hardware internet

System Ryzen Threadripper 3970X – check efficiency and intake. Comparison of three generations Ryzen

Accrued passion

This subject matter written via the customer, and the amassed remuneration.

Frankly, for relatively a very long time consciously and unconsciously not on time the discharge of this a part of the fabric. And just like the exams nonetheless to the NG carried out, and the puzzle is “how one can write” in my head have been, however as they are saying – “the fabric isn’t”. Long may just now not perceive why, however the explanation why, as same old, lies at the floor. The indisputable fact that the fabric isn’t vintage, isn’t acquainted to the reader, and as a end result marginally helpful exams in some Linux, in any no transparent goals. So, I can you have to be as informative as conceivable and provide an explanation for what, how and for what we can do.


Of route one thing acquainted, too. For instance, we examine the efficiency of three generations Ryzen, measure how a lot the device consumes in Threadripper 3970X from the socket and check the sport. In General, I drove.

Will start with a short lived description of the check programs. Since I don’t faux to be educational and now not working exams “on science”, then describe intimately the entire device does now not see the purpose. The comparability can be most commonly estimates. Processors of all programs can be within the drain — ie, no acceleration won’t use, even PBO. With the exception of reminiscence.


In instances the place a substantial impact of SSD drives, exams will do in order that this impact is to exclude or decrease.

System Threadripper 3970X(detailed description is in my profile):


Memory: 64GB Corsair Vengeance LED, 3000МГц running on 2733МГц

System SSD: Samsung EVO 960


Motherboard: Gigabyte AORUS TRX40 XTREME

Video card: ASUS GeForce RTX 2080Ti ROG STRIX OC GAMING



System Ryzen 1800x:

Memory 16GB GOODRAM IRDM X — 2666МГц running on 2666МГц

System SSD: Corsair Neutron GTX 240GB

Motherboard: ASRock Pro4 B450

Video card: now not put in.


System Ryzen 2700x:

Memory: a hodgepodge of two HyperX Fury 4Gb and two inexperienced modules, Samsung 4Gb. All the modules function at 2400MHz

System SSD: 120GB AMD Radeon R3

Motherboard: Gigabyte AX370-Gaming K7

Video card: ASUS GeForce RTX 2080Ti ROG STRIX OC GAMING

Memory frequency displays the utmost strong running months, which was once accomplished. Not all exams can be in comparison to all three programs.

In my packages steadily must procedure massive units of textual content knowledge, specifically to variety and take away duplicates traces from a dossier. But with this process, and start our checking out.

To do away with the affect of CDs, all exams can be carried out in reminiscence. Temporary information can also be saved in reminiscence /dev/shm/. So take a small dossier with the intention to examine other programs. The dossier isn’t marked-up textual content corresponding to the next:

“divine savior Aaron’s kiss collection e-book two via kathi s barton international citadel publishing httpwww

worldcastlepublishingcom it is a paintings of fiction

names characters puts and incidents are merchandise of the writer’s creativeness or are used fictitiously and aren’t to be construed as actual

any resemblance to precise occasions, places or organizations individual residing or useless is fully coincidental”

File: x01part1-text-dataset-for-test.txt

The collection of traces in dossier depend command “wc” with the-l transfer:

wc-l x01part1-text-dataset-for-test.txt

38735813 x01part1-text-dataset-for-test.txt


du -hs x01part1-text-dataset-for-test.txt

2.3 G x01part1-text-dataset-for-test.txt

For the 16 move processors, the check was once performed at 4, 8, 16 threads.

For Threadripper 4, 8, 16, 32, 64 threads.

In General the command seems like this:

time cat x01part1-text-dataset-for-test.txt | LC_ALL=C sort-u –parallel=16 -S 3G -T /dev/shm > /dev/null

That’s what makes this difficult line:

The command “time” – measures the execution time of the entire construction;

“cat x01part1-text-dataset-for-test.txt” – merges the dossier line via line to straightforward output (display screen);

“|” – pipe — the pipeline, “transfers” all of the knowledge to the type command;

“LC_ALL=C” – quickly units the locale simplified, as variety is way quicker sorting slightly worse;

Parameters variety:

key “u” – implies that you wish to have to take away duplicates;

“–parallel=” – units what the collection of threads to variety;

“-S 3G” units the buffer reminiscence 3Gb;

The “-T” – specifies the listing for brief information. In our case it’s the RAM.

