Some Early AMD Ryzen Threadripper 1950X Linux Benchmarks

[eva2000]

looking forward to the tests

would be nice to see besides an Apache web server test, a test for Nginx web server too and watch how far Nginx scales with number of cpu cores/threads

[satai]

Great!

I am looking forward compilation tests. Threadriper looks like a reasonable CPU for a developer workstation, mainly because of the ECC (sadly unregistered only AFAIK) support.

[duby229]

Great!

I am looking forward compilation tests. Threadriper looks like a reasonable CPU for a developer workstation, mainly because of the ECC (sadly unregistered only AFAIK) support.

Just my opinion here, I agree with you about ECC, I think all RAM modules should be ECC. People here disagree with me, but I don't care, I know for an absolute fact that all non-ECC RAM modules produce errors every single day. Every single time. So yeah, ECC should've been made standard as soon as memory modules starting hitting over 1GB.

But registered DIMMS are a complete waste. The additional IC in the PCB is usually -the- bottleneck. In basically every single case that IC is way too slow for what the DIMM is rated at. They suck ass. They definitely aren't worth it at all.

[satai]

But registered DIMMS are a complete waste. The additional IC in the PCB is usually -the- bottleneck. In basically every single case that IC is way too slow for what the DIMM is rated at. They suck ass. They definitely aren't worth it at all.

I agree with you partially, mainly for the developer workstation user case. The issue I solve is that I sometimes run a simulation, that needs huge amount of memory. If I run 10+ of them in parallel (That is no issue CPU-wise, because Threadripper has 12-18 cores now), it needs quite a lot memory in total. Unregistered ECC DIMMs are limited now for up to 16GB. That limits the total installable memory to 128GB for all the available motherboards.

[rene]

I doubt they are blanks. In the chip world they would be foolish to be cutting up good silicon to fill the spaces with. I would be they take dead die that failed at probe and pull them and just them as spacers. So while they wouldn't be blank die, they would be about as useful.

You want to get as many useful die on a wafer as possible and when you have blanks in the mix it would mess up etch uniformity and mess some with the CMP.

I know just a small detail but I do really get a kick out of the stories on windows sites and the people who comment that talk about making chips, you can tell most have no idea what they are doing and the ones that semi know most of them are in college or professors that have not been in the real world and don't realize how far off their ideas are. (Yes I work for a company that makes chips at small geometries and high speed) Though not all are that way, there are some you can tell that know their stuff.

also in the future they could replace the two dead dies, with working ones and again double the thread/core count !!! ;-)!

[torsionbar28]

I doubt they are blanks. In the chip world they would be foolish to be cutting up good silicon to fill the spaces with. I would be they take dead die that failed at probe and pull them and just them as spacers. So while they wouldn't be blank die, they would be about as useful.

AMD refers to them as "spacers" for the purpose of supporting the heat spreader. Agree they're probably not made of blank silicon wafers, but not sure it really matters what they're made of. Point is they're not simply disabled cores or some such, it sounds like they are purposefully inert pieces.

[asdfblah]

Michael, do you plan to look into the KVM NPT that affects all AMD CPUs?

Also, are you going to look into the PCI bug that affects GPUs on threadripper.

I wonder... Is this really a CPU bug? Doesn't it only affect KVM, while Xen, VirtualBox, ... work just fine? If so, how come the KVM people haven't taken the issue seriously?

[edwaleni]

Gamer benchmarks on server chips have just as much value as NAS benchmarks on desktop chips.

You don't focus on the method, you focus on what drives the result. If Crysis makes a large use of some function that requires the CPU supply "X", then the bench has merit.

Likewise when running specialized function tests (ie: NAS) on a desktop CPU, it can expose how well a particular function behaves.

Also, one can validate across benchmarks on certain behaviors the CPU provides. Why does one bench show well on this CPU and bad on another when they seem the same?

Gamer benchmarks have exposed more issues in datacenter oriented hardware than I ever imagined.

[Marc Driftmeyer]

Why in the hell are you benchmarking against dual Intel Xeon Gold 6138 Tyan Server?

40 Cores/80 Threads vs 16/32 and the best part, when EPYC arrives for Server v. Server you better test against the Tyan or people will find your `objectivity' skewed.

The EPYC will win out and anyone still buying Intel hardware will look like fools.

[drSeehas]

... because of the ECC (sadly unregistered only AFAIK) support.

Are there any Socket TR4 mainboards with working ECC? I can see only gaming mainboards.