13-Way IBM POWER9 Talos II vs. Intel Xeon vs. AMD Linux Benchmarks On Debian

[curaga]

Aww, all this teasing. I'd really like one of those 18-core POWER beasts, and they've talked about additional cpu options in Q3. Maybe a sweet spot between 8 and 18? The idle power usage seems a bit high right now, and with just two pci-e slots (without the bifurcation support), I think I'd need to grab a pci-e to several pci adapter, then put in a sound card and a SATA card via pci (keeping the other pci-e for a gpu). Any Raptor folks around to comment on these? It's still way over my budget, but with IBM's bounties it might be possible to get one, and then do a few more with local testing ability.

[willmore]

Or people how to use OpenBenchmarking.org

I figured you'd understand my reference.

[bridgman]

You claim PC users would not buy a GPU like that without "video encode/decode technology." and without "DRM" but I am a PC user to and i would buy a closed-source microcode free product. And i would use it for Graphics and Video to.

I claimed nothing of the sort. What I said was that the high volume OEMs who account for most of our sales would not buy them.

this means you(AMD) plan to manufacture REAL HARDWARE without Evil Treason to the user/customer and you need to call it "datacenter-only" or the Capitalists in your country kill you for this move...

and in the end we can have a Desktop PC without Evil Treason ...

i call this big news... how long do we need to wait for this?

No, it means nothing of the sort. What I said was that I had been LOOKING INTO THE POSSIBILITY of identifying one or more products that could allow open microcode without being a money-loser, and that so far it hadn't gone very well.

[bridgman]

you want a compute only card to be a microcode free solution to limit "money-loser" we understand that ...

any functionality what goes beyond the horizon of this functionality (Compute only) is called a money-loser in your world.

Actually no - "money loser" is a function of development costs vs gross margins... nothing to do with definitions, horizons or functionality. The problem is that developing a GPU which could have open microcode without putting our current (OEM PC) products at risk is quite high, needing something close to a billion dollars of sales at current margins to recoup the development costs.

[DavidKL]

I registered to respond to some of the comments here mainly. However, I also have a basic question about the article:

Which Samsung SSD was used in the test system? We know which one was used for the IBM tests (960 EVO) but not for the rest.

As for the comments:

1) Stop solely blaming the marketing department. Blame the people who manage them even more. Yes, marketing's role is to confuse people with emotion — to sell them dreams and feelings instead of solidity. However, marketing extends right to the top of a corporation. Lisa Su, for instance, is marketing when she goes in front of the public to speak about AMD. People rely on emotion to justify continuing to live (i.e. people need myth to give them purpose), so marketeers can soothe their consciences a little.

2) All corporations are evil by definition. They exist to sell more for less. That means confusing people enough to part with more of their money than they should. That is the basis of profiteering. And, guess what? Blame the public. Blame everyone. In a rational culture this kind of corruption certainly wouldn't be not only condoned but promoted as the backbone of the economy. The corruption is justified by Social Darwinist elitism. Unfortunately, instead of meritocracy — humanity typically creates plutocracy. Resources, in the latter, are not distributed according to merit. The two are not all that compatible, although all "ocracies" have some meritocracy involved by absolute necessity.

3) It is hyperbolic, and simply false, to claim that parts that OEM corporations will sell must be loaded with anti-consumer features, like black box spyware/DRM. Firstly, it should be possible to disable those things, in hardware, during chip runs without driving up costs beyond the benefit of giving people the ability to buy what they want. Imagine that? Consumers being able to buy the product they want, one that actually works for them instead of for 3rd parties. Crazy idea, I know. Secondly, what about enabling OEMs that want to sell what consumers actually want to compete with those that don't? Apple was on the verge of going out of business. At least, it was barely relevant. Now, look at how huge it is. Corporate dynasties rise and fall. Simply because the status quo exists doesn't mean it justifies itself.

Regardless... sure, let OEMs like Dell, ASUS, and Apple sell their DRM spyware packages so they can profit in a variety of ways, including via government happiness. However, give the DIY market, especially the prosumer (e.g. Threadripper) subset of that market the ability to control the products. That means have them be fully documented and fully controllable. If you must have a spyware black box built into products to satisfy governments then consumers need to know that. It needs to be labeled as that. Other governments will likely then demand documentation.

4) Someone here said don't assume computer users are stupid. Well, all people are stupid in a lot of very serious ways (hence the heap of corruption known as humanity, both today and in the past). And, also, the fact that IQ has a fairly wide range. However, more to the point is the advice to not assume the readers of technical articles are stupid. Don't try to confuse us with marketing BS. Just speak the truth as you know it, plainly.

