IBM Power11 CPUs Launching In 2025 - Linux 6.13 Preps KVM Nested Guests For Power11

[coder]

I didn't say Amigas are alive, I said AMIGAOS is alive. There's a very stark difference between software and hardware development, you know?

It's funny how I'm learning far more about Amiga/AmigaOS in this thread than about POWER.

[Vistaus]

]n1508317]
So, the newest, best machine it officially supports is a 2.2 GHz quad core from 10 years ago? That doesn't seem very "alive" to me. Most of the companies that have had anything to do with Amiga are defunct or at least in some kind of precarious financial situation.

What's so great about it, other than nostalgia?

I didn't say Amigas are alive, I said AMIGAOS is alive. There's a very stark difference between software and hardware development, you know?

AmigaOS 4 is a very modern OS, it even has a built-in update manager and software store. It receives updates every week.
It can also run Linux apps through a nested X server and Qt apps natively. It's super customizable, even more so than KDE, and has a well-thought out look and feel. And ye olde AmigaOS 3 is kept alive by them as well.

I don't know exactly what nostalgia has to do with it, because at least in my case, I started using it in 2012 as a secondary OS. Never used it before 2012. And I'm currently 32 years old, so not exactly an old geezer either (and I was 20 when I bought my first Amiga with AmigaOS 4.1).

[mos87]

I wasn't aware that the PowerPC was still being developed.

I wasn't aware this shtick is still going. Maybe it's some sort of right of passage who knows

[coder]

Still don't have a clue as usual.

If that were true, you could actually post a correction instead of merely feigning superiority.

[brad0]

The only current Debian downloadable POWER images I see appear to be for little endian.

And if you look at the Gentoo options, it certainly appears like they're headed towards little endian, with two of the three PPC/POWER options being for LE and the latter two focusing on modern hardware. There BE download sounds like it's oriented mainly towards legacy support.

https://www.gentoo.org/downloads/#ppc

Still don't have a clue as usual.

[brad0]

Steve Jobs, the PR genius, caused great harm to PowerPC consumer/public image. What he meant was, IBM completely ignored Apple's specific needs and they broke their promise of higher Mhz G5. Nothing more. The CPU isn't fit to mobile, simple as that. Otherwise, they passed 5ghz a couple of years later with a new generation. I remember submitting that news to slashdot at that time.

Apple did the same thing again, to Intel. Public seem to figure how hot/power consuming Intel is right after their M1 switch.

Irrelevant rambling mess.

[brad0]

As if Amiga the company wasn't dead.

That's completely irrelevant.

[coder]

You're making comparisons using POWER9 (2017) and POWER10 (2020). No shit they're behind. POWER11 comes out next year, meaning those generations of POWER are nearly EoL. Would you like me to compare them against an intel Kabylake server and see how they do? No?

The main comparison I made was a contemporaneous one with Zen 2, which launched in 2019. My point about socket-scaling was that it's no longer so important like it used to be.

OMI had higher bandwidth than any DDR technology that existed when POWER10 came out.

Doesn't matter, because it couldn't scale in performance and had worse energy-efficiency. That's why it died.

Frankly, it's an embarrassment that DDR still uses a parallel interface

This just shows you don't understand hardware.

OMI was a massive step in the right direction. No shit it's dead, it's being replaced by CXL which wasn't a thing at the time

See, this again shows you don't even know why OMI died. CXL didn't kill it, because CXL was merely an idea at the time.

CXL.mem will happen, but only after on-package memory becomes the norm for server CPUs. At that point, CXL will be used to scale capacity, while on-package DRAM and memory tiering efficiently scales performance.

it's very likely POWER11 will ship PCIe 6.0 next year. Any ideas when AMD well get around to that? No?

Next year would be my guess. They're due for a new server socket and CXL 3.0 uses the same PHY standard as PCIe 6.0.

Those POWER10 cores each do 8-way SMT btw,

Only up to 15 cores, according to what I'm reading.

so you're talking about 240 POWER threads vs 256 from AMD.

Leaving aside the above discrepancy (120, by my count), I think it's funny you're trying to equate SMT threads as if each POWER10 core has 4x the backend throughput as each Zen 5 core (or Zen 4C, I'm not sure which AMD CPU you're comparing against, at this point).

The reason to go extra wide on SMT is usually as a substitute for a deep reorder buffer. Otherwise, having so many threads is just a waste of die area. GPUs have lots of SMT threads, but they're typically in-order.

