What is the price difference?
And how productive are said engineers that a theoretical 10% increase in cpu performance can make up for the price difference or not?
-----
This is dramatically oversimplifying things a bit but just think of EDram as spending more money on faster speeds of ddr4 memory, and the benefits of ddr4 memory over ddr3 memory. The technologies are not quite the same but the goal of barely faster latency (the number is lower), while seeing real world increases in bandwidth is effectively the same even though we are talking about different techs on how to achieve this goal.
I do not claim to be an expert in this. I guarantee many people here can school me on this with a better understanding of how the cpu works and a more detailed computer science background, regardless my goal is to just talk simply.
-----
In most situations you get little speed increase for going with the edram and often it is less than 10%
with the exception of graphics. This is because the latency between edram is not much faster than ddr4 (see chart below). In addition the bandwidth difference between ddr4 and edram is shrinking.
This above chart is talking about memory latency and not bandwidth. Notice the latency even with the 5775C in purple is comparable with most ranges with the other CPUs this is due to the edram. Now the chart shown above is using the 2133 memory for DDR4 Skylake and I forgot which timings, if you now how to tweak the timings of your ram you can bring it closer to the rest of the chart.
The benefit of edram besides a slightly smaller faster latency (which is really not that much) is that you double the memory bandwidth for its 50 GB/s bidirectional vs a 20 to 30 GB/s bandwidth with DDR4 memory depending on what the speed of the memory is when you run the modules in dual channel mode.
If you ever read a memory speed article you will see the improvement of using faster memory for cpu tasks is almost always less than 10%, and is often less than 4% in specific workloads that benefit from faster memory. In other workloads there is less than a 1% improvement and that is to such a small number that the testing methodology will introduce at least a 1% margin of error.
Remember you are not getting the faster memory bandwidth overall but just in stuff that can fit in that 128mb l4 cache buffer. If the workload is bigger than that 128mb buffer then you will still be waiting to retrieve data from your RAM. Now 128mb is actually a big deal for some things for while your program that you are running may be measured in GBs what you are actually modifying on the fly is much smaller and with intelligent enough prefetching you can put the stuff you are modifying in the faster edram.
-----
Now memory bandwidth is much more of a big deal when you are dealing not with quad cores but server parts where you can see up to 24 cores per an individual single cpu, and even more cores once you go multiple sockets and thus multiple cpus. This is why you often see quad channel memory in servers. Even then the swap from ddr3 memory to ddr4 memory is the goal is often not trying to improve CPU instructions per clock but instead lowering the voltage and causing power savings as well as higher density of GBs per ram module.
-----
Please remember I do not claim to be an expert, so for all the people here that are smarter than me on the nuts and bolts of how the cpu works underneath it all, if I misspoke with anything feel free to correct me but just use a nice tone please.
Facebook
Twitter