3 ms·
AMD's efficiency cores are much more similar to their mainline cores than for either Intel or ARM. Zen 4c cores have the exact same pattern of transistor conne
by Symmetry 3y ago
AMD's efficiency cores are much more similar to their mainline cores than for either Intel or ARM. Zen 4c cores have the exact same pattern of transistor connections as a regular Zen 4 core, it's just that the transistors much more of a consistently low width and so laid out differently, with those transistors that have to push signals long distances or to many consumers not allowed to grow as much. This means you can't drive the clock rates as high, but at the benefit of a smaller core and generally lower power usage.
- alphablended 3y ago[dead]
- rewmie 3y ago> AMD's efficiency cores are much more similar to their mainline cores than for either Intel or ARM. I apologize for going off on a tangent, but now that we have the benefit of hindsight what was the verdict on AMD's Bulldozer microarchitecture? I recall that Intel invested a great deal on its anti-Bulldozer smear campaign with their "no true cores" accusations, but nowadays it seems that cheaper non-performance many-cores are the winning recipe. https://en.wikipedia.org/wiki/Bulldozer_(microarchitecture) https://en.wikipedia.org/wiki/Bulldozer_(microarchitecture)
- The_Colonel 3y agoCompanies now do mix of fast cores and slow cores. The problem of Bulldozer was that it only had slow cores.
- rewmie 3y ago> The problem of Bulldozer was that it only had slow cores. I don't think that was the case. Bulldozer shared the FPU and cache, and if I recall correctly it was proven that the shared FPU was rarely a bottleneck because even in compute-intensive operations the number of FPU operations was relatively low.
- The_Colonel 3y agoThat's not what I was hinting at. Bulldozer was slower than its predecessor Phenom in single-threaded tasks where shared FPU/cache shouldn't have a detrimental effect. IIRC AMD made a trade-off to focus more on the multi-thread performance at the cost of single-thread performance (so basically more slower cores). But then many tasks are largely single-threaded and at that time the parallelization efforts were still very young.