AMD Threadripper 2990WX Linux Benchmarks: The 32-Core / 64-Thread Beast

[juanrga]

Which have to be proven.

Since latency depends on distance, two adjacent dies connected directly will send/receive signals fasters connected through interposer.

[nuetzel]

Since latency depends on distance, two adjacent dies connected directly will send/receive signals fasters connected through interposer.

They are in _both_ cases NOT directly connected.

[chithanh]

The Computex demo used a hacked Xeon on a modified server board. Since both Xeons and server boards aren't designed for overclocking (*), Intel had to use a chiller to push clocks to 5GHz and simulate the performance of the forthcoming 28 core Skylake-A chip.

(*) Same reason why der8auer used a chiller to overclock a EPYC on a SP3 board to simulate the performance of a 32 core ThreadRipper.

I replied to that elsewhere already. The Intel Computex demo had nothing to do with reality - neither with power consumption nor with performance - of the product that you would be able to buy later this year. No reasonable configuration can reach 5 GHz on all 28 cores for sure.

I said that some Zen systems fall to satisfy the marketing TDP, not that all Zen do.

You specifically replied to a post of that discussed 7980XE vs. 2990WX power consumption, and claimed that Intel is within TDP under "ordinary load" while many Zen systems aren't.
The Phoronix numbers suggest that 2990WX is within TDP under load. Also Tom's Hardware agrees:

AMD stays exactly within their TDP of 180 W and 250 W respectively, under all conditions and all loads. Only if you enable PBO, the 2990WX will consume 500 W or more if the cooling and mobo are up to it.

Intel on the other hand succumbs to the AVX powervirus (as you call it) and exceeds the rated 165 W TDP by a wide margin at stock configurations.

[juanrga]

I replied to that elsewhere already. The Intel Computex demo had nothing to do with reality - neither with power consumption nor with performance - of the product that you would be able to buy later this year. No reasonable configuration can reach 5 GHz on all 28 cores for sure.

You specifically replied to a post of that discussed 7980XE vs. 2990WX power consumption, and claimed that Intel is within TDP under "ordinary load" while many Zen systems aren't.
The Phoronix numbers suggest that 2990WX is within TDP under load. Also Tom's Hardware agrees:

AMD stays exactly within their TDP of 180 W and 250 W respectively, under all conditions and all loads. Only if you enable PBO, the 2990WX will consume 500 W or more if the cooling and mobo are up to it.

Intel on the other hand succumbs to the AVX powervirus (as you call it) and exceeds the rated 165 W TDP by a wide margin at stock configurations.

First, I will quote myself, because you are ignoring relevant parts of what I said:

I said that some Zen systems fall to satisfy the marketing TDP, not that all Zen do. E.g. '15W' ryzen mobile are usually 25W or even 35W. The 2700X has a real TDP of 140W. The older '65W' 1700 was a 90W chip,... other Zen chips work within the official TDP.

Early engineering samples of first gen threadripper violated the rated TDP. E.g. one of the first 180W samples had a real TDP above 200W. Final chips worked within the official TDP of 180W, but they include a power limit mechanism in the TR4 socket that underclocks the core under the base clock under full loads. Here you have a 1950X in action

https://www.hardware.fr/medias/photo...IMG0054462.png

I suppose second gen threadripper work similarly because using the same socket.

Second, TDP is defined for ordinary workloads, not for AVX-based power virus, whose goal is to stress CPUs beyond stock limits. Power virus are often used to check stability of overclocked chips. If an overclocked chip is stable under a power virus, then it is stable for 24/7 operation.

Third, Intel chips have 256bit or 512bit AVX units. The 7980XE has a pair of 512bit AVX units per core, whereas Zen core has only two 128bits FMAC units. I.e. Skylake-X core has 4x bigger units than Zen core and so consumes much more under AVX-based stress tests.

Fourth, Prime 95 isn't a power virus for Zen systems as the Stilt proved time ago:

Note: Current versions of Prime95 (28.10) do not stress Ryzen CPUs properly. The resulting power consumption is abnormally low, and both Firestarter and Linpack result in significantly higher power consumption.

That is why the 2700X consumes only 104.7W under Prime95 but consumes 140W under ordinary workloads such as encoding/rendering.

