Figure 1: (top) Multi-chip Package, back view of Intel GT3e GPU containing eDRAM and the Intel Haswell processor; (bottom) A stitched image showing the SEM cross-section of the multi-chip package. Click here for larger image. (Source: Techinsights)
Refresh seems to consume about 0.5W. The L4 latency is about half that of DDR channels, and the bandwidth about the same as 4 DDR4. Typical hit rate 95% (that would be remarkably good in general, but maybe the gamer market is what they aim at where locality might be better).
Not clear this is better than a Memory Cube. But probably much cheaper.
The DRAM will leak charge without a refresh, which requires power frequently every second. SRAM doesn't. Maybe SRAM doesn't sleep well enough anymore to conserve standby continuous power? What happened to near-threshold operation?
Interesting. The article only mentions "The eDRAM also required 1/5 of the keep-alive power compared with an SRAM device." Could you explain more about eDRAM's constant refresh and is that something that would be fixed someday, do you think?