“>” – specify the place to direct the output of the command variety. In our case, to not use neither reminiscence nor disks all output to /dev/null.

For instance, Ryzen 1800x will obtain roughly such output:

time cat x01part1-text-dataset-for-test.txt | LC_ALL=C sort-u –parallel=16 -S 3G -T /dev/shm > /dev/null

actual 0m20.801s

consumer 1m10.018s

sys 0m3.789s

We have an interest within the “actual” – conditionally how a lot time took to execute all instructions within the string. The time in seconds.

Here are the effects I were given:


Please notice that the chart Y-axis begins now not from 0 however from the 15 seconds. So the time distinction is healthier visual.

It is noticed that the sorting time is lowering relatively all of a sudden with expanding numbers of bodily cores. But this “WOW” impact, as can also be noticed from “column” to Threadripper, disappearing after 16 cores. Ie 32 nucleus has now not appreciably scale back the sorting time. What is the rationale we will be able to handiest bet. It is conceivable that the software of variety is solely now not optimized for extra cores might prohibit one thing else, for instance the velocity of the reminiscence.

Also, we will be able to see how AMD has advanced each and every era of processors. Ryzen 2700x 1800x quicker via about 5%. Threadripper 3970x 2700x quicker via about 4%, with the similar quantity cores in fact. Improvement, despite the fact that now not vital, however they’re. It’s time to take into accounts the notorious 5% in step with era. 🙂


For exams use the multi-threaded pigz archiver – complete analogue gzip, handiest mnogopotochnoy. In the arena of unix/linux gzip archives are nonetheless commonplace. All types of RAR, ZIP, 7Z, cross to the furnace. They have a large drawback with mnogopotochnoy in Linux. 🙂

The dossier will use the similar as within the earlier exams — x01part1-text-dataset-for-test.txt. All the exams do additionally in reminiscence, as a result of check power isn’t integrated in our duties.

To check, use a easy script:

for i in `seq 1 16`; do time pigz -p $i -k x01part1-text-dataset-for-test.txt && rm x01part1-text-dataset-for-test.txt.gz; carried out

What makes this set of instructions:

The frame of the loop “for i in `seq 1 16`; do” runs from 1 to the utmost collection of threads.

Inside the loop:

exhibiting at the terminal the collection of streams which is compressed: echo “Number of threads: $i”;

time – the measured time;

compress a dossier with a selected collection of threads “$i” – “pigz -p $i -k x01part1-text-dataset-for-test.txt”. Key-k manner stay the unique dossier.

“&&” – situation. If the command was once finished as it should be, carry out the next.

“rm x01part1-text-dataset-for-test.txt.gz” delete the archive.

“carried out” – end the collection.

For Threadripper’and the collection of threads will do 64, let it sweat.

The output of the command can be like this:

Number of threads: 1

actual 1m54.367s

consumer 1m53.911s

sys 0m0.425s

Number of threads: 2

actual 0m58.683s

consumer 1m59.683s

sys 0m1.472s

Number of threads: 3

actual 0m39.615s

consumer 2m0.487s

sys 0m1.242s

Then simply accumulate all of the knowledge within the dossier and construct the graphics.


Looking on the charts I used to be come what may now not inspired. 2700x and Threadripper despite the fact that forward of Ryzen 1800x, however equivalent to one another, as much as 8 cores. Further Threadripper takes and simply killing all of the brute pressure. The enlargement of the nuclei is for sure spectacular. But so we write the kernel to this actual drawback is 2700x. Will attempt to dig some additional info from the information accumulated.

We can calculate the dimensions issue of the archiver to the collection of bodily cores.

Ryzen 1800x:

1 core — 147,15 sec, 8 cores — 20 h, the dimensions issue is got of 7.36 — nearly linearly.