5) Does performance-per-dollar get complicated by various deals made between server OEMs and customers? I don't know much about this market but I know that a serious liability of performance-per-dollar in the DIY segment is sale pricing. AMD GPUs (e.g. 470 and 290), for instance, were available well below MSRP when on special sales (i.e. slickdeals front page listings), prior to the latest crypto wave. While I doubt there is that kind of volatility in enterprise pricing I wonder how much sweetheart deal-making can come into play.

6) Don't tell others what they're allowed to talk about. I saw at least two posts that were complaining about subject changes, referring not to the overall article but, instead, to smaller matters mainly brought to the fore by forum posters. People are allowed to talk about whatever they like that is related to the subject matter of the article. Likewise, readers such as myself are free to decide what angles are interesting to us when we read all the different comments. If it doesn't fit your investment into a particular axe being ground that's too bad.

7) In light of point #3... Like the ridiculous outrage over CTS ("We absolutely must muzzle the tech press like Google does!"), the popular notion in computing circles right now is that, since knowledge is power — why should any consumers want any?. Instead of being able to leverage that power in their favor, consumers are supposed to feel safe via the bliss of ignorance — protection from being able to leverage power for their benefit. So, instead of a free tech press, we're supposed to just love the idea that insiders can dump stock and hackers can exploit our systems — for whatever arbitrary amount of time megacorps/governments set. Google, let's all remember, was caught hacking into iOS and OS X to install a spyware payload — the kind of behavior that would send actual people, like 18-year-old "kids", to prison. Sony used the bogus "bad apples" defense ("no one cares about ___") and other blatant ad hominems in court to defend its bait-and-switch Linux on PS3 decision — the same company notorious for its DRM. But, corporations aren't just people. They're superior people, according to today's corruption status quo. Not only do they deserve the right to break laws as long as the penalties are smaller than the gain, they are supposed to replace the free press.

Are we all captured, we tech enthusiasts? Are we satisfied with being passively captured or are we going to leverage our power to keep what we have and perhaps gain a bit more? I can't answer that for you but I can tell you that I don't think naiveté is going to win you much in our world (which is dominated by Social Darwinism). And, even some altruists will argue the elitist point of view, where Pollyanna doesn't deserve all that big a pie piece. Better save some for the brighter DNA. Put simply: If you don't fight for yourselves, who will? Corporations are not people, my friend.


Also, I have to disagree about the aluminum cooling fins. It seems, from what I've read, that they're really ideal from a cost-benefit perspective for quite a lot of different cooling solutions, usually as a supplement to something else. Copper is a bit nice but tends to lose its shine when one has more physical space to put aluminum. I can agree with the argument that aluminum cooling fins should only be used when they are efficient from a noise-cooling perspective and in terms of throttling avoidance. They shouldn't be used as a way to fake-out consumers by giving them tinnitus and/or throttling equipment.

Also, since it looks like memory performance is an area where Power9 shines, are there things that could be benchmarked (that are relevant enough) that would show this advantage more than what's already here in the article? These systems don't have the Centaur L4 eDRAM, correct? I just read an article about how those systems have yet to hit the market: "IBM Readies Big Iron with 'Cumulus' Power9 Chips":

"Now, it is getting close for “Zeppelin” and 'Fleetwood/Mack' to take the Power9 stage with the “Cumulus” Power9 variant, which sports up to 12 cores per chip and eight threads per core and which also makes use of the 'Centaur' memory buffer chip and L4 cache memory to provide roughly twice the memory capacity and bandwidth per socket."

[DavidKL]

I tried to post but it said it's unapproved. Hmm...

[bridgman]

I tried to post but it said it's unapproved. Hmm...

I believe new users (under 5 posts) go through moderation automatically. You're almost there

Occasionally the same mechanism kicks in for people with thousands of posts as well, but I haven't seen that happen as much recently.

[PeterB]

I'd be very interested to see the results of a floating point benchmark built to use quad-precision. I gather Power9 implements this type in the hardware.

[poorguy]

By the way @Michael,

I noticed that your benchmark of Zstandard is only single-threaded.
Fair enough, but know that you could *also* test zstd in multi-threading mode,
with just a very small command line update :
-T2 for 2 threads, -T4 for 4 threads, or -T0 for as many threads as there are cores (which can be handy when different packages feature different amount of cores).

Typically, your `install.sh` benchmark script would become :
```
#!/bin/sh

tar -xvf zstd-1.3.4.tar.gz
cd zstd-1.3.4/
make
cd ~
cat > compress-zstd <<EOT
#!/bin/sh
./zstd-1.3.4/zstd -19 -T0 ubuntu-16.04.3-server-i386.img > /dev/null 2>&1
EOT
chmod +x compress-zstd
```

For reference, I tested it on an ARM server platform, and discovered that the increase in nb of cores largely compensated for the small decrease in single thread performance compared to pricier Intel Xeon packages. Granted, it takes a well designed multi-threaded application to get there. But zstd seems to be one of them.