Intel’s Haswell is a bit of a puzzle. On the one hand, this is the fastest single-threaded chip in the world — but on the other, it’s hard to get excited about a chip that’s only a few percent faster, consumes more power, and has weaker overclocking potential than its predecessor. Where did it all go wrong for the master chip maker?
To understand why Haswell is so lackluster, we have to go back to the beginning, to Haswell’s inception a few years ago. Right from the outset, the Haswell microarchitecture was designed from the ground up to be a mobile-focused chip. While there are some intrinsic boosts to baseline performance, the matter of the fact is that Haswell simply wasn’t designed for high-load and overclocking scenarios.
One of the biggest additions in Haswell is the full integration of the voltage regulator onto the CPU die. Instead of relying on the motherboard to produce a variety of power rails for different parts of the chip (memory controller, PCIe controller, GPU), Haswell has a fully-integrated voltage regulator (FIVR) that takes a single input (~1.8V) and splits it into all the required rails. The end result is an overall reduction in system power (~20%), comparable idle CPU/GPU power, but a sizable increase in load power consumption (~10%).
This obviously makes Haswell ideal for mobile computing, where there’s a lot of idling and puttering around the interwebs. These power savings are so extensive that they could equate to a 25% battery life improvement over Ivy Bridge. On the desktop, though, where you might be playing games, editing photos, or encoding videos, Haswell only just scrapes a victory from Ivy Bridge.
The fully-integrated voltage regulator and higher power consumption under load mean that Haswell is actually worse than Ivy Bridge at overclocking. As our hardware analyst mentioned in his Haswell review, Intel’s new parts struggle to get past 4.5GHz on air, while Ivy Bridge could reliably hit 4.7GHz, with some parts reaching 4.9GHz. In reality, the picture is even muddier than that: Early reports suggest that some Haswell chips can only reach 4.3GHz, while others can get to 4.7GHz or higher (again, on air).
If we look at the bigger picture, though, overclocking has been slowly dying for a decade. For the most part, every process node shrink, and the accompanying die shrink, reduces overclockability. This is just a fundamental restriction of physics: As components get smaller, not only does transistor density increase (consuming more power) but there is also less surface area to radiate heat. If you take two theoretical 3.5GHz chips, the one with the larger die size is going to be easier to cool, and thus capable of reaching higher overclocks.
Ivy Bridge and Haswell are both based on Intel’s 22nm FinFET process, but the integrated voltage regulator exacerbates the issue. It also seems that Haswell’s built-in thermal throttling is much more aggressive than Ivy Bridge: Where the Core i7-3770K is happy to sit at 3.7GHz under full load at 90C, the Core i7-4770K throttles back to 3.5GHz within moments of starting Prime95.
Given the wide range of reported overclocks, it would seem that there are some yield issues at play, too. Basically, if you want to overclock Haswell, pray for a good chip — and remember, the power premium (100 watts or more), for maybe 20% more performance, is the largest yet for any of Intel’s chips.
Back to reality
Overclocking aside, though, it’s important to remember that Haswell is still the fastest processor that Intel has ever produced. For the same price as an Ivy Bridge chip, you get around 10% more performance. The overclocking proposition might not be quite as enticing, but in reality Haswell’s 10% clock-for-clock advantage over Ivy Bridge means that a Haswell-based system at 4.5GHz should still beat out IVB at 4.9GHz.
It’s also very important to remember that Haswell is a much more advanced chip than Ivy Bridge. In terms of transistor counts, disregarding any changes to the GPU, a quad-core Haswell CPU has roughly 200 million more transistors than a quad-core Ivy Bridge CPU (1.4 billion vs. 1.2 billion). A lot of these transistors were spent on increasing the chip’s IPC (instructions-per-clock) by adding more execution resources and beefing up out-of-order execution (OoOE) capabilities, but most of them are dedicated to brand new features such as AVX2, FMA3, and TSX. These are very powerful features, but for the most part they only boost performance when software has been specifically written/compiled to take advantage of them. Today, we are merely seeing the ~10% speed-up provided by Haswell’s reworked execution core; tomorrow, when software uses AVX2, FMA3, and TSX, the speed-up could be 25% or more.
In this light, Haswell is a monster of a chip. Intel has produced a chip that decimates power usage on the low end — which, let’s be honest, is the market that Intel is really interested in — and yet will also provide a huge speed-up for power users once software and compilers are updated. It’s easy to be disappointed by Haswell’s performance on paper, but remember: There really is no other chip out there, especially from AMD, that comes anywhere close. Still, between the death of overclocking, Intel’s focus on mobile, and the shift to soldered-on chips, it’s clear that the writing is on the wall for desktop PC power users.