That narrows the gap significantly, and those POWER cores are being fed by memory bandwidth that AMD could only dream of.

How many nodes are you talking about (I gather there are 4 CPUs per node) and what's the Stream TRIAD score for it?

Interestingly, it recently came out that AMD made a version of the the MI300 with just CPU cores + HBM. Microsoft is using it and getting 1.75 TB/s per 96 Zen 4 cores.

No, you don't. Not at the scale POWER machines operate at. That's joe blow consumer logic. That sheer number of sockets should have tipped you off.

The constraining factor on CPUs at this scale is power dissipation.

Data movement wastes a ton of power, though. I think that's among the reasons the industry prefers fewer sockets. Plus, it seems like those 15-core Power10 CPUs have a base clock speed of 3.55 GHz. If we use the somewhat contemporaneous example of Sapphire Rapids, the 60 core Xeon 8490H (which scales up to 8 sockets) has a base clock of 1.9 GHz. I'd take 60 cores at 1.9 GHz vs. 15 cores at 3.55 GHz any day, especially when those 60 cores have AVX-512 and AMX.

They have so many cores crammed onto one die and their cores are so inefficient that essentially every core needs to fall back the instant one or two P-cores start to turbo.

FWIW, Sapphire Rapids has about a dozen chiplets. Emerald Rapids uses basically the same node but has just 3 tiles and can run 64 cores @ the same TDP with the same base clock.

By spreading out the system across more sockets and making use of high-bandwidth OMI-style links between them (technically, it's a different fabric between sockets) you get more cores, more cache per core, more system memory, higher sustained clocks per core,

If you have any substantial data sharing between those cores, performance will be worse than having the same (or more) cores in fewer sockets. 3.55 GHz really isn't that high, considering the Xeon 6458Q can do more than twice as much (32 cores) at 3.1 GHz. AMD's EPYC 75F3 could run 32 Zen 3 cores at 2.95 GHz, but in TDP of just 280 W (Intel Sapphire Rapids' is 350 W).

And if you don't have substantial data sharing, then you'd probably be better off just using multiple dual-socket machines.

Pretty much the only argument for having 16 sockets is to host lots and lots of RAM, but even this use case will fall away when switched CXL arrives.

The result is that current (soon to be replaced) POWER chips have been able to hit top speed across all cores on all sockets at the same time, something AMD only just caught up to.

Back in 2020, AMD could also hit that speed, if you limit the CPU size to just 16 core EPYCs.

When was the last time you saw an AMD server with 16 TB of ram in 2020?)

AMD mainly shoots for mainstream use cases, which is why they stop at only 2S scalability. However, the Intel Xeon 8490H, that I mentioned above, can host 4 TB per socket and has 8S scalability. If you want to go back even further, the Xeon 8280L had 8S scalability and claims support for up to 4.5 TB per socket (although I suspect some of that is using Optane DIMMs).

BTW, how much power do these POWER10 CPUs actually use? You keep talking about running these cores "flat out", but the main reason Intel and AMD don't do that is that it'd burn too much power. Nvidia's H100 burns up to 700W of power, so it's clearly possible to dissipate that much heat from a server socket, if you really wanted to. As a matter of fact, I'm pretty sure I've read of multi-chip mainframe CPUs that burned as much as 2kW from a single socket!

[coder]

AmigaOS is very much alive and has been using PowerPC on new hardware for years.

https://www.amigaos.net/content/72/supported-hardware

So, the newest, best machine it officially supports is a 2.2 GHz quad core from 10 years ago? That doesn't seem very "alive" to me. Most of the companies that have had anything to do with Amiga are defunct or at least in some kind of precarious financial situation.

What's so great about it, other than nostalgia?

[hiryu]

Raptor have hedged their bets as best they can. If you look over their twitter timeline, they're a lot more positive about POWER11.

Making anything at all is absurdly difficult, even moreso if it has to be fast (just ask Red Semi, who are ALSO supposedly making an OpenPOWER high-performance CPU, and have been for many years now). If Solid never deliver, then it has no real impact on Raptor since they need the chip to even start development of their own systems. It costs Raptor nearly nothing to hedge their bets in this way.

We shall see! At least with IBM, they're a known quantity. We know they're going to deliver a minimum a solid server CPU. I've been very happy with POWER9, but it has gotten very long in the tooth!

I haven't been able to find out much about Solid Silicon in general, which is a bit of a yellow flag for me.