Forum:Choose between AMD and INTEL Server CPU in 2021-2022
2
1
Entering edit mode
3.4 years ago
JustinZhang ▴ 120

Ladies and Gentlemen, I'm now planning to buy a new high-performance computing server for bioinfomatics. I prefer the tower type because the server is supposed to be laid at the corner of our office as we don't have a single computer room.

There are several options for me as follows:

1) Dell T640, with two Intel Xeon 8280 CPU (each 28 cores, 56 thrds, 2.70~4.00GHz)(best tower type of server in market)

2) H3C UniServer R4950 G5, with two AMD EPYC Milan 7763 CPU (each 64 cores, 128 thrds, 2.45~3.50GHz)(though rack type, but a silent cabinet only costs less than 1000 CNY = 155 USD at 2021/07/03)

3) Wait for a tower server to come to market with 3rd Gen Intel Xeon CPUs (especially Intel® Xeon® Platinum 8380) or AMD EPYC Milan CPUs (especially AMD EPYC™ 7763)

I wonder if there are any compatibility issues between AMD cpu and common-used softwares? The AMD seems much cheaper and could leave me more budget for RAM, but compatibility is more essential, right?

Also, have anyone tested computing performance of these CPUs(INTEL 8280 vs. AMD 7742; INTEL 8380 vs. AMD 7763)? Some people said the calculating capability of 2 AMD threads equals that of a single Intel thread.

Besides, is the rack type of server really very noisy?

Thank you in advance for sharing your opinion and your experience.

hardware server • 4.0k views
ADD COMMENT
0
Entering edit mode

Relevant reviews from Hexus: Milan 7763, Ice Lake 8380. My impression is that the AMD offering is better.

I wonder if there are any compatibility issues between AMD cpu and common-used softwares?

Why would this be the case?

Some people said the calculating capability of 2 AMD threads equals that of a single Intel thread.

Somebody should send this to Intel's marketing department?

Besides, is the rack type of server really very noisy?

You don't want to have a rack server in your office.

ADD REPLY
1
Entering edit mode

Thx for your reply.

Why would this be the case?

As most discussions about this issue on the Interent are contaminated by AMD fans, Intel fans and hardware merchants, i cannot get enough valuable info. And that's why I raise this question. There are only two useful points I got as below: 1) GATK team said they have worked with Intel team to optimize code. 2) BUG reported on Github page.

You don't want to have a rack server in your office.

OK. I don't want to apply hydrocooling system on reck server either, so it seems that tower type is the only feasible option.

ADD REPLY
1
Entering edit mode

Thank you for sharing those links. Didn't know about either of these things. I wonder why there isn't a GATK-AMD partnership--perhaps it's in the works.

That bug report is pretty concerning also.

It remains in the model(images) forever with gpu utilization go to 0% (memory is occupied, not 0), three cpu cores go to 100%, and other cpu cores go to 0%.

This comment from the bug report stands out to me. I've had something similar happen to me on a couple of AMD Rome machines (2x 64 cores, 128/256 threads). I'd submit a job to the machine (via Slurm), come back and take a look, and there'd be no job-related CPU activity, but some 50-60 GB of RAM would have been occupied. But I presume the issues I had faced probably stem from Slurm mis-configurations, network file storage problems, and maybe Linux kernel issues than the processors themselves. (These machines were recent acquisitions.) The problem never occurred on the much older Intel Haswell hardware that co-exists with these machines.

OK. I don't want to apply hydrocooling system on reck server either, so it seems that tower type is the only feasible option.

If systems integrators don't offer tower-builds, you could try throwing a machine together yourself. You'll be on your own when it comes to troubleshooting over the lifetime of the device though.

ADD REPLY
0
Entering edit mode

Depending on the bandwidth available to get the data to and from your server, a remote/cloud server could be a competitive option.
Also if you have the resources, you could test a few different hardware set ups with some cloud providers before buying your server.

ADD REPLY
0
Entering edit mode

Since you mention cloud compute: do you actually know of any good cost-efficient providers in Europe that don't happen to be Google, Amazon, or one of the other big companies?

ADD REPLY
1
Entering edit mode

I can't speak about cost because this really depends on the specifics of the tasks and the provider. However, besides the commercial entities, there are also a number of academic clouds such as the EMBL-EBI Embassy cloud or the German de.NBI cloud. You can find moe providers by browsing the European Open Science Cloud marketplace.

ADD REPLY
4
Entering edit mode
3.4 years ago
  • go for AMD. Much better processors than Intel in the last two years
  • it's better to max RAM than go for a slightly better CPU
  • get as much SSD as you can afford, and then some. Bioinfo is more about IO than CPU in most cases.
  • don't put it in the office. Heat, noise, try to find a broom cupboard with cooling somewhere ....
ADD COMMENT
1
Entering edit mode
3.4 years ago
GenoMax 147k

We have been happy with AMD CPU's of late. They are going to give you the best performance for your yuan.

Large server CPU's generate plenty of heat and thus have a need for noisy fans. You can invest in proper liquid cooling solution if noise is going to be a consideration. You can also go with CPU's that are a few rungs below the top tier and/or designed for low wattage. You are not going to notice the additional time it is going to take to complete your jobs over a period of time.

A server is sum of all of its parts. While the CPU is important RAM/storage is more critical for scientific workloads. Pay adequate attention to those areas otherwise your cores may keep waiting for data to come to them.

Some people said the calculating capability of 2 AMD threads equals that of a single Intel thread.

That is not correct.

ADD COMMENT
0
Entering edit mode

Lesson learned, and thx very much for your reply.

ADD REPLY

Login before adding your answer.

Traffic: 2690 users visited in the last hour
Help About
FAQ
Access RSS
API
Stats

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.

Powered by the version 2.3.6