An Initial Look at the IBM POWER9 4c/16t CPU Performance on the Blackbird | Hacker News Reader

An Initial Look at the IBM POWER9 4c/16t CPU Performance on the Blackbird

phoronix.com

45 pointsMrRadar18 comments

The interesting thing about Raptor's Power9 boards is their emphasis on open firmware.

But I'm disappointed that the Blackbird wasn't at a lower price point (such as if IBM heavily subsidized it, to try to get more open source labor supporting their architecture). It's hard to get excited about paying big bucks to volunteer for IBM.

Personally, I'm now more interested in RISC-V boards that are even more open than what Raptor has been doing with Power9. Something like the Raspberry Pi, but RISC-V and more open, would be awesome.

> But I'm disappointed that the Blackbird wasn't at a lower price point

This sentiment is why I believe open systems, like the Raptor or RISC-V, are doomed to never become truly popular. There are too many people more concerned with cost and/or performance than freedom.

There's a balance. For example, buying into a huge black box component/architecture effectively controlled by IBM is limited freedom for your dollar.

Does having four threads per core cause more problems for speculative execution attacks? I seem to recall that one of the fixes proposed for Spectre was disabling hyperthreading.

Those attacks can be mitigated. They're not fatal to hyperthreading. HT is a great way to improve performance and is not going away.

They're fatal in the sense that you can't mitigate all the side-channels. It's intrinsic to HT/SMT that certain critical resources are shared without an ability to mask the side-channels. What you can do is make sure that only processes and threads with the same privilege (e.g. same UID) share a core. But AFAIK such mitigations haven't made their way into OS schedulers.

You're almost certainly right that HT isn't going away, the performance benefits are too great. But to fully minimize the security impact operating systems in particular, but also some userspace applications (e.g. browsers), will need to be aware of the problems. This is already the case more generally for cryptographic algorithms and similar software--e.g. how memcmp can leak passwords.

Unfortunately same UID is not sufficient, if my program is running some other script or respnsing to input provided by an external user or attacker. JavaScript is the easiest example of this.

It depends ... AMD seems to not require disabling SMT due to a different implementation than Intel.

Am I missing something or were those benchmarks really as brutal as they looked? It was beaten by a Core i3 and lower end first gen Ryzen chips.

"For those curious about the performance potential for low-end POWER9 parts compared to the more common high-core/thread count POWER processors we have benchmarked before..."

They are benchmarking a low end 4 core part that sells for $375. They have 24 core Power9 chips. I believe, though, that they all lag current Intel and AMD parts. Their market is mostly mainframes and high-reliability specialty servers.

I thought that they were wonderful few years ago, compared to those 4 core CPUs. Now, that we will have 64 core ThreadRipper soon for quite reasonable price, that's a very different landscape. AMD really is changing the world. What a time to be alive.

IBM advertise their "big iron" systems in Transactions Per Second. A measure of the total system performance; IPC, Buses, caches, etc. They get shit stomped in a traditional CPU bound benchmark. Remember Apples marketing using "The Megahertz Myth" in their later PPC days because of how one sided benchmarks (and clock speeds) were getting?

It's a fine arch with plenty to offer but a 1:1 comparison with an x86 desktop is like saying that a 50 ton bulldozer just got wrecked in a race with a Bugatti Veron.

The bulldozer would even lose from a Fiat 500 or a bicycle if it came to raw speed.

It's a lower-end part, but regardless, POWER9 isn't really anything special in terms of instruction throughput / calculations (it's only got 4-wide SIMD for example), and its FP performance is not great.

Where it really shines compared to x86/64 is memory bandwidth / throughput.

Now that modern POWER hardware can be bought for roughly commodity prices, how likely are optimizations to these workloads?

It is unlikely that significant optimization will happen outside of companies payong for it to happen, most of the tinkerers that do this type of optimization for fun are working on projects like Armbian, as a $15 SBC is a more useful target than a piece of big iron.

What is that black cube on top of the heatsink?

Probably a piece of rubber that keeps some margin between the heatsink and the cover of the server.

Reply on news.ycombinator.com