[DavidKL]

(Note: This post was also posted in the Linux Scaling topic)

You have been quoted by PCMag. The reviewer, though, apparently on a quest to support his headline (which makes no mention of Windows flaws), uses a Microsoft product (Visual Studio) to see whether or not another Microsoft product (Windows) is hampering the 2990. This reviewer also doesn't bother to do any Linux tests himself nor does he, after pointing out that your 7 Zip version is older code, bother to check to see if the more recent 7 Zip code could be an issue.

https://www.pcworld.com/article/3298859/components-processors/how-memory-bandwidth-is-killing-amds-32-core-threadripper-performance.html?page=2

[chithanh]

First, I will quote myself, because you are ignoring relevant parts of what I said:

Second, TDP is defined for ordinary workloads, not for AVX-based power virus, whose goal is to stress CPUs beyond stock limits. Power virus are often used to check stability of overclocked chips. If an overclocked chip is stable under a power virus, then it is stable for 24/7 operation.

Third, Intel chips have 256bit or 512bit AVX units. The 7980XE has a pair of 512bit AVX units per core, whereas Zen core has only two 128bits FMAC units. I.e. Skylake-X core has 4x bigger units than Zen core and so consumes much more under AVX-based stress tests.

That is all fine and dandy, but actually you were ignoring the relevant aspects of what the original post which you replied. Here it is again for reference:

On topic:

Amazing how the Intel Core i9 7980XE, which is supposedly a 165W TDP consumes almost the same energy of the TR 2990WX, which is 250W TDP, even taking the whole system in account

The difference in system power consumption between 2990X and 7980XE (381 W vs 315 W, Δ=66 W) is smaller than the difference in CPU TDP (250 W vs. 165 W, Δ=85 W) And Phoronix was using Stockfish benchmark here, not Prime95. The measurement is therefore incompatible with the assumption that Intel stays within TDP and AMD doesn't. You bringing up how Zen elsewhere exceeds TDP and Intel doesn't, serves therefore no other purpose than throwing shade on the 2990WX.

Fourth, Prime 95 isn't a power virus for Zen systems as the Stilt proved time ago:

Note: Current versions of Prime95 (28.10) do not stress Ryzen CPUs properly. The resulting power consumption is abnormally low, and both Firestarter and Linpack result in significantly higher power consumption.

That is why the 2700X consumes only 104.7W under Prime95 but consumes 140W under ordinary workloads such as encoding/rendering.

(emphasis added)

That is bollocks. The Ryzen Prime95 power consumption is not abnormally low. It is precisely (within 1 W) matching the TDP as Tom's Hardware measured (105 W for 2700X, 180 W for 1920X/1950X/2950X, 250 W for 2990X).

This reviewer also doesn't bother to do any Linux tests himself nor does he, after pointing out that your 7 Zip version is older code, bother to check to see if the more recent 7 Zip code could be an issue.

Ugh, it always surprises me how tech reviewers can expose such astonishing levels of incompetence. I mean if you can't make informed statements about or perform tests on Linux fine, don't do that then. But what the reviewer said is just asinine.

[andrei_me]

You bringing up how Zen elsewhere exceeds TDP and Intel doesn't, serves therefore no other purpose than throwing shade on the 2990WX.

I said exactly the opposite, I was putting emphasis on how Intel consumes more energy than advertised, you even confirms my claim showing how the total energy delta is smaller than the CPU's TDP delta.

Read my post again...

[chithanh]

andrei_me I was using your post (which indeed noted better power consumption in 2990WX system vs. TDP compared to 7980XE system vs. TDP) to show how juanrga was throwing shade on the 2990WX, sorry for causing misunderstanding.

[andrei_me]

Aaah, now it makes sense, sorry for being rude earlier, we are all fans of TR

[drdwo]

Did some experiments with numactl with scientific code I wrote myself. Found out that I got most out of the 2990WX running 4 experiments in parallel each using 16 parallel threads. The first two have to be called with numactl -l -N0 and numactl -l -N2, the other two using numactl -l. The first two run as fast as on my 16 Core 1950x, the other two suffer from a 10-15% slowdown because of the missing direct memory connection. Without using numactl all four parallel runs suffer from a similar slowdown. In this setting utilizing all cores TR2 seems not faster than TR1 - may be the 2950x - which I don't have - can slightly be higher clocked than the 2990wx or the 1950x. On the positive side: I could solve a scientific problem which required a 100 CPU cluster before with 1 CPU in about 14.5 hours. Partly because the 2990WX is quite powerful, partly because of the superiority of my algorithm. See http://scientificcomputation.blogspo...90-wx-for.html