Some Early AMD Ryzen Threadripper 1950X Linux Benchmarks

[duby229]

Some delidding info to consider...

https://www.extremetech.com/computing/253248-amd-threadripper-delidded-multi-core-surprise-hood?source=Computing

According to der8auer, who spoke with AMD, Threadripper is an Epyc CPU configuration with four dies, each packing eight cores. At first, we thought this meant AMD had gone for a similar arrangement to what it used for Ryzen 3: four cores active on each die (with SMT enabled in this case) and 16 cores/32 threads on the entire chip. But that’s not the case if der8auer’s information is accurate. According to him, Threadripper only uses two of the four Ryzen dies mounted on the PCB.

I would love to hear some confirmation about this! What's up with the 2 unused dies?

[Adarion]

With that beast one could probably even play games on software rendering.

I'm also very interested in the usual multithreaded benchmarks (as well as some single-threaded things) and in power numbers: especially idle and good load on all cores (e.g. compiler running).

why the stupid environmentally polluting packaging? For what all the plastic foam crap? Can they not put it in a nice paper box? :-/
a AMD fan since my first am386, but this plastic waste stupidity should be boycotted to save our only planet, …!

Yes on plastic packaging in general, though one must always include all factors into an equation. (is it technically necessary to keep the item intact, will it be recycled or upcycled, is it from recycled materials, what overall water and power consumption would a cardbox / paper equivalent have and so on)
Apart from that a part of the plastic goes actually into the CPU socket and remains there (kind of an insertion aid as far as I saw).
Moreover - not everyone needs a Threadripper. Just for writing and surfing some <= 25 W chip power Kabini system might be sufficient.
Another point: You might want to engage in other places first where large amounts of plastic are wasted daily, countries that have no recycling, or you might want to engage in something like the Ocean Cleanup Project.

[chuckula]

Threadripper handily beats the intel i9 in every test, sometimes by a little, and sometimes by a lot. Impressive!!

You obviously didn't read the review then where it loses in at least one of the NAS parallel computationally intensive tests [including losing by practically 50% in NAS CG.B to a chip with 6 fewer cores and a lower clock speed, I wouldn't call that 'impressive' in the slightest] and is a statistical tie with a chip that has 6 fewer cores and a lower base clockspeed in others.

Incidentally, it didn't seem to win by the expected 60% (or more) margin that one would expect for a truly 16 core vs. 10 core chip in any benchmark whatsoever.

Now that we've seen how RyZen scales (or doesn't scale) to 16 cores I'm extremely interested to see how the 16 core Skylake X chips do. I'm expecting substantially better scaling compared to a 10 core Skylake vs. what we've seen from Threadripper.

[Brisse]

Actually, no. In this case, it's the retail kit for Threadripper - https://www.newegg.com/Product/Produ...82E16819113447

Oh!

[Filiprino]

Most benchmarks do not scale beyond 10 cores. A lot of other benchmarks are optimized for Intel and Intel's AVX/AVX2 (with intrinsics). No -mtune nor -march where used.

You obviously didn't read the review then where it loses in at least one of the NAS parallel computationally intensive tests [including losing by practically 50% in NAS CG.B to a chip with 6 fewer cores and a lower clock speed, I wouldn't call that 'impressive' in the slightest] and is a statistical tie with a chip that has 6 fewer cores and a lower base clockspeed in others.

Incidentally, it didn't seem to win by the expected 60% (or more) margin that one would expect for a truly 16 core vs. 10 core chip in any benchmark whatsoever.

Now that we've seen how RyZen scales (or doesn't scale) to 16 cores I'm extremely interested to see how the 16 core Skylake X chips do. I'm expecting substantially better scaling compared to a 10 core Skylake vs. what we've seen from Threadripper.

[darkbasic]

It's the benchmark which doesn't scale well, not the cpu.

[jrch2k8]

You obviously didn't read the review then where it loses in at least one of the NAS parallel computationally intensive tests [including losing by practically 50% in NAS CG.B to a chip with 6 fewer cores and a lower clock speed, I wouldn't call that 'impressive' in the slightest] and is a statistical tie with a chip that has 6 fewer cores and a lower base clockspeed in others.

Incidentally, it didn't seem to win by the expected 60% (or more) margin that one would expect for a truly 16 core vs. 10 core chip in any benchmark whatsoever.

Now that we've seen how RyZen scales (or doesn't scale) to 16 cores I'm extremely interested to see how the 16 core Skylake X chips do. I'm expecting substantially better scaling compared to a 10 core Skylake vs. what we've seen from Threadripper.

For anyone relatively aware of how storage works that result is perfectly normal and as expected as the eventual mp3 encoding.

What I mean is NAS is certainly parallel but is not intensive, is more concurrent than parallel. aka it tries to keep I/O operations non-blocking or at least distribute them on as many lock independent threads as possible(is very dependent of how the I/O subsystem and bandwidth are available) but the actual code running on each thread is very frequency sensible and the i9-7900x has higher turbo frequencies(and we don't know yet how the turbo in TR works on linux or which was was used) and a bit better latency while TR and Ryzen for that matter is very affected here for the low load on each thread vs latency.

Now in the NAS scenario is an overall Win for TR in more complex scenarios, like.

1.) ZFS is very compute intensive under high I/O pressure and AES-NI(as shown in many other benchmarks) is way faster on Zen, so for complex RAID60+encryption+dedup+High Bandwidth caches/log NAS system, Ryzen/TR and Epyc should show very strong performance but for smaller less enterprise class NAS systems Intel have the upper hand due to higher frequencies and lower latency.(this one should be very tricky to benchmarks tho)

2.) X399 and sR4 have way more available bandwidth, as a platform intel M.2 and U.2 are raid(ish) capable with ZFS or with the bundle usb thingy but the bandwidth is capped by a single chipset to CPU x4 lane while X399 and sR4 each M.2 and U.2 busses go straight full x4 to the CPU. Hence if you test ZFS on raid1 on each platform TR/sR4 will show otherworldly read/write result while X299 will be a bit better than a single M.2.

As an additional note here, even with that "win" on intel 7900x it should never be recommended/used for any decently able NAS system simply because X299 don't support ECC memory making this test entirely educational but invalid for any real scenario. If anyone need a NAS and prefer an Intel platform go for an equivalent Bronze/gold Xeon instead unless your NAS is strictly for home use

[CuriousTommy]

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.

[hauberg]

IT would be great to see benchmarks comparing different implementations of blas on ThreadRipper. That's what matters for scientific computing.

[torsionbar28]

As an additional note here, even with that "win" on intel 7900x it should never be recommended/used for any decently able NAS system simply because X299 don't support ECC memory making this test entirely educational but invalid for any real scenario. If anyone need a NAS and prefer an Intel platform go for an equivalent Bronze/gold Xeon instead unless your NAS is strictly for home use

^ This, exactly. Server benchmarks on a desktop chip are pointless. It's a fake use case. Nobody is using desktop peecee chips in a server. It's no different than running gaming benchmarks on a server chip - nobody cares how Crysis 3 runs on Xeon because nobody is playing Crysis 3 on a Xeon.