Hacker News Reader

Comment by pinewurst | Hacker News Reader

There's often a long lead time between writing up benchmarks/white papers and publication. At my last employer it could take well over a month to let everyone have their 2 cents and get through Legal for approval. This is a long enough interval to explain the old/new GROMACS usage.

As the saying goes, don't attribute to malice that which can be adequately explained by stupidity (including corporate bureaucracy).

View on HN

The author pointed out that they published benchmarks using the same test that had not been optimized for AMD yet. But they included that information in the publication, including on the graphs that would otherwise appear to show the outright superiority of Intel in that regard.

If the publication was delayed to allow time for due diligence then this disclosure should have made the cut. I like to think that Intel had some top-notch benchmarking nerds who were superseded by unscrupulous executives and would not have published a dishonest benchmark otherwise.

Also, the difference in sub-NUMA configuration actually favors AMD as well. NPS=4 is optimal for AMD, as shown by ServeTheHome's own work earlier. Both systems are tested in their proper (fastest) configuration.

Despite ServeTheHome's own previous work showing this, they're whining about it being different. But if Intel hadn't tested AMD in the proper configuration they would have complained about that too.

Yeah, you take first-party benchmarks with a big grain of salt like always, but STH is just looking to stir some shit here.

> Both systems are tested in their proper (fastest) configuration.

AMD is running at half the threads it can. I'm pretty sure that's not its proper nor fastest configuration.

oh really one thread per core is the optimum for EPYC server? I don't know why AMD care about hyper threading then.

We're not talking about the general case. One thread per core might be optimal for that specific workload.

BTW turns out that was a typo and they used two threads per core anyway.

I don't see where I said one thread per core was optimal for Epyc or not.

Unless that company is well-noted for their malice.

That's not to deny the possibility of both. :)

Haha true enough. I'm sure the truth in this case in somewhere in the middle. Reading any marketing material should always be done with a healthy dose of skepticism.

Reply on news.ycombinator.com