Ryzen 2700x:

1 core — 114,36 sec, 8 cores — of 15.55 seconds, the dimensions issue is got of 7.35 — rather worse, however nearly linearly.

Threadripper 3970x:

1 core — 122,34 sec, 32 cores — 3,98 sec, zoom ratio of 30.74 seems — as it’s not fully linear. Count to eight cores: 1 core — 122,34 sec, 8 cores — 15,29 h, the scaling issue is got 8,001. Wow! Even a bit higher than simply linearly. It seems that if cores so much, Cycling a smaller collection of them, you’ll be able to get 100% linear scaling.


Division of the textual content into engrams.

I within the remaining section promised to not send exams of their explicit duties, however I feel one drawback you’ll be able to check.

Has to paintings so much with texts. Frequent drawback – number of textual content n-grams. Conditional n-grams are combos of phrases. There are other two, three, five, ten phrases. I do such duties in Python with the Toolkit NLTK – Natural Language Toolkit.

Since the duty isn’t easy and in most cases takes an afternoon, then we rather simplify your lifestyles, a mouthful from our dossier “x01part1-text-dataset-for-test.txt” the primary 8 000 000 traces


head-n 8000000 x01part1-text-dataset-for-test.txt > 8000000-text-dataset-for-test.txt

File dimension:

du -hs 8000000-text-dataset-for-test.txt

483M 8000000-text-dataset-for-test.txt

What do you assume the volume of textual content in MB or KB, will after laying on engrams? Write your personal model within the feedback.

Next, let’s arrange our dossier with ideas at the collection of items equivalent to the collection of bodily CPU cores.

For 2700X and 1800X – it is 8 items, for Threadripper 3970X – 32 piece. Do this with the command break up:

For 8 cores:

break up –additional-suffix=text-dataset-for-test.txt -l 1000000 -a 4 –numeric-suffixes=0001 8000000-text-dataset-for-test.txt

For 32 cores:

break up –additional-suffix=text-dataset-for-test.txt -250000 l-a 4 –numeric-suffixes=0001 8000000-text-dataset-for-test.txt

If you break up the textual content information the use of the break up there’s one essential level — the textual content dossier it will be significant to separate line via line, the use of the-l transfer. If you smash a byte-by-byte the use of the-n transfer, you are going to obtain the items of partial textual content in information.

Next, get started precisely at the collection of bodily cores replica of this straightforward script

#-*- coding: utf-8 -*-
import time
start_time = time.time()
import random
import string
import os
import sys
import nltk
from nltk.util import ngrams
from nltk.tokenize import sent_tokenize, word_tokenize filename = ".sign up for(random.selection(string.digits) for i in vary(8))
dossier = open(("time-test-" + filename + ".txt"), 'w')
tmpfile = open(("tmp-file-" + filename + ".txt"), 'w') def extract_ngrams(knowledge, num): n_grams = ngrams(nltk.word_tokenize(knowledge), num) go back [ ' '.join(g) for grams in n_grams] for stringin in sys.stdin: knowledge = stringin.strip().decode('utf-8', mistakes='exchange').encode('ascii', mistakes='exchange') words = sent_tokenize(knowledge) for g in vary(0,len(extract_ngrams(knowledge, 4))): twogram = extract_ngrams(knowledge, 2)[grams] threegram = extract_ngrams(knowledge, 3)[grams] tmpfile.write(twogram + "n") tmpfile.write(threegram + "n") dossier.write("%f seconds" % (time.time() - start_time) + "n")
dossier.shut() tmpfile.shut()

In a nutshell, this script will get the enter sentence, decomposes it into bigrams and trigrams, writes them to a dossier paulrandal identify, additionally in a separate dossier is written to the script execution time. The precise script execution time is solely what we want.

In a General view can be about this design:

cat <piece of the dossier> | ./ – will run concurrently, one replica in step with bodily core. This can also be carried out in numerous tactics.

For instance, making a bash script of the shape:

cat <piece 1> | ./ &

cat <bite …> | ./ &

cat <bite N> | ./ &

I ran during the display screen. There design is extra difficult, nevertheless it does the similar factor. I in order that handy – every so often you wish to have to keep watch over the method of scripting. In General, used their achievements.

