA1155. The new socket should come as no surprise to anyone who reads Maximum PC, as we’ve been reporting on Intel’s plan for Haswell for a while, but here it is officially: If you want the new CPU, you need a new motherboard. AMD /AM3+ fanboys can feel free to unleash a big Nelson Muntz–style “ha-ha!” in the faces of Intel fanboys.
You're going to need a new 1150-socket Mobo for Haswell
Got a Spare FIVR, Buddy?
We’ve long said that Intel’s CPUs are gravitational black holes sucking everything into them. Nehalem ate the memory controller. Lynnfield swallowed PCIe. Sandy Bridge gobbled up graphics. And Haswell has a new fully integrated voltage regulator, or FIVR, inside the package. By integrating the voltage regulator, Intel simplifies power inputs into the CPU but also takes a lot of the control out of the motherboard makers’ hands. The FIVR doesn’t eliminate all voltage regulation on the motherboard, as the power to the CPU must still come from somewhere, so you’ll still find boards with beefy caps and voltage regulation circuits.
By integrating the VR, though, Intel is able to regulate power to a much finer degree than has been possible on even the best motherboards. Voltage ripple is practically nil on the design, and the performance of the FIVR outstrips anything that can be done externally. The FIVR also technically lowers the cost of a motherboard, as some of the external voltage regulators are no longer needed.
The FIVR isn’t a free ride, though. It adds more heat to the CPU and raises the TDP of the 4770K to 84 watts compared to the 77 watts of its predecessor.
The bigger question for desktops users is, why? The hard truth is that we suspect the change was not done to benefit desktop PCs. It’s to benefit laptops, all-in-ones, and anything else that really needs precise control over power and voltage in a very thin package. In other words, it’s a move that’s all about mobile and small computing. The good news is that it doesn’t seem to hurt enthusiasts very much. Yes, you’ll need beefy cooling to run Haswell overclocked, but you always needed that.
Overclocking Give and Take
When Intel jumped from Lynnfield to Sandy Bridge chips, it left behind the ability to pump up the base clock to overclock a chip. Overclockers still got unlocked “K” CPUs, but even non-K parts could overclock by four bins through the multiplier on Z-series boards. With Haswell, that feature is now gone, so non-K parts are truly clock-blocked in every way possible.
The good news for enthusiasts is that Intel has added more knobs to K-chip overclocking. Borrowing from the Sandy Bridge-E chips, Haswell K chips will now offer additional CPU straps for overclocking. Rather than being limited to just 100MHz and a few megahertz above it, additional ratios of 125MHz, 160MHz, and 250MHz should be available to help overclock the CPU without overclocking PCIe and other clock-sensitive components.
Big Fat L4—Just Not for You
One of the most exciting developments in the Haswell parts list is the new Core i7-4770R. This one chip features a massive 128MB of embedded DRAM, or eDRAM, to ameliorate memory bandwidth issues in graphics. Don’t care about integrated graphics? You should, because the R part’s eDRAM also acts as a massive L4 cache, which, according to some developers, offers a pretty big boost in performance outside of graphics. The really bad news is that you can’t get it in anything other than a BGA chip today. After hearing the objections of the enthusiast tech press (see, we help you sometimes), Intel is looking at the option of offering a socketed R chip.
TSX for Only Some of Us
Much has been said about Intel’s transactional memory feature, or TSX, in Haswell. TSX essentially makes it easier for programmers to write multithreaded code by addressing the complexities of having to lock portions of an array of data. TSX lets the processor handle much of the grunt work. Now for the bad news: TSX is apparently only available on some Haswell chips. Intel wouldn’t say which chips had it and which didn’t, but a leaked chart on Tom’s Hardware indicates that the only two chips we care about—the two unlocked K parts—don’t have it.
And the Good News?
