Intel has been releasing information about Haswell, its next-generation CPU architecture, for months now. Our coverage has already been fairly extensive—we've already got a nice overview of Haswell's CPU architecture itself, along with a primer on its brand-new integrated GPUs. All we need to know now is specific product information, and Intel is finally giving us our first official taste of that today.
This morning's announcement will revolve around high-end quad-core chips in the Core i7 and Core i5 families, 12 for desktops and ten for laptops. If you're looking for specific information about U- and Y- series low-voltage chips for Ultrabooks or anything belonging to the Core i3, Pentium, or Celeron families, you'll have to wait a little while longer. We'll be sure to pass that information along as we have it.
What we'll do here is present a high-level recap of the CPU, GPU, and chipset enhancements Intel is introducing in Haswell. After that, we'll break down the specific CPUs that Intel is announcing today, and the kinds of systems you're likely to find them in. Note that all of this information is coming directly from Intel—they're not going to out-and-out make things up, but they're definitely going to present their CPUs in the best possible light. We'll be putting all of the below performance claims to the test as we begin to review Haswell-based systems later this summer.
The CPUs: Performance is up a little, power consumption is down a lot
The biggest part of the fourth-generation Core launch is Haswell itself, the new CPU architecture that superseded last year's Ivy Bridge (which was itself only a modification of 2011's Sandy Bridge). Haswell is a "tuck" in Intel's overarching "ticktock" refresh strategy, meaning that it is a new architecture built on Intel's established 22nm 3D tri-gate manufacturing process. Next year's refresh, code-named Broadwell, will modify the Haswell architecture for manufacture on Intel's forthcoming 14nm process.
David Kanter has already done a deep dive into the particulars of Haswell's architecture, and I won't attempt to duplicate his explanation and analysis here. Like previous architectures, Haswell should increase performance relative to Ivy Bridge while consuming similar or lower levels of power.
This time around, Intel is actually much more interested in telling us about that lowered power consumption, as is evident in the use of phrases like "biggest [generation-to-generation] battery life increase in Intel history." By the company's measurements, a laptop based on Haswell should in some circumstances be able to get as much as a third more battery life than the same laptop based on Ivy Bridge. We'll need to wait for actual systems to bear that supposition out, though—it's entirely possible that OEMs will choose to keep a battery life level while reducing the size of the battery (and thus, their devices' weight and thickness) or adding another power-hungry features like high-resolution displays.
Some of these power savings come because Haswell is the first architecture tailored specifically for the 22nm process rather than being designed for a larger process and then shrunk down, as Ivy Bridge was. Others come from improvements that Intel has made it to idle power usage and the amount of power used when switching between those active and idle states. A new "active idle" power state, also dubbed S0ix, allows the system to continue sending and receiving data without powering the CPU and GPU all the way up, increasing the amount the system can do without being "active" and reducing the amount of power needed to switch to a fully active state.
A side effect of Haswell's low-power leanings is that the architecture will find itself in more form factors than ever, from workstations and servers at the top end to Ultrabooks, convertibles, and tables at the low end. There will even be some overlap with Atom processors, as we've already seen in some Ivy Bridge-based Windows 8 tablets. That area of overlap will only get larger as Atom continues to get faster (see: Bay Trail, due later this year).
Intel has been pushing Ultrabooks and convertibles hard through the life of the Sandy Bridge and Ivy Bridge architectures and expect that push to continue with Haswell—a source at Intel even told Ars that Haswell-powered Android tablets and laptops are in the cards as the company refines its Android strategy.
Of course, most of this information applies to dual-core, Ultrabook-ready CPUs that haven't yet been announced. Given that Ultrabooks and low-power CPUs are a large and growing piece of Intel's pie, we doubt we'll make it past Computex without those CPUs making an appearance as well.
The GPUs: More power, more choices, more potential for confusion
Intel has improved its integrated GPUs by leaps and bounds in the past two years, and Haswell will continue that trend. With this increased power comes increased complexity, though: there are now at least five separate integrated GPUs that will come with Haswell CPUs, up from three in Ivy Bridge.