Since the run time of each and every replica of the script isn’t strictly deterministic, then for higher effects, let’s variety the time in ascending order.

Since I did the measurements for 3 processors: 1800x Ryzen, Ryzen 2700x and Threadripper 3970x, then there’s one small drawback. Due to the truth that the dimensions of the items are other, the result of 8-core processors does now not correspond with the 32-core. In order with the intention to correlate the efficiency of all processors, for starters we can check it on 8 cores and for Threadripper 3970x, thus we will be able to relate the efficiency of various generations of processors.

Here’s a ravishing chart grew to become out:

For the duty “lay out the textual content on engrams” acceleration for three generations greater than vital.

Ryzen 2700x on moderate 9% quicker than Ryzen 1800x. Threadripper 3970x on moderate quicker than 2700x via an excellent 24%.

But all of the effects seem at the total chart, the place Threadripper 3970x displays all its energy:

In different phrases, what used on my Ryzen 2700x regarded as hour, I can carry out the similar process in 13 mins. Acceleration is ready 4.7 instances.

It is not going an ordinary consumer makes use of your house laptop to retrieve n-gram from the textual content. Right? And I’m now not the one engrams I pick out, every so often flexing. So I used to be additionally questioning what’s there within the video games. 🙂

Compare handiest took place Ryzen 2700x and Threadripper 3970x, after which handiest in one check that I used to be now not that upset and somewhat confused.

Linux device is omnivorous, and in case you have the whole thing put in as it should be, you’ll be able to simply insert the outdated power into the brand new device and quiet to run. That’s what I did. For passion I’ve examined Superposition Benchmark v1.1 other variations of Ubuntu.

In Ubuntu 16.04 Ryzen 2700x and Threadripper 3970x made beautiful shut.

The outcome 2700x — 10150 parrots:

The outcome Threadripper 3970x — 10118 parrots:

The distinction of 0.3% is even not up to the mistake.

And this is the end result 18.04 in Ubuntu with new Nvidia drivers I used to be confused — it’s decrease via 1.3%. Although it’s anticipated that the brand new can be higher than the outdated. The result’s repeated continuously — that implies now not some random surge.

The first Lark examined handiest Threadripper. So, what is a excellent effects or now not. Settings and integrated check effects the next display screen pictures.

Temperature and electrical energy intake.

In the off state, the pc consumes from a community of about 8W.

In the on state with a loaded Linux device eats about 170Вт.

You can throw 20-30W if you happen to trade the keep watch over frequency of the CPU from “efficiency” to “ondemand”. By default, Ubuntu makes use of “ondemand”. The device works a bit slower, whilst eating 130-150W. I exploit “efficiency”, so with out load the device eats a bit extra. The state of “efficiency” and “ondemand” – can in first approximation be regarded as as analogs of efficiency plans in Windows.

Further, the intake within the load.

The CPU was once loaded with High95 v29.8,construct 6 64 thread Smallest FFTs.

The most intake, which controlled to {photograph} — 507Вт. Caught specifically.

Usually the intake is strolling within the house of 430 — 450W.

The temperatures I shot in two variations of the case:

Option vintage processor from above however from underneath it heats the graphics card;

Option “Changeling” – processor from underneath, blown via the chilly air. The processor itself warms up slightly graphics card.

For CPU load used High95 v29.8,construct 6 64 thread Smallest FFTs to load the video card GPU_BURN –

All classes for 20 mins, then got rid of the displays. The temperature of inlet into the housing of the air was once within the vary 25-26 gr. C.

To get the utmost intake of all of the device at the graphics card twisted energy prohibit to 325Вт.

The most intake of the entire device below this type of load, many weren’t acquainted 763Вт.

In the tip, if you happen to come with PBO or podrzavati, you’ll be able to nonetheless 100 and a part Watt to throw. In General, when I used to be doing all my calculations and measurements at the 700w — it seems, isn’t up to now off – handiest 8.5%.