So, Haswell runs a bit hotter, takes some voltage control out of your hands, eliminates the non-K overclocks, doesn’t give enthusiasts access to the large L4 cache version, doesn’t have TSX in the K parts, and, well, requires a new motherboard, too. You’re probably wondering just where the hell the good news is for enthusiasts with Haswell.
Despite all our bitching, we will say that Intel has at least paid attention to the one metric that counts most: performance. Intel didn’t just take an Ivy Bridge die, erase the name, and pencil in Haswell. The company has added new instructions to Haswell, including AVX2 and FMA2, that will eventually benefit you. The company has also increased the execution ports and generally made a lot of nips and tucks in the name of performance. What this means is that, clock for clock, Haswell offers a noticeable performance boost over Ivy Bridge. The full skinny on Haswell’s performance follows, but let’s just say it again: It’s fast. The apparent lack of TSX, fat L4, and multiplier overclocking might give you a frowny face, but maybe the only people who should really have a frowny face are those who just bought into a full-boat LGA1155 system with a top-of-the-line Core i7-3770K chip.
Click the next page to read about Haswell's integrated graphics.
Haswell Graphics Finally Good Enough? We’re going to kick that football to the moon
Intel’s x86 prowess has almost always been the envy of the free world. We can say that without anyone except the most ardent AMD fanboy objecting, because when it comes to x86 performance, Intel has thoroughly cleaned everyone’s clocks. From PowerPC to 68000K and from SPARC to MIPs, Intel’s x86 parts have sent everyone packing.
When it comes to graphics, though, most agree it’s been the exact opposite. Old salts will remember Intel’s disastrous entrance into discrete graphics with the i740 in late 1990. Despite analyst predictions of an Intel-led graphic-card-ageddon, Intel instead withdrew with its video card between its legs two years later. Though a sucktacular non-success, Intel turned lemons into gold-plated lemonade when it embedded i740 into the 810 chipset. Believe it or not, today, Intel’s integrated graphics dominate in market penetration over AMD and Nvidia graphics chips.
But Intel wants more than to merely excel at sucky graphics, and has been on a steady march to gain some respect. Every year, Intel proclaims its integrated graphics much improved, and usually people shrug it off as Lucy trolling that sap Charlie Brown. With Ivy Bridge, though, Intel’s graphics indeed got better, but overall still couldn't compete with even low-end discrete parts. This time, Lucy says, it’s different. But is it? Are Haswell graphics really, finally good enough to replace discrete graphics? We can’t say—yet.
First, to be fair to Intel graphics, our ability to judge its performance is only based on the HD4600 part embedded in the Core i7-4770K chip—a CPU no one is likely to use without a discrete GPU. The HD 5000 and HD 5200 are actually the “real deal,” but we couldn’t get our mitts on them. HD4600 is indeed better than Ivy Bridge, though. To compare, we used the same rigs we used for the other benchmarks in this story, but ripped out the discrete cards. We saw Haswell offering a healthy increase over Ivy Bridge. In fact, in Portal 2, which we considered unplayable with Ivy Bridge at 1080p resolutions, Haswell gave us reasonable (but not fantastic) frame rates with image quality turned down a few notches. Not bad, honestly. Resident Evil 6 at 1360x768 saw Haswell with twice the performance of Ivy Bridge. Other games also gave Haswell a good leg up over its predecessor. Enough to replace discrete graphics? Hell no. At least, not with HD4600.
The Bantha in the room is really Intel’s HD5200 graphics. With its 128MB of eDRAM, we suspect that it’ll give even AMD’s best APUs a sound thrashing and might finally be the football Charlie Brown has been waiting to kick all these years.
Haswell vs. Ivy Bridge Graphics
Click the next page to see how well Haswell compares to Ivy Bridge and Sandy Bridge-E.