But all five of them should support the same basic features: Direct3D 11.1, OpenCL 1.2, and OpenGL 4.1 on the API side; DisplayPort 1.2, which should improve support for 4K resolutions; and a faster version of Intel's QuickSync video encoding engine are all among the highlights.
Intel's new "Iris" GPUs are the ones we'll probably hear the most about, though as we'll see in a bit they're certainly not the ones that will ship in the majority of its chips. The Iris 5100 will roughly double the performance of the Intel HD 4000 GPU in most cases, but it will be confined mostly to performance notebooks and larger Ultrabooks because of its 28 watt TDP (ultra-low-voltage Haswell Ultrabook chips will generally have TDPs of 15 watts or less). The Iris Pro 5200 GPU adds 128MB of eDRAM directly to the processor die to boost performance even further—about 2.5 times the Intel HD 4000 in laptops, according to Intel—but those CPUs' even-higher TDP of 47 watts will confine them to larger notebooks. In desktops, the sole announced CPU with the Iris Pro 5200 delivers almost three times better performance than the HD 4000 by Intel's metrics, likely due to the increased thermal headroom available in a desktop chassis.
The other GPUs, which retain the Intel HD Graphics branding, are less glamorous but remain a reasonable upgrade over the Intel HD 4000. Intel's slides all seem to suggest that the HD 5000 and the HD 4600 GPUs both deliver performance roughly 1.5 times better than the HD 4000. The difference is that the HD 5000 part is destined for as-yet-unannounced 15 watt TDP Ultrabook parts, while the HD 4600 is the mainstream GPU you'll find paired with most of the CPUs being launched today. The differences between the HD 4600, 4400, and 4200, all of which are lumped together on Intel's slides, aren't immediately apparent using the information Intel is providing, but we'd guess that the difference may be in features and perhaps maximum clock speeds rather than raw 3D execution resources.
Finally, there's the lowly Intel HD Graphics at the bottom of the slide, about which we know very little. If past is prologue, this will be a basic GPU with most of the extra features (QuickSync, for instance) disabled, and it will ship only in Haswell-based Pentium and Celeron processors. Typically, new architectures don't trickle down into entry-level CPUs until a few months after their introduction, and I expect that to remain true for Haswell.
The chipsets: Say goodbye to PCI
As usual, a new generation of CPUs brings with it a new chipset: Intel's 8 series. Unlike the 6- and 7-series chipsets introduced with Sandy Bridge and Ivy Bridge respectively, which could support all CPUs from both architectures given BIOS support, all Haswell CPUs will require 8-series chipsets. The Haswell desktop CPUs also use a new socket, LGA 1150, which will supposedly be compatible with next-next-generation Broadwell CPUs as well.
At a high-level, the 8-series chipsets build on the basic foundation laid by the 7 series. Both chipsets support a total of 14 USB ports and 6 SATA ports, but the 8-series can provide six USB 3.0 ports (up from four) and six SATA III ports (up from two SATA III and four SATA II). All USB 2.0 and 3.0 ports in the 8-series chipsets are now controlled by xHCI controllers, which simplifies the driver stack and allows for lower-power operation when idle than EHCI (previously, only the USB 3.0 ports were controlled by xHCI).
Finally, the PCI Express situation will be the same: the PCI Express lanes controlled by the chipset (normally the smaller PCI Express x1 slots used for non-GPU peripherals) are still based on version 2.0 of the specification. Dedicated GPUs will support PCI Express 3.0 thanks to support baked into the CPU itself, just as with Ivy Bridge. The memory controller also continues to be integrated into the CPU; 1600MHz DDR3 is Haswell's highest officially supported speed, same as Ivy Bridge.
One final change that has been a long time coming but may cause problems for desktop users with older expansion cards is that the 8-series chipsets completely drop support for legacy PCI cards. At this point in history the 20-year-old PCI specification has been well and truly superseded by the faster and more flexible PCI Express, which was introduced in 2004. But just as with floppy drives and optical drives we're sure that some of you will have trouble letting go.
Perhaps the biggest difference with the Haswell chipset isn't in what it does, but in how it's integrated. The lower-power 15 and 27-watt mobile processors can have the chipset integrated onto the same package (though the CPU and chipset dies remain separate). Separate CPU and chipset packages are still necessary for the higher-power CPUs, though, including the models that pack the highest-performing Iris Pro 5200 GPU.
