Haswell Architecture – CPU side:
At the higher level, Haswell architecture is mostly based on the evolution of what was introduced with Sandy Bridge and Ivy Bridge. However, on the lower level, Intel has tweaked a lot of stuff to increase IPC (instructions per clock) performance while keeping the power consumption in check. There are no major improvements in the key pipelines. However, Intel has improved branch prediction along with significantly increasing the size of buffers and data structures. The out-of-order window has also been widened which now feeds even more parallel instructions to the execution unit.
Furthermore, two new dispatch ports have been added, increasing the number from 6 to 8. This leads to the addition of another ALU, branch predictor and store address.
Intel has also added support for AVX2 instruction on the ISA side doubling the floating point throughput of the machine. AVX was first introduced with Sandy Bridge which doubles the single precision and double precision performance over Nehalem’s SSE instruction set. Now, that has been doubled once again. In order to accommodate twice the FP performance per clock, the cache hierarchy has been overhauled as well. This resulted in doubling of the bandwidth that is offered by L1 and L2 cache.
Another big change on the architectural side is the introduction of TSX instruction set. This IS helps developers with efficient development of highly multi-threaded programs as TSX improves the concurrency and multi-thread handling on the hardware level. Unfortunate part of this change is that Intel has selectively disabled it on some of the key SKUs including the K-series desktop parts.
Haswell Architecture – GPU side:
Haswell’s integrated graphics is probably the most talked about thing in the new processors. CPU side of the Haswell got fair bit of architectural changes but the performance improvements are not as big. On the GPU, its other way around. For most part, Haswell’s GPU is based on the foundation laid out by last year’s Ivy Bridge but the performance improvements are phenomenal and that’s mostly because Intel has increased GPU’s real-estate share on the die. Sandy Bridge die has around 17% of area for GPU; it gets bumped up to 27% on Ivy Bridge and with Haswell, it is increased to roughly around 32%. There are no major architectural changes made to the GPU this year. That’s reserved for Broadwell which is expected in 2015.
Intel has also introduced a new naming scheme for the IGP with Haswell. The Haswell GT1 class graphics are now simply known as Intel HD graphics, the Haswell GT2 class is now HD Graphics 4200/4400/4600. In addition to this, the new GT3 class graphics are introduced. GT3 class chips with 1.1 GHz or below are now named as HD Graphics 5000 and from 1.1 GHz to 1.3 GHz, its called Iris 5100. And finally, there is GT3e GPUs which is basically GT3 with on-die eDRAM. This is the fastest of the Haswell integrated graphics and it is known as Iris Pro 5200. New graphics also add support for DirectX 11, OpenCL 1.2 and OpenGL 4.0.
Haswell GPUs basically consists of two building blocks; non-slice and slice. It has been designed with modularity in mind. One non-slice part and a slice part is essential in every GPU. However, Intel can add another slice to considerably increase the performance and that’s exactly what has been done in the case of GT3 graphics. The GT1 and GT2 chips consists on one non-slice and one slice but in GT3, simple another slice is added in the series giving a considerable performance improvement.
Here is a table showing basic specifications of Haswell’s GPUs:
|Intel Haswell GPU Specifications|
|Cores/EUs||Max GPU Frequency|
|Intel Iris Pro 5100/5200||40||1300MHz|
|Intel HD Graphics 5000||40||1100MHz|
|Intel HD Graphics 4600||20||1350MHz|
Power management has been dramatically improved with Haswell and as a result, we should be seeing considerably better power efficiency. Major management changes include finer power gating, lower chipset power and swift transition between power states. The transition is said to be about 25% faster than the Ivy Bridge. This enables a more aggressive selection of power modes according to the load on the processor.
Furthermore, Intel is also introducing a thing called FIVR. That’s short for Fully Integrated Voltage Regulator. It is a combination of some on-package and on-die circuitry which takes the task of voltage regulation from the motherboard. Haswell CPUs take a single voltage rail (with around 1.8v at default) from the motherboard and then FIVR provides it to the individual components on the chip. This helps Intel with the aggressive power gating and finer voltage control.
Moving the voltage regulation to the CPU saves some real estate of the motherboard as well. It also decreases the motherboard manufacturing costs.
All of this truly shines in mobile form factors only but desktops are also seeing some benefits.