That effort takes the form of Intel's most powerful graphics core yet. There are lots of different versions of the new graphics core, but the underlying technology hasn't changed much. What has changed is the amount of silicon Intel is willing to spend on graphics. It's gone up. Massively.
The quickest existing Sandy Bridge version of Intel's HD Graphics tops out at 16 execution units and 1,150Mhz. With Haswell, that explodes to 40 units and 1,300MHz. Well over double the raw computational performance, in other words.
But wait, there's more. Intel has also added an optional 128MB slab of eDRAM to Haswell. It's a separate slice of silicon, but fitted to the same package as the rest of the processor and actually operates as a shared L4 cache for both the graphics and the CPU cores. But one of its key roles is offsetting perhaps the biggest bottleneck for integrated graphics performance. Bandwidth.
Exactly how much difference it makes to performance is something we'll have to wait and see. We don't yet have our hands on a Haswell processor with the eDRAM chip. But even on paper, it remains a little off the pace of a mid-range discrete graphics chip for raw memory bandwidth.
Anywho, there are a metric tonne of slightly different graphics cores with Haswell. That eDRAM effort with 40 execution units is one of two new cores bequeathed to us via the new Intel Iris brand. So, that's Intel Iris Pro 5200. Take away the eDRAM but keep all 40 units and you have Intel Iris 5100. Keep the 40 units again but lower the clocks a bit from 1,300MHz to 1,100MHz and the Iris bit is dropped in favour of plain old Intel HD Graphics 5000 branding.
Drop to 20 cores, but up the clocks to 1,350MHz and you have Intel HD Graphics 5000, which is what's in all three chips we've so far test. There are further variations, but we suspect you're beginning to lose the will to live. No, we don't know who makes this nonsense up, but we can only assume their primary aim is to make the branding as baffling as possible. At the very least, why on earth aren't all the 40 unit cores called Iris?
Haswell architecture
But never mind, there's a pretty siginificant part of Haswell we've yet to discuss. Yup, it's the CPU cores themselves. It's indicative of Intel's priorities that the cores come last in this discussion and that there's actually not all that much to say.
Yup, there's a list of tweaks we could go through in forensic detail. Would could explain, for instance, how Intel has massaged Haswell's front end (ooer!) for better branch prediction. Intel has also improved Haswell's instruction level parallelism (ILP) courtesy of upping the szie of various data-level structures, including the out-of-order window and the in-flight loads and stores. Intel has also added a pair of execution ports, bringing the total to eight, and added the obligatory extra floating point instructions, known as AVX2.
But it's still pretty much the same four-wide execution engine as before with the same 14 to 19-stage pipeline as Ivy Bridge. And that means that the vast majority of the time, it performs a very similar amount of work each clock cycle.
Which really leaves clockspeeds and cores as the best way to improve performance. Except Intel hasn't done that. The top Intel Core i7-4770K has the same 3.5 default frequency as its Core i7-3770K predecessor. And the Core i7-2700K that came before that.
Nor has the core count changed. We're still talking about a maximum of four cores for Intel's mainstream socket (which incidentally has changed again, more on which in a moment). All of which means you have to go all the way back to the likes of the In fact, you have to go back to the likes of the Intel Core i7-875K almost exactly three years ago to find a comparable chip from Intel that offers tangibly less performance.
