Lynnfield to Haswell upgrade

The long awaited platform upgrade


4 years without a platform upgrade is really unprecedented for an hardware enthusiast. But the decision to upgrade aren't because of a new killer application that requires more performance. It's simply the enthusiast in me itching to build another system. And what's a better excuse when your significant other's PC died and you offered to pass your Lynnfield to her. :)

I also took the chance to change the chassis and jump on to the AIO watercooling wagon, which present its own issues that will be explained in a future post.
Spending close to 800SGD, I got:

• Faster CPU (i5 750 to i5 4670k)
• PCIE 3.0 ( won't see any improvement until next GPU upgrade)
• SATA 3.0
• USB 3.0
• UEFI bios
• Faster boot (due to UEFI bios and a little boost from SATA 3.0)

Performance improvement in bench is pretty much what I would expect from a single generation jump. In case it slips your mind, Lynnfield is actually 3 generation older than Haswell. Other than encoding and photo editing task, no other task feels any faster than a Lynnfield. General usage and experience did improved a little due to the removal of SATA 2.0 bottleneck on my SSD.

Lynfield runs on 2.66 stock and I get to overclock it to 3.6 ghz. Haswell runs on 3.4ghz stock but I currently sitting on 4ghz due to excessive amount of heat and power required to drive it above 4ghz. After 2 full node shrink, overclocking did not significantly improve. If you have been following hardware development, you would realize that 22nm is a step backwards in terms of overclocking. This has nothing to do with Intel manufacturing capability and node design but simply the laws of physics stepping in. It simply means subsequent node shrink is going to be worse as more transistors is packed into a ever smaller physical space.

That's right. After 4 years, 3 architecture generation apart and 2 full node shrink, performance of CPU have not improved significantly. If this isn't the clearest indicator that silicon based computing device have hits its wall, I don't know what is. And Intel knows this, Haswell is build from ground up to target mobile devices to stop pursuing higher performance. Its successor will be BGA packaged only, since no one will be bothered to upgrade their ivy bridge or haswell to it for another 5% performance improvement.

Another disappointment is the number of cores. Due to lack of market competition, Intel have no incentive to increase the amount of CPU. Remember the days manufacturers uses 2 die to aggressively increase CPU cores? At 2009, when I first got my Lynnfield, I would have expect a minimum of octa core when I upgrade in the future. But I've never been so wrong. Even with a native octa core design in the server space for a while now, Intel have no intention of releasing them to consumer as it would push the prices of quad core down, thus reducing their margin. Even a hexa core is limited to a server platform masqueraded as a enthusiast platform (with outdated chipset due to how server platform is updated).

Core count got stagnant. CPU cache however, got reduced. A 45nm CPU can squeeze 8MB of cache while a 22nm reduces is 6mb? Now I know how Intel's margin keeps going up, even in such sluggish economy and demand...

In fact, I personally believe Moore's law will be obsolete in the next few years. Die shrink requires money, lots of it. As IT industry is aware of the performance wall on traditional computing environment, they have been expanding aggressively into mobile, which is also aided by consumer demand. In fact, after being neglected for years, the current mobile push have brought many innovations of desktop CPU into the mobile space.

One thing a node shrink is still good for at the moment, is reducing heat and power consumption at the lower clock speed spectrum. This greatly benefitted the mobile space as ever bigger design can be validated to fit the mobile device power and space requirements. But how long can this last? I will give it a couple of years before it hit the similar wall the consumer PC is facing.

We will all reach the "fast enough" performance level on smart phones and tablets and no one will have the incentive to upgrade. At that point, even Intel wont have the money nor motivation to do another node shrink. Even before that, I personally doubt the low margin mobile CPU industry is able to singlehandedly fund die shrinks at similar pace as the past decade.

Consumer PC market is totally stagnant at this moment. The only innovation on the horizon is AMD's HSA and HUMA. Being reduced to a niche player and finances in the red, AMD is in a limited position to aggressively promote this computing model to developers. After the release of hardware, it would still take years for developers to take full advantage of it. Until then, I will stick with Haswell platform for probably many more years than what I did with Lynnfield.

In conclusion, the PC hardware enthusiast is nothing but dead. Its about time I find myself a new hobby.