Originally Posted by Blameless
Economics. While eDRAM isn't particularly expensive, transistor for transistor, it's still far from free. eDRAM controller and substrate/interposer complexity adds to that.
It's a choice between more logic or another cache level, and for most workloads, more logic will win.
We'll probably see eDRAM in x86 CPUs again, but I'd expect it to primarily remain an aid to IGPs, even if it has some side benefits as a L4 cache.
Just to clarify that the substrate is the part the die is connected to and an interposer, if used, is an adapter between the substrate and mainboard socket or connection. At least that's always how I've interpreted it.
Does seem it's primary function was to support the integrated graphics and I still think that the eDRAM on subsequent families such as Broadwell and Skylake it is still the same 22nm device although designed to work a little differently for Skylake. Hard to find results for Skylake though, maybe if the processors such as the 6785R weren't locked. I wouldn't be surprised if the non-k overclock left the eDRAM disabled. So costs should be lower by re-purposing the Haswell eDRAM however as DRAM gets faster I suppose there is less advantage of such a device, especially if it is still the old 22nm one. Also some software will benefit more than others, Winrar shows this well. An example with an i7-4980HQ running the same clocks except for eDRAM, and RAM timings. 125MHz strap used and DRAM set to 8-11-11-29-102-2 at 1113MHz (2226MT/s)
So around 30% increase with default eDRAM and another 13% or so for overclocking the eDRAM by 33%. Would have expected nearer 10% rather than 13% but those are the results.
Originally Posted by 8051
I really wish I had bought a 5775c system instead of my 5820k system.
Me too, I waited for launch day for the 5775C then waited some more and some more again. It never sold here, it was like it had never existed.
You mention X99 and someone else mentioned about quad channel bandwidth. I too have an X99 and bandwidth only seems to be realized when running multiple threads. Single thread bandwidth is shocking, around 10GB/s for 2133MT/s quad channel. This can be seen with maxxmem2 3.0 http://maxxpi2.de/pages/results/top-...gs---ddr-3.php
although cache has an effect on reads but write speed would be near the mark. Because of the cache effect you might find 5775C's can take top places for DDR3 because of the eDRAM, well until the author deletes them.
Originally Posted by rluker5
I saw a 5950hq for $164 and a 4980hq for $160 and jumped on them. They were leftover ES samples that apparently...
When was that? I bought my i7-4980HQ end of 2018 when they were selling around the $100 mark and it's a production CPU although it might well be secondhand. It can be run fully unlocked via not running microcode patching until later on but I have microcode 0x17 modded into the BIOS, it's the same length as one of the unneeded patches so makes it an easier replacement. The 1.3V problem sounds like the FIVR over voltage protection isn't being disabled. Disabling OVP for vcore/vcache should see the limit at 1.925V which above that sets voltage back to default VID, well at least it does that on a couple of cheapo Haswell desktop CPU's I tried. Still, should be able to do a lot with 1.3V so not a huge loss.