3 ms·
The Intel ISA has some very powerful instructions that allow it to keep performance up, but unfortunately decoding n x86 instructions involves doing a number of
by pclmulqdq 6y ago
The Intel ISA has some very powerful instructions that allow it to keep performance up, but unfortunately decoding n x86 instructions involves doing a number of things that are O(n) in latency and O(n^2) in chip area/power - both instruction boundary detection and micro-op scheduling (since instructions can emit more than one uop) are included here. This isn't too bad when n is 4, but it has stressed both AMD and Intel's capabilities to make it to 8, and I don't think anyone would ever seriously consider building a 16-wide x86 machine. Once memory and I/O buses get fast enough to feed a 16-wide general-purpose computer, x86 as an architecture is in trouble. They have quietly gotten fast enough to feed an 8-wide machine.
LPDDR4 very quietly took the speed crown from DDR4 due to the fact that the signal integrity problems are less demanding (soldered vs DIMM being the standard interface), and it looks like the DDR interface may be near the end of its useful life. Transceiver-based memory modules using standards like CXL and CCIX may be the long-term fix to the problems of DDR4. In the short-term, those are going to hurt, but long-term they should be a lot better.
- gpderetta 6y agoCommon knowledge is that going wider quickly hits diminishing returns so generally there hasn't been a lot of pressure for x86 to go wider. It seems to be working for Apple though possibly because they are targeting differen trade offs (lower clock speed for example); but maybe their secret is a better modelling of common workloads revealing this common knowledge to be outdated.
- ece 6y agoI'd like to see this explored more. I think AMD/Intel might be historically looking at C++/compiled workloads more than JS/Java/python interpreted workloads. JITs very well could take better advantage of wider cores than compiled code. Diminishing returns with increasing cache sizes and going wider doesn't seem to hold back the M1.
- gpderetta 6y agoMaybe. JITs and runtimes in general might add some "overhead" code that does book-keeping and verifies assumptions (to rollback conditional optimistic optimizations). Maybe wider engines allow running this overhead without impacting actual runtime. It probably also help hide the cost of type checks and bound checks.
- ece 6y agoIt might just be the combination of better branch prediction, TSMC 5nm, 8-channel memory, 16k pages, along with 8-wide cores+larger structures especially cache are enough to account for the more perf/watt of the M1.