Leo wrote: ↑Mon Aug 19, 2019 5:47 pm
What does I/O and memory channel mean? How does it benefit chess engines?
I/O is the number of PCIe connections to the CPU. Each PCIe 3.0 "lane" connection supports 985 MB/s to... well... anything. Ethernet, GPUs, NVMe SSDs, etc. etc. On modern computers, you use x2 or x4 connections to SSDs (1.9GB/s and 3.9GB/s respectively), and x8 or x16 connections to GPUs (15.7GB/s). The M.2 slots usually use x4 connections.
Note that some CPUs support PCIe 4.0, which doubles the bandwidth. However, GPUs and SSDs which use PCIe 4.0 are extremely new and extremely expensive, so the typical CPU-builder will probably stick with PCIe 3.0 parts.
Your typical CPU (Ryzen 7 or Intel i7) will only support x16 lanes + x4 lanes to the "southbridge". The "southbridge" is all of the miscellaneous features of your motherboard lies (USB connections, SATA, maybe a legacy PCIe 2.0 or even PCI / ISA port). Some motherboards lay out the NVMe drives "behind" the southbridge, which is annoying... to say the least.
A high-end CPU like Threadripper supports x64 lanes of I/O, allowing stupifying amounts of I/O bandwidth:
https://www.guru3d.com/news-story/eight ... 8-gbs.html
-------------
DDR4 Memory channels is similar: its the number of DDR4 channels that are supported. DDR4 is clocked at different rates, but typical DDR4 is 2600 MT/s, or roughly 21GB/s. Two channels of 2600 RAM will give you 42GB/s, while Threadripper supports 4x channels, or 84GB/s.
Note that these memory channels work in parallel. Which means you only really access the full bandwidth if you use multiple threads. A single core typically only has a fraction of the memory bandwidth. This depends on details, such as NUMA architecture and microarchitectural details (in particular: cache hierarchy, load/store units, etc. etc.). Its
very difficult to write a program to use all memory bandwidth available, you pretty much have to go AVX2 and very carefully lay out your memory accesses to work with the prefetchers. But more-bandwidth will (in general) help out with multi-threaded programs.
When you have 16x cores or 32x threads running on one CPU, you'll probably start to run out of memory-bandwidth.
---------------
I/O could benefit a chess engine if you use large amounts of GPUs, SSD drives, or Network (Ethernet) connections. You could feasibly put 4x GPUs (in 8x lane configuration) + RAID0 4x NVMe SSDs + 10GbE on a Threadripper for instance, giving 4x GPUs worth of compute and maybe 10GB/s read/write speed to your Tablebase.
Memory bandwidth is far more difficult to figure out, because that's engine and data-structure specific. The easiest way to take advantgae of memory bandwidth is to increase the number of threads operating on your program.
EDIT: Actually, both I/O and Memory are very difficult to fully take advantage of. Although you have very high bandwidth available on modern machines, I/O has huge latency (and even DDR4 to an extent). So you need to learn how to write asynchronous programs, or maybe heavily multithreaded programs, to fully utilize the I/O and Memory bandwidths available today.