The POWER and PowerPC General Discussion / News Thread

Hardware CPU
, ,
872

Is there anywhere I can read docs on these chips? A lot of my info is based on other sources and info from a lawsuit that happened 4 ish years ago, so some of my memory could be hazy. However, I don’t forget documentation. I can save that :slight_smile:

Yeah I don’t have that sorta info. And by “Not having control of their fab” I meant, they literally didn’t own it anymore. It was its own company, and with things being split up like that, the logistics of the company moving forward was to fill the empty spaces that made sense in the market.

i think we’re on the same page we just both have different levels of understanding.

873

Blackbird does not use POWER8 chips

The original Talos I design would have used POWER8 chips, but it did not succeed in crowdfunding on CrowdSupply. It was a dramatically different board, and even needed to embed IBM’s Centaur memory buffer because POWER8 was not designed to interface with DDR RAM directly at all.

Both Talos II and Blackbird boards use POWER9 Scale Out (native-DDR4 not Centaur) chips, packaged in Sforza processor modules, socketing into a 2601-pin LGA-style socket.

There was an effort by Raptor to design a third POWER9 board (Condor) that would socket the larger LaGrange module, still containing a Scale Out (DDR4) chip, exposing IBM’s fancy OpenCAPI interface rather than purely PCIe, but that was cancelled.

1 Like
874

I’m honestly not sure where we can read about it in specific. I have been in the scene quite a long time and keep up with the markets on a daily basis.
Maybe there are some tech history websites, but I am a history buff and like to do deep dives into these things as well.

875

Please write some posts!

876

Depends on what you want to learn about. I can’t just write posts randomly without any context lol

877

Sure you can, this forum is crawled by search engines fairly well, so any bit of information would also benefit the world wide hivemind those who do not frequent the forum. If you worry it would make too meandering a thread, you could always post it in the #blog category. #wiki is a good choice for particular kinds of informative posts as well.

For example, I had a thread in the #blog section keeping track of orbital launches last year; 2021 Orbital Launches (COSPAR quick reference) perhaps a bit niche, but I thought it could be of interest to others.

I would say, however, that further discussion about AMD chips probably belongs elsewhere than this thread.

1 Like
878

Apparently suspend to RAM is not currently available on POWER9 boards, but the newer revision of Blackbird boards (1.02) makes it at least feasible.

3 Likes
879

It would be a bit curious why a workstation / server mobo needs to be able to hibernate, that’s a feature for laptops. Actually more curious why they are even trying to implement this.

1 Like
880

presumably so you don’t have to reboot and start up all your stuff from scratch every time

the alternative is running your desktop 24/7

2 Likes
881

Not exacty, the floating point/SIMD engine was a shared pipline, otherwise seperate L1 caches, decoders, reservation station, and integer/address pipelines. Some workloads suffered greatly, others ran quite well.

5 Likes
882

Hi, I am believe something has become corrupt on my POWER8 system it will boot a crippled version of petitboot and nothing changes when i flash it anyone have ideas. Its a TN71-BP012 Habanero system
log.txt (28.1 KB)

883

You need to give us more information about what kind of “crippled” behaviour you are seeing.

Regarding your log file, while I have not booted an IBM POWER chip myself, and I do not know what error messages are considered normal for Habenero boards, this is where I would start looking if I had not other information:

-START OF FILE-
  2.00640|ERRL|Dumping errors reported prior to registration
…
 33.50456|ISTEP 14. 3
 33.74930|================================================
 33.74930|Error reported by unknown (0xE500)
 35.12626|================================================
 72.84226|ISTEP 10. 1
…
 74.10333|ISTEP 14. 3
 78.40002|ISTEP 14. 4
…
[50175362446,5] FDT: Parsing fdt @0xff00000
[50176726814,3] DT: dt_attach_root failed, duplicate sensor@58
…
50247388963,5] PLAT: Detected Habanero platform
[50273063231,5] CENTAUR: Found centaur for chip 0x0 channel 5
[50273141324,5] CENTAUR:   FSI host: 0x0 cMFSI0 port 6
[50279353930,3] MFSI: Error status=0x00018000 !
[50355752886,3] CENTAUR: MFSI write error -6 writing CMD
[50359243977,3] CENTAUR:   FSISCOM error -6 reading ID register
…
[1618937594,3] NVRAM: Partition at offset 0xffff0 extends beyond end of nvram !
[1619128477,3] NVRAM: Re-initializing
…
[3644302379,3] PHB0: Base location code not found !
[3646325219,3] BT seq 0x03 cmd 0x48: Retry sending message
[3729381845,3] PHB1: Base location code not found !
[3791801471,3] PHB2: Base location code not found !
[3853684239,5] PCI: Resetting PHBs...
[4928972951,5] PCI: Probing slots...
[4929090914,5] PHB0:00:00.0 [ROOT] 1014 03dc R:00 C:060400 B:00..00
[4929243336,5] PHB1:00:00.0 [ROOT] 1014 03dc R:00 C:060400 B:00..00
[4929746911,5] PHB2:00:00.0 [ROOT] 1014 03dc R:00 C:060400 B:00..00
…
[11367266859,3] FLASH: No ROOTFS partition
…
[11582518318,3] OPAL: Trying a CPU re-init with flags: 0x2
[[11601604493,3] BT seq 0x05 cmd 0x06: Retry sending message
…
[    0.000000] Kernel command line:
…
[    0.687594] pci_bus 0000:00: root bus resource [bus 00-ff]
[    0.687780] pci 0000:00:00.0: bridge configuration invalid ([bus 00-00]), reconfiguring
[    0.687884] pci 0000:00:00.0: PCI bridge to [bus 01-ff]
[    0.687942] PCI: I/O resource not set for host bridge /pciex@3fffe40100000 (domain 1)
[    0.688014] PCI host bridge to bus 0001:00
…
[    0.688169] pci_bus 0001:00: root bus resource [bus 00-ff]
[    0.688343] pci 0001:00:00.0: bridge configuration invalid ([bus 00-00]), reconfiguring
[    0.688553] pci 0001:00:00.0: PCI bridge to [bus 01-ff]
[    0.688611] PCI: I/O resource not set for host bridge /pciex@3fffe40000000 (domain 2)
[    0.688679] PCI host bridge to bus 0002:00
…
[    0.688833] pci_bus 0002:00: root bus resource [bus 00-ff]
[    0.689011] pci 0002:00:00.0: bridge configuration invalid ([bus 00-00]), reconfiguring
[    0.689115] pci 0002:00:00.0: PCI bridge to [bus 01-ff]
[    0.689223] pci 0000:00:00.0: PCI bridge to [bus 01]
[    0.689354] pci 0001:00:00.0: PCI bridge to [bus 01]
[    0.689412] pci 0002:00:00.0: PCI bridge to [bus 01]
…
[    4.028974] ipmi device interface
[    4.029019] ipmi-powernv ibm,opal:ipmi: Unable to map irq from device tree
[    4.030945] ipmi-powernv ibm,opal:ipmi: Found new BMC (man_id: 0x000000, prod_id: 0x0000, dev_id: 0x00)
…
[    4.187527] netconsole: network logging started
[    4.189003] Freeing unused kernel memory: 6784K (c000000001040000 - c0000000016e0000)
[    4.202399] udevd[2213]: starting version 3.0
-END OF FILE-

Overall, it looks like a fairly successful boot; by the end, 80 threads are started, I see multiple USB controllers, and OCC seems to be functional. I really need to understand what problem you are seeing.

For all