2 ms·
It doesn't seem any worse than existing NUMA systems today, where memory latency depends on what core you're running on. In contrast, the proposed system would
by loeg 6y ago
It doesn't seem any worse than existing NUMA systems today, where memory latency depends on what core you're running on. In contrast, the proposed system would have the same performance for on-board vs plugged DIMM regardless of which CPU is accessing it, which simplifies scheduling — from a scheduling perspective, it's all the same. I think that's easier to work with than e.g. Zen1 NUMA systems.