Originally Posted by Liranan
eDRAM is video RAM, not general use cache.
eDRAM is L4, accessed by both the CPU and GPU. It's nothing more than a tier in the memory hierarchy between RAM and L3. Designed with the integrated graphics in mind, sure, but it still benefits the CPU. I imagine it helps with the Xeon E3-1200 v4 series to some degree as, say, web or file servers. More L3 would be expensive, but a DRAM die isn't too much extra.
Assuming all else is equal, it really depends. For smaller files that can fit within 60MiB of memory, the Xeon will win since its L3 ought to be faster. For files between 60MiB and 128MiB, the i7 should win, as it has access to L4. For files beyond that, it should be about equal (if we use similar memory configurations of course).
Or so you'd think. In reality, it's never that simple. Even adding L3 isn't a sure fire way to increase performance thanks to how cache works. A friend explained it to me (and 90% of it went over my head) but it has to do with how cache is accessed and I think "predictions" of where some data might be.
Techreport did some synthetic benchmarks
as well and includes some interesting results. Skylake and Haswell have similar latencies, but Broadwell and its L4 show a significant gap where its L4 would be most effective, compensating for its reduced L3 (6MiB rather than 8MiB). Keep in mind that it's a logarithmic scale, so the L4 is accessed in about half the cycles compared to Skylake and Haswell needing to access system memory, probably because it's on the CPU package rather than through the motherboard. If you look at the very first graph, there's a very tiny gap where the 5775C edges out vs the 5960X. But compared to a processor with three times the L3, I'd assume that even the L4 couldn't help it and the E7 would win hands down.
Well, except price. $7000 for a quad-core is a bit much, don't you think?
EDIT: Also, for some history and explanation of cache, read this
. It goes into cache hits and misses as well as associativity.
EDIT 2: Neat, part 2 here
. The die shots really emphasize just how much space is dedicated to cache on die and why L3 is so expensive. My Xeon (Haswell GT2, same as the 4790K) contains exactly 9 699 328 bytes of SRAM acting as cache - the vast majority being L3 - made up of 465 567 744 transistors (~465 million). The die itself contains 1.4 billion in total, meaning about a third of the transistors on this CPU are dedicated to cache. And given that a significant part of the die goes to the integrated graphics, I'd guesstimate that about half of the CPU portion of the die is cache.Edited by CynicalUnicorn - 8/6/16 at 9:24pm