When we wrote that Haswell was a graphics monster a few months ago, we mentioned there was more to the story than that. What we couldn’t tell you about at the time, SemiAccurate values the fingers of their moles more than a headline, was a technology called Crystalwell (or possibly Crystal Well).
Crystalwell is not L4 cache on package as many have been surmising, it is more GPU memory on an interposer. It looks like Ivy Bridge won’t be the first time this technology is productized, Haswell now has that honor. We are told the GT3 variants of Haswell will have 64MB of on-package memory connected through an ultra-wide bus. While it won’t necessarily eliminate memory bandwidth problems, it will go a long long way toward minimizing the problem.
Remember when we said something was missing from Haswell? That was Crystalwell, and it very likely reduces the pin count of the CPU by a lot. Less memory channels means less pins. Less I/O power means less power pins. Less pins means lower package costs and routing complexity, offset by the huge interposer costs for now, but it is progress. There is even enough memory for on-die frame-buffers, lowering power use during low CPU C-states, saving lots of power too. Less power means less cooling, smaller batteries and lower BoM costs for a laptop. In general, there are substantial savings that will offset a lot of the added cost.
In the end, the massive bandwidth, coupled with the 5x increase in shader performance, will mean Haswell is a real graphics monster. 5x the shader performance is only the beginning of the story, memory bandwidth will likely have a larger impact than the shaders in the real world. If the drivers work. Luckily for AMD and Nvidia, there is a 10 year record of that not happening at Intel.S|A