Haswell Meets the Benchmarks: A Mobile-friendly Design Doesn’t Diminish Desktop Performance
Picking which CPUs to test was fairly straightforward: We went with the top-end Core i7-4770K, a Core i7-3770K, and to give people a perspective on how LGA2011 chips compare, a Core i7-3820. Even though it’s out of the price band of the three others, we also decided to test the hexa-core Core i7-3930K to show the quad-core boys what they’re missing, or not, by foregoing the two extra cores.
While differing sockets made it impossible to use identical motherboards for our tests, we at least stuck with one vendor in the hopes of achieving some uniformity. For LGA1155, we used an Asus P8Z77-V Premium, for LGA2011 an Asus Sabertooth X79, and Asus’s new Z87-Deluxe board for the LGA1150. Each motherboard was outfitted with 16GB of Corsair DDR3 clocked at 1,600MHz. All three systems were outfitted with identical 240GB Corsair Neutron GTX SSDs, which had the same firmware and were Trimmed before testing began. For graphics, matching GeForce GTX 580 cards were used, with the same driver on each. For OS, we used Windows 8.
As previously mentioned, in performance, Haswell doesn’t disappoint. For example, we ran Cinebench 10 across just one core to gauge the per-core performance. Haswell gave us about a 17 percent bump over Ivy Bridge. Against the older Sandy Bridge-E cores in the Core i7-3820, Haswell was about 30 percent faster.
In the other benchmarks, Haswell’s advantage remained constant, with the new CPU performing faster than Ivy Bridge in just about every category, and the margins widening against the Sandy Bridge-E chip.
For the most part, Haswell offers an 8 to 15 percent performance boost over Ivy Bridge. In the heavily threaded Premiere Pro CS6 test, Haswell enjoyed a 12 percent advantage. The same 12 percent advantage also popped up in our ProShow Producer test, which tops out at about four threads.
In a search for more real-world results, we decided to throw HDRsoft’s popular PhotoMatix at the CPU. We used a nine-shot RAW file from a Nikon D800 and ran a batch convert on it. From what we can tell, that particular load is heavily multithreaded and literally stops everything else while being run. Haswell pulled out a big win, processing the HDR about 22 percent faster than Ivy Bridge.
In the synthetic 3DMark gaming tests, both CPUs pulled even in the graphics test—no surprise. But in the physics test, Haswell again had a 10 percent gain over Ivy Bridge. In actual game engines, the difference between the two CPUs varied. We ran our tests at low quality and low resolutions to take the GPU out of the equation. In Dirt 3, we saw an astounding 34 percent difference between Haswell and Ivy Bridge, and in Total War: Shogun 2’s CPU test, a 10 percent gain. We could go on, but for the most part, Haswell is across-the-board faster. Part of the performance improvements come from Intel’s under-the-hood tweaks to the chip, but some of it also comes from the Turbo Boost modes. We monitored the chip’s clock speed in some tests and found that sometimes it would run at 4.1GHz for long stretches of time, while the Ivy Bridge chip hung back at 3.8GHz. Combining both the higher Turbo Boost speeds and the efficiency enhancements explains most of its wins against Ivy Bridge.
Haswell already has a leg up over Ivy Bridge, but up against the even older Sandy Bridge-E cores, it’s a bit of a slaughter. If you were to peg the advantage Haswell has over Sandy Bridge-E, we’d say it’s about 20 percent overall when its four cores versus four cores. The only advantage Sandy Bridge-E picks up is when you get to six cores. The Core i7-3930K has an advantage in heavily multithreaded tasks such as 3D rendering and video encoding. In gaming and apps that really can’t exploit all six cores of the Core i7-3930K, Haswell comes out on top.
What Should You Buy?
First, Haswell doesn’t invalidate Ivy Bridge systems. Machines built on Ivy Bridge are still quite fast and offer plenty of performance, but there’s simply no reason to build on an older-generation CPU with a dead socket (good luck getting another LGA1155 out of Intel) with an inferior chipset. Sorry, but that’s the truth, and we know some of you just built Ivy Bridge boxes, too, despite our warning you to wait a month or so.
