Xeons 8461V/6338, EPYC 7R13, AWS Graviton tested by minio on erasure encoding perf

[gigatexal]

Erasure Coding Impact on CPU Utilization in MinIO

The answer to the burning question you always wanted to ask - how does erasure coding utilize CPU?

blog.min.io

Seems the intel chips are rather performant here where I was thinking the AMD parts would walk away with this test.

 

[ano]

interesting, also interesting selection of SKU's on the amd part, I would have expected 75F3 and 7763, not 7r13 (225w tdp) when testing that generation of cpus

They are also mixing in the 8461 intel, which is 4th gen intel, against a 3rd gen amd, here the obvious choice is epyc 9474F for comparison, or 9454P

for power, the high sku intel 8461 is 300w tdp, the dual are 2x205 for 410w.. and the 7R13 at 225W.. so at watt per performance, some of the intel comes in quite bad for performance really



I really wish they would do a proper comparison for stuff like this, Ive always been with intel, but recently become friendly with AMD, for our workloads, and test clusters, and systems deployed to datacenter last 16 months, our AMD 2nd and 3rd gen (and now 4th) has been absolutly crushing our intel stuff :| both performance, and power cost.

We are not that big a company, we have genoa machines in labs. having access to 8461 lab, and not genoa lab is interesting. Not saying Intel sponsored this, but..... yeah seems like a showcase to show off some cpu features, more than a comparison.

 

[gigatexal]

true I've been out of the game a bit so I was not aware of the generational differences but seems they didn't control for tdp or benched the same generation against the equivalent from a competitor etc etc

 

[Patriot]

"We are happy to see that Intel improved throughput on their latest platform. We look forward to testing the most recent AMD platform, and we expect its AVX512 and GFNI support to provide a further performance boost."

Yeah they picked up and optimized off an intel specific avx crypto instruction set. Which is why the mid-end Milan fell off in certain test. Though oddly overall smoother performance, less spikey with thread count... Their data is odd.

Galois Field New Instructions (GFNI) Technology Guide

The Galois Field New Instructions (GFNI) introduced in the 3rd Gen Intel® Xeon® Scalable processor were designed to accelerate cryptographic and security applications. With some imagination, one of the instructions – the affine transformation instruction - also can be used to perform a host of...

networkbuilders.intel.com

 

[111alan]

It's not only in the case of accelerators. When we say 4th gen Xeon scalable is "bad", what we mean is that in per core, even IPC standpoint there is a regression compared to the last gen Xeon. It doesn't mean with the same core count it can't absolute smoke EPYC in most situations. Unless the scenario is very L3 centric like some benchmarks like Cinebench(that's why I always say that it's very important for the "Scale" and "Structure" of benchmarks to match real use cases), or it doesn't care about core performance at all and only care about thread count, like in some cloud rental use cases.

Render time(left to right: EPYC 7B13@3.2-3.3GHz/64c, Xeon 8480+(E5)@3.1GHz/56c, Xeon 8375C@3.5GHz/32c)

 

[RolloZ170]

even IPC standpoint there is a regression compared to the last gen Xeon

intel says 15% more, have you any proofs/links/documents of that ?

 

[111alan]

intel says 15% more, have you any proofs/links/documents of that ?

By my own tests. Ran several tests on a dual 8490H system, one of my friend posted a video about it on a Chinese site.

视频去哪了呢？_哔哩哔哩_bilibili

undefined, 视频播放量 undefined、弹幕量 undefined、点赞数 undefined、投硬币枚数 undefined、收藏人数 undefined、转发人数 undefined, 视频作者 undefined, 作者简介 undefined，相关视频：

www.bilibili.com

15% is probably about core architecture itself, or simply calculated from some selected favorable situations. There is also some MCC single die units that I haven't got my hands on yet. These may have some positive perf gains core on core, GHz on GHz.