The products: All quad-cores, all the time (for now)
And now for the best part: making sense of the many-headed Hydra that is Intel's product matrix. We'll start with the desktop chips and move on to the mobile ones. You can click the charts below to enlarge them.
Intel is announcing six Core i7 CPUs for desktops, and no fewer than five of them all share the "i7-4770" model number despite not sharing all of the same features. When you buy an i7, you get four CPU cores that can process two threads each thanks to Hyperthreading. The CPUs differ mostly in clock speed and TDP, but the i7-4770R has 6MB of L3 cache instead of 8MB and neither it nor the i7-4770K supports Intel's vPro, TXT, VT-d, or SIPP features.
Most of Intel's model number suffixes are the same as in Ivy Bridge. The T suffix denotes low-TDP models with lower power consumption and a lower base clock speed, the absence of a suffix indicates a faster CPU with a higher TDP, and the S suffix denotes something in between the two in both speed and TDP. The K suffix continues to be used for unlocked, overclocker-friendly CPUs.
A new addition is the R suffix, which indicates a high-performance GPU—in this case, the top-of-the-line Iris Pro 5200. Note that this i7-4770R is only available in a BGA (read: soldered to the motherboard) package, implying that you'll find it mostly in highly integrated or all-in-one computers and you won't be able to buy one to pop into your own desktop. All of the non-R CPUs use an Intel HD 4600.
Finally, the 35W i7-4765T CPU is slow, but offers the lowest TDP of the lot, making it best-suited for low-power systems where having more, slower threads is preferable to fewer, faster ones.
Moving on to the i5 CPUs, you retain four CPU cores and save about $100 compared to most of the i7s, but you also lose Hyperthreading and drop to 6MB of L3 cache rather than 8MB. As in the i7s, you can usually tell how fast one i5 will be relative to the others by the suffix used. Note however that the i5 lineup lacks both an R-series part with higher graphics performance and a 35W TDP part. All of the i5 parts still use an Intel HD 4600 GPU.
All of the mobile parts being announced today belong to the i7 family, and like their desktop counterparts they're all quad-core chips with Hyperthreading enabled. The differences here come in the suffixes: the MX and MQ parts (reversed from Ivy Bridge's QM and QX suffixes) come with the Intel HD 4600 GPU, while most of the HQ parts come with the faster Iris Pro 5200 GPU.
Note that this isn't universally true, though. The i7-4702HQ and MQ and the i7-4700HQ and MQ all share the same GPUs, TDPs, and clock speeds; the only difference appears to be vPro support, which is present in the HQ models but not the MQ ones. Intel's reasoning here is beyond me, and moves like this are why its product lineup has grown as unruly as it has. The upshot is that you need to be paying close attention to the CPU you buy if you're looking for a particular feature.
Among the MQ and HQ families, most of the differences again come down to clock speed and TDP. Note that the HQ family chips with the faster Iris Pro GPUs have considerably slower base clock speeds than the MQ family chips; the extra processing power and eDRAM in these CPUs means that Intel has to ratchet back the CPU speeds a bit to stay within its stated power envelope.
A predictable upgrade, and that's mostly a good thing
Haswell is the sort of CPU upgrade we've come to expect from Intel: a whole bunch of incremental improvements over last year's model, all delivered basically on-time and as promised. Again, we'll need to have test systems in hand to verify all of the lofty claims that the company is making here, but at least on paper Haswell looks like a big push in the right directions. It increases GPU power to fight off Nvidia and AMD, and it decreases overall power consumption to better battle ARM.
Unfortunately, the parts that will actually be doing battle with ARM haven't even been announced yet. Powerful quad-core chips are all well and good, but if you're waiting to see what Intel has in store for Ultrabooks you're going to have to wait a bit longer. For our part, we'd guess that Computex will bring those CPUs along with a flood of accompanying Haswell Ultrabooks, convertibles, and tablets. Once we see those computers, we'll have a better idea of what Haswell can do that Ivy Bridge couldn't.
Exact retail availability for the processors has not yet been announced, but expect that information to be available in the coming weeks.