When it comes to Haswell Core i7-4770K vs. Core i7-3820, it’s a bit more tricky. Haswell flat-out leaves quad-core Sandy Bridge-E in the dust. The only reason to even build a Core i7-3820 box at this point is with an eye toward future upgrades. Intel is expected to push out Ivy Bridge-E or Ivy Town CPUs later this year.
That’s pretty much our guidance for Core i7-3930K, too. The chip is almost twice as expensive as Core i7-4770K but if you work for a living pushing pixels in a 3D renderer, or video encoder, it’s worth it. If you can cut a four-hour video encode down to three hours with Core i7-3930K, you can work on other projects that pay, rather than wait. The only caveat is that Haswell’s higher clock and improved core can actually be as fast, even on multithreaded workloads. Peep our PhotoMatix test that saw the Core i7-4770K actually beat the six-core Core i7-3930K.
Despite all our negativity, we have to say, Haswell can still make us smile. We can’t help but wonder, though, as impressive as Haswell is in performance, what it could have been like if Intel had focused on making Haswell a purely desktop-enthusiast part. We fear those days are behind us, though.
Top Procs Compared
Click the next page for Haswell vs Ivy Bridge comparison pics and to read about the new 8-series chipset.
Haswell vs. Ivy Bridge in Pictures
The move from Sandy Bridge to Ivy Bridge saw graphics die space eat up a lot more area. On casual examination, Haswell looks almost the same proportionally as Ivy Bridge, but does have an additional 200,000 transistors tucked inside. The onboard fully integrated voltage regulator isn’t actually on the die but is inside the package. Intel chips with the 128MB of embedded DRAM will also show an additional strip of silicon inside the package but outside the actual die area.
8-Series Erases Previous Mistakes
Z87 finally gives us SATA 6Gb/s across all ports!
The new 8-series chipsets finally bring SATA 6Gb/s across all ports!
We’ve bitched about Intel not giving us this or not giving us that, but the company has at least finally fixed our biggest complaint since, well, forever. Intel’s P67 first introduced native SATA 6Gb/s that made third-party drive controllers seem horrible by comparison. The only problem? SATA 6Gb/s was only supported on two ports. Back in 2011, this wasn’t an issue, as who the hell had more than two SSDs that could even push a SATA 6Gb/s interface? Today, with SSD prices truly affordable and capable of saturating SATA 6Gb/s ports, it’s a problem. Oh, that and AMD has had native 6Gb/s across all six SATA ports since 2010 (cue Nelson Muntz again).
The new Z87 chipset corrects that. All six ports are SATA 6Gb/s. Intel has also upped the USB support, going from four USB 3.0 ports to six ports, and from 10 USB 2.0 ports to 14 USB 2.0 ports. The Z87 chipset now also supports per-port disabling for security purposes.
There was news earlier this year that the USB 3.0 support in the Z87 chipset was botched. The apparent bug would put USB 3.0 devices into sleep modes. We’ve queried board makers and OEMs who believe that the issue is mostly resolved now, and only existed in earlier revisions of silicon, which shouldn’t reach consumers' hands.
Other than the elimination of support for PCI, the Z87 chipset doesn’t look too different from its predecessor.
Click the next page for our in-depth Haswell benchmarks.
To see how the four top contenders stack up, we’ll get into a deep dive on just how each particular CPU performed.
Cinebench 10 Single Core
Maxon’s Cinebench is based on the engine used in its Cinema 4D 3D modeling program. We use the older version to measure performance on a per core basis. As you can see, it’s a nice generational increase in performance. We don’t get a massive boost but it’s a worthwhile performance difference that likely comes from the under the hood changes to Haswell as well the higher clock speeds the chip seem to hold for far longer than Ivy Bridge.
