oleotiger

Here is one situation that I wanna to collect metrics with pcm tools when I run a benchmark. And running the benchmark is really time consuming. If pcm tools can...

According to the wiki [PCM Column Names Decoder Ring](https://www.intel.com/content/www/us/en/developer/articles/technical/intel-pcm-column-names-decoder-ring.html), `TIME(ticks)` here means `Number of invariant clockticks`. As I set time interval as 1s, there should be 3G ticks during per...

So the value of TIME(ticks) should be multiplied by 1 million? Or the domain should be changed to TIME(1e6 ticks)?

Should the value of `INST` be multiplied by 1million as well? In my experiment, the number of instructions is far blow my desired value.

Exporting conf of parent node: ``` [json:timescaledb_instance] enabled = yes destination = localhost:14866 remote write URL path = /write data source = as collected prefix = netdata update every =...

I reproced it. Error log of netdata: ``` 2021-01-27 10:30:15: netdata ERROR : MAIN : EXPORTING: failed to write data to 'localhost:14866'. Willing to write 3289916 bytes, wrote 2154360 bytes....

8张卡成本很高了，我们手里只有两张A100。 我们通过以下方式测试机器性能，减少Transformer的层数，tp改为2，保证模型size能放到两张A100中运行的。 1、 但是模型不能支持cuda unified memory吗?哪怕性能下降，也至少保证正确运行？ 2、intel新出的spr，epr处理器都支持AMX，ArmV9也支持SVE、SME，都可以做高效矩阵乘，有些超算CPU也挂有HBM，CPU在大模型推理上开始有一定的优势，这里能不能出个CPU版本？