39 ms·
I wish we had more details about the on-die internconnect ( Hypertransport successor) in terms of latency and bandwidth, and even topology. We run a NUMA .. c
by drewg123 9y ago
I wish we had more details about the on-die internconnect ( Hypertransport successor) in terms of latency and bandwidth, and even topology. We run a NUMA .. challenged .. OS, and depending on the interconnect that may or may not matter so much for our workload.
- valarauca1 9y ago<rumor> Initial leaks point to a PCIe 3.0 64x interconnect </rumor>
- dom0 9y ago^ Well we do already know that the interconnect between modules uses the same PHYs as PCIe 3.0.
- drewg123 9y agoSure.. but I'm wondering about the on-die interconnect between the 4 different packages that share the die. Eg, everything seems like an EPYC is 4 Ryzen packages, and a "thread ripper" is 2 Ryzen packages. So it seems logical that there is something connecting those 2 or 4 packages, and that fabric as bandwidth and latency characteristics. If it is somehow infinitely fast, then that's great for me.
- lightedman 9y agoActual verification shows up to 128 PCI-E lanes.