Cinebench 10 Multi Core
As appoint of reference, here’s Cinebench 10 run across all threads available on the chips. Haswell has a very decent lead but obviously the Core i7-3930K wins. We will note that the even with the four extra threads (two physical and two logical from Hyper-Threading) the 3930K isn’t ahead as far as you would expect in something so heavily multi-threaded.
Cinebench 11.5 has been updated to support up to 64-threads and obviously the newer engine likes cores. The 6-core Core i7-3930K represents itself a bit better here and should make those who spent $550 on the chip feel a bit better.
POV Ray is a ray tracing engine that’s older than some of our readers. Like any rendering app, it’s heavily multi-threaded but look at how much the Core i7-4770K steps away from the older Sandy Bridge-E/Sandy Bridge core as well as from the Ivy Bridge part. The Core i7-3930K saves much face here and proves that six is still better than four for some tasks.
For our Handbrake run we used 0.9.9 to transcode an existing 1080P resolution MP4 file using the Android profile. The result is in seconds and lower is better. Handbrake loves threads and cores. The Haswell Core i7-4770K again opens up a decent lead over the Core I 7-3770K and Core i7-3320 parts but the 6-core Core i7-3930K steals the show with a hefty, hefty performance advantage over the three quad-core parts. Still, Haswell shows its generational advantage over the two other quad cores here with a respectable lead.
We’ve been using TechARP’s X264 HD 5.01 encoding test as a benchmark for more than a year and we’ve been pretty with it as a general indicator of encoding performance. The program performs two passes multiple times. We’ve found the first pass to be slightly more sensitive to clock speeds and memory bandwidth in the past. Here, Haswell again come out on top of all the quad cores by a decent amount. The Core i7-3930K again comes out as Charlie Sheen overall, but it also costs almost twice as much too.
TechARP X264 Pass 2
The second pass of X264 HD 5.01 has been more sensitive to the amount of threads so it’s no surprise the Core i7-3930K wins. Still, Haswell steps ahead of Ivy Bridge and Sandy Bridge-E/Sandy Bridge by a noticeable amount.
Adobe Premiere Pro CS6
Our final video encoding test uses Adobe Premiere Pro CS6 to edit and render out a 1080P video with multiple timelines using source material shot at 1080P on a Canon EOS 5D Mk II at 30 fps. It’s no surprise-Hawell steps ahead of the others by noticeable amount. We’d be OK if we had an Ivy Bridge part, but the performance of the Core i7-3820 against Haswell would make us consider popping it out for a Core i7-3930K part.
PC Mark 7
Unfortunately, PC Mark 8 wasn’t out in time for our testing but we did run the older PC Mark 7 across four chips. PC Mark 7 is a synthetic test, but it’s supposedly based on real-world workloads. Haswell is the winner again with the older Sandy Bridge-E cores including the 6-core Core i7-3930K part, bringing up the rear.
Gigapan Stitch.EFx 2.0
In casting about for some real-world, heavy duty workloads, we used Gigapan’s Stitch.EFx 2.0 to stitch together about 263 images shot with a Canon EOS 7D and GigaPan Epic Pro head to make a 1.1 billion pixel image. We like Stitch.Efx 2.0 because the app is an interesting mixture of single threaded and multi-threaded performance. The first two thirds of the process where images are aligned is single-threaded with the remaining third is where the images are actually stitched is multi-threaded. Higher clocked parts should be favored for the first portion with multi-threaded parts favored for the end. The higher clock and efficiency of Haswell gives it a pretty big win over the others. Interestingly, the Core i7-3930K’s two more cores plus two Hyper-Threaded cores just barely pull even with the Core i7-3770K.
