Last year, Intel (INTC +1.14%) announced that it had designed a special embedded memory to be paired on package with its highest-performing Haswell notebook processor. In a nutshell, this on-package memory improves chip performance, particularly in memory bandwidth-hungry graphics workloads.
It has been reported that Intel plans to deploy this eDRAM across its notebook processors, but I'm not convinced the company's eDRAM ambitions end with high-end consumer notebooks. I believe Intel will eventually bring this technology to its Xeon server processors.
Modifying IBM's playbook a bit
According to microprocessor analyst David Kanter, IBM (IBM 0.58%) is "the biggest proponent by far" of using eDRAM, employing the technology as "massive last level caches" in its server processors "to improve array density." The idea here is that one can pack much more memory into a given area with an eDRAM built on the same process generation as a comparable SRAM. The downside, though, is that SRAM is generally faster than eDRAM.
I'm not convinced Intel would replace the less-dense-but-faster SRAM-based L3 caches in its Xeon processors with denser-but-slower eDRAM, but I like the idea I have seen floating around that the chipmaker could sell variants of its Xeon processors with lots of eDRAM on package. This would be similar to what is being done with the high-end laptop chips, but on a much larger scale.
A pretty nice n-1/n-2 technology fab filler
Intel tries to transition to new manufacturing technologies as quickly as possible. That said, Intel has made it clear that any given technology generation typically lasts 10 years, which means the company needs to find product to keep those older factories full.
One nice way to sustain utilization of these older factories, as Kanter and many others have suggested, is to use them to build eDRAM. With the Skylake processor generation, leaked Intel plans suggest that even the higher-end 15-watt and 28-watt Ultrabook chips (think MacBook Air/MacBook Pro) will get on-package eDRAM. This definitely seems like a long-term play.
About that average selling price increase
According to Kanter, Intel sells its 128MB eDRAM chips for approximately $50. For server applications, Intel could sell multiple of these 128MB chips or build a larger eDRAM, potentially allowing it to generate incremental revenue per unit. This would increase the average selling price per unit of its server chips, ultimately growing revenue.
I imagine such a solution would be too expensive and/or esoteric for Intel's more mainstream Xeon E5 family of processors. However, for the company's ultra-high-priced, high-margin Xeon E7 processors, a few very high-end stock-keeping units with lots of eDRAM might do quite well in the marketplace.