Okay, time to transport directly to the temperatures. Load handiest the CPU. Let me remind you that every one acceleration is grew to become off.

And thus, the model “vintage”:

For Windows customers a bit acquainted to look such difficult screenshots.

If you glance from left to proper:

1. Testimony tracking, which was once maintained on the time the exams are run. Okay10temp — temperature of the processor. On a Board two chips of tracking. Here you’ll be able to see handiest one chip. Therefore, the velocity of the CPU cooler and one from the corpus we will be able to handiest bet. Now in fact improve for the second one chip set, however to retake exams won’t.

2. It seems like working High95 in Linux

3. Frequency of processor cores. With the readings in Windows, they didn’t correlate. The software calculates those frequencies. I will handiest say that, on moderate, the frequency of the displays underneath – about 150-200MHz on Windows.

4. Settings the video output of nvidia-smi.

5. Small propcheck iron. Someone right here wrote that no processor I’ve. 🙂

Dear haters, the processor is! Since the announcement Threadripper 3990x my already was a pumpkin, then began saving cash from the varsity Breakfast 3990x. 🙂

Okay, jokes apart. What will we see?

And we see relatively excellent and in my opinion I anticipated the location: well-planned, designed and examined air-cooled and the frame might nicely take care of the cooling of the processor with out overclocking. Let me remind you that the utmost temperature for brand new Threadripper-s — 95гр. With, subsequent is a reset of frequencies. I’ve in a sealed enclosure grew to become out 82,2 gr. C. a small inventory of 10-12 g. In concept, this newsletter may just finish, as this load is 90% of my wishes. But every so often it so occurs that the processor and graphics card are loaded concurrently and the eyeballs. So take a look at the location.

Here the location is way much less rosy. The card feels nice, and the processor is heated to 87.5 grams. S. This is in fact legitimate, however too tall.

Decided “to reserve” to check in Windows 10 (LTSB). The handiest load at the processor.

As you’ll be able to see from the screenshots, Windows for High95 heats the CPU a bit higher. But Hwinfo displays that trolling no. But those indications can’t be thought to be a reference, as I’m now not a large Customizer Windows. The motive force didn’t put, meal plans aren’t set, all of the main points of settings now not know. Just the Windows updates set to run as soon as High95.

In General, the temperature I satisfied, can are living. But sought after to test what the temperature can be within the type of “Changeling”. For this I needed to disassemble all of the device and reassembled. The first time was once accumulated correctly, because it was once now not sure that I can disassemble.

Option “Changeling” was once attracted via the possibly decrease temperatures at the CPU when loaded processor and graphics card. And “Changeling” met all my expectancies.

Option “Changeling”:

If loading the processor, the temperature diminished now not considerably. Was 82,2 gr. With, was once 82 grams. C. Not nice win.

But the whole thing adjustments when within the device concurrently loaded processor and graphics card.

All as sought after, even higher.

The CPU temperature is at 87.5 grams. With, was once handiest 84. As you’ll be able to see the video card had nearly no impact at the CPU temperature. And the video card warmed up via handiest 2 levels upper.

The cooler at the processor started to paintings a lot quieter. However, it’s my opinion in all variants didn’t cross to the max as a result of within the BIOS the max for it was once written at the 90g. But the device within the variant “vintage” was once noticeably extra noisy.

Conclusions, in keeping with the per 30 days utilization revel in:

1. The fan at the Silver Arrow TR4 despite the fact that tough, however at 2500 rpm is relatively noisy. So I changed it a pair slow-moving. But that is any other tale. Talk one at a time. The exams within the article is in fact performed with the local fan.

2. As I anticipated, nicely deliberate air cooling copes with cooling Threadripper 3970x with out overclocking.

3. Managed to do all 4 chassis enthusiasts.

4. I love the whole thing, I’m satisfied. 🙂 But the reminiscence was once in a position to earn handiest 2733МГц. But AMD informed us that they’d advanced running reminiscence. In General, be expecting a couple of releases of the BIOS, if now not proper, I modify to 128GB UDIMM. 64GB nonetheless every so often isn’t sufficient.

