How is the "e-core" scheduling going these days? I am not an Intel guy, but I remember reading about it causing a lot of issues for multicore workloads and people having to explicitly pin processes away from those cores.
There are interesting core scheduling questions with a CPU this large, considering its 4-way-shared L2 design and profusion of thermal domains, however these issues are so complex that no system I've ever heard of — certainly not Linux or Windows — attempts to optimize it.
I wonder how they'd work for AI compared to a GPU? That's all about memory bandwidth so I'm guessing the GPU would still win unless it was an AI type that is more compute-bound.
I did see mention of Ponte Vecchio.
It seems like Intel is getting closer to its Larrabee vision of massive MP low power chips (which were 486 cores in the shelved 2010 project).
While Intel's E cores share some history with the Atom line, they have almost no resemblance to the Atom line you might remember from the netbook days. In fact, Intel hasnt made a consumer Atom in a decade (there are still Atom processors for the embedded market).
The first E cores came out a few years ago with performance similar to Skylake and performance has only increased from there. Intel's first Xeon E core processor from last year even outperformed the some of the big core Xeons from the year before: https://www.phoronix.com/review/intel-xeon-6780e-6766e/10
The evolution of the Atom cores has been continuous in the series ... Silvermont, Airmont, Goldmont, Golmont Plus, Tremont (also included as E-core inside Intel Lakefield), Gracemont (from Alder Lake/Raptor Lake/Alder Lake N/Amston Lake), Crestmont (from Meteor Lake/Sierra Forest), Skymont (Lunar Lake/Arrow Lake), Darkmont (Clearwater Forest).
All the cores of this series have been designed in succession by the same team and the improvements between them have been incremental. These cores have always been intended as competitors for various Arm Cortex-A cores, e.g. Goldmont Plus for Cortex-A75, Tremont for Cortex-A76, Gracemont for Cortex-A78, and now Darkmont/Skymont for Neoverse V3/Cortex-X4.
Before Tremont, these Atom cores were used only in homogeneous CPUs. In recent years only those sold for industrial applications have retained the brand "Atom". The most recent CPUs branded as "Atom" have been launched last year.
For now, it does not matter whether one refers to any of these cores as "Atom" cores or as "E-cores", because either name refers unambiguously to the same cores. "Atom" seems preferable, because it is more clearly Intel-specific. One could refer as "efficient cores" to cores in CPUs made by others.
The first E-cores were the Tremont cores of Intel Lakefield, which was launched in Q2 2020, which were paired with Ice Lake cores, and which had a performance much less than Skylake cores. Only the next generation of Atom cores a.k.a. E-cores, Gracemont, has reached performances comparable with Skylake. While Gracemont was first launched in Alder Lake, later it has become widely available in Atom-branded CPUs.
Anyway as you note these are excellent cores, and having 576 slightly wimpy cores in a box is an idea worth considering.