HDRsoft’s PhotoMatrix is one of the most popular and powerful applications for creating HDR imagery. For our test, we used a nine-shot sequence shot with a Nikon D800 DSLR. The RAW files were then processed by Photomatix in batch mode using the “realistic” option. This option is heavily multi-threaded and so resource intense that HDRsoft says it’s only available as a batch mode to keep people from thinking their weak machines have locked up (our words actually.) If you look at Haswell’s performance, it easily steps away from the Ivy Bridge part and makes the elderly Sandy Bridge-E/Sandy Bridge core look positively ancient. Even the four extra threads of the Core i7-3930K can’t put it ahead of the Haswell part even though this app is heavily multi-threaded. This is a pretty big win for Haswell.
ProShow Producer 5
Our final photography-realted benchmark uses our old stand-by ProShow Producer 5 to produce and encode a 1080P video photo slide show using images shot with a Canon EOS 5D Mk II. ProShow tops out with four cores so it makes it a pretty good way to measure a more realistic application load as few actually exploit more than four cores today. Hawell has a pretty sizeable lift over Ivy Bridge as well both Sandy Bridge-E chips here. Interestingly, the Core i7-3930K actually outpaces the Core i7-3820 part despite its lower clock speeds. We’re not sure why but it’s possible the larger cache is a factor.
We run an overall score with 3DMark11 to show you that, well, gaming is mostly about the GPU these days. There’s no reason to even read anymore because we’re just writing this to fill space since there’s not much of a difference. Oh, we don’t get paid by the word? Doh. Move along.
The physics portion of 3DMark11 is multi-threaded and would give you an indication gaming physics. Interestingly, despite its slightly higher clocks on boost and improved core, it’s pretty much a wash with Ivy Bridge. The 6-core pull out in front, but we do really have to wonder if very many games are using all those cores.
The new 3Dmark is an interesting beast and is a bit of a changeup for us. We’ve long seen the 3DMark series as a GPU test but clearly the CPU matters a lot because look at the length of these bars man.
Oh wait, that was just a scale issue. When you actually see the whole chart, you can see the CPU’s impact on overall 3DMark scores is pretty much meh. Move along.
Valve Particle Test
Valve’s Particle test is an older test of how well a CPU will run physics in a Valve game. We’ve long suspected that the test is very cache-speed sensitive as well as very main memory latency sensitive but we’re now wondering if it doesn’t just like CPUs with very larges caches since the Core i7-3930K with its 20MB of L3 cache demolishes the others. Haswell still shows a decent improvement though.
Resident Evil 6
The Resident Evil 6 is based on the actual engine from the popular third-person shooter series. We run it at low resolutions and with image quality turned down to take the GPU out of the equation. The end result shows Haswell with a small lead over the Ivy Bridge part but really nothing to write home about.
Total War: Shogun 2
TWS2 has many options to test performance with. For this, we selected the DX9/CPU test. It’s a CPU-heavy filling a battlefield with hundreds of warriors. The result shows Haswell with a decent lead but again, nothing to write home about. The Total War series is one of the few that will use more than four cores so the Core i7-3930K actually pulls up alongside the higher-clocked Ivy Bridge CPU.
How much more bandwidth do the LGA2011 CPUs have? A ton more than the LGA1156 parts. It’s a neat synthetic to show you that, yes indeed, quad-channel is actually working. The bigger question is what exactly does that extra bandwidth get you in most of our apps? It’s apparently not helping that much.
Update: Leaked Intel Roadmap Reveals New Batch of Haswell Chips
Intel's Haswell architecture hasn't been on the market all that long, but such is the advance of technology that something newer, better, and faster is always right around the corner. In this case, it's still Haswell we're talking about, but according to a leaked roadmap, Intel is getting ready to refresh its Haswell family with a few new chips in the third quarter of 2013, including two that fall into the company's "Premium Performance" category and one that's an "Extreme" part.
As previous roadmaps have shown, Intel appears to be targeting the second quarter of 2014 for a Haswell platform refresh, perhaps indicating that the Santa Clara chip maker will skip over the 14nm Broadwell architecture on the desktop except for BGA chips.
What do you think of Intel's new Haswell processor? Let us know in the comments below!