5. Noise. Any particular measurements now not performed. There is handiest subjective. To admire noise to be anxious / now not anxious / do not listen. In the case of the local fan at the CPU — pressure. In the case of low-speed two — now not anxious. I listen a rustle. What to the infamous jumps of temperature — there’s not anything intelligible cannot say, for the reason that laptop is continuously loaded. Jump every so often, however because the device is beautiful quiet, the trade of the rotation pace of the enthusiasts anxious.

Headphones noise cancelling noise can not listen. 🙂

6. If you’re going to collect a identical device to you laziness one thing to believe, check and Refine, and you’ve got a sizzling area, I’d counsel to be aware of liquid cooling. Having the chance of leakage and all programs fail, you are going to scale back the temperature 10 levels. Might scale back the noise. But finally, the device on such tough processors might not be utterly silent. Perhaps after a while there can be new air cooling device custom-tailored to new Threadripper-s.

7. 64GB of reminiscence will not be sufficient. Anyway, I had little kernels, I for the yr a few instances confronted with a scarcity of reminiscence. And with 32 cores per thirty days 4 instances already needed to optimize their systems. Therefore, I will counsel to start out from a minimum of 1.5 GB at the waft – 96Гб. Reasonable 2GB at the waft — 128GB. However, the whole thing depends upon your goals. In “titrable” have you ever noticed the video chuchkov that disperse to 16GB of reminiscence on Threadripper-Oh, and really proud. But, for me, this configuration is admittedly now not usable for such processors.

8. Someone within the feedback mentioned that there are issues putting in Linux on NVMe drives. I responsibly claim: no issues of Linux set up on NVMe drives no! Talking about Kubuntu 18.04. However, there’s a small drawback with the set up — default kernel distribution bundle isn’t loaded. Need to put in the boot loader within the kernel parameters, specify “mce=off”

Now I exploit kernel 5.3.18. In General, all this is wanted works. With the exception of the sound by way of the SPDIF however it’s one thing I want. But promise handiest within the nucleus 5.5. So whilst the use of Bluetooth. Analog sound isn’t any drawback. Everything works — paintings each your sound card. There was once an issue with some “unique” bluetooth headphones — the primary time attached as a “mono headset”, the second one time the whole thing is okay. The drawback is solved via putting in the most recent model of the blues. I amassed from supply, however they are saying there are in programs.

9. Uptime (with out reboots). While not anything a lot is alleged, for the reason that device is within the means of putting in and exams, so there was once a wish to restart. While the file — week, totally loaded, all bodily cores — load moderate 33-38. Overall, a cheap advice at the parameter load moderate for Linux is <collection of bodily cores> multiplied via 1,25. Or <the collection of bodily cores> plus 25%. Can be shipped and above, nevertheless it applies to optimized code, for instance, the renderers (rendering in Russian).

10. VRM warmth within the running and most quite a bit does now not exceed 60g. Heating of the chipset additionally does now not exceed 60g. The enthusiasts aren’t spinning.

Behind the scenes.

Further exams carried out in Blender within the phoronix check suite. Temperature now not got rid of, however reminiscence — every so often a few levels may well be upper than in High95.

You can examine at the hyperlink.

My numbers grew to become out such:

Scene BMW27 — via 45.11 sec.

Stage Classroom — 114,99 sec.

The Pabellon Barcelona scene — 148,45 sec.

My device was once slightly slower than a Phoronix’. Perhaps because of the reminiscence pace, possibly on account of the off PBO. And possibly on account of reminiscence and PBO in combination.

While fidgeting with putting in the device, ran of their duties with enabled and disabled PBO — the adaptation is no more than 1% and the temperature above 5 levels. In General, grew to become off. No PBO is quieter and cooler. AMD did an excellent processor which acceleration is in concept now not essential, as no vital build up in efficiency.

Previous portions:

Part 1

Part 2

Part 3

Part 4

Part 5

PS the Site is closely compressed photos in some photos can endure within the legibility of textual content. If anyone is extremely essential, write in feedback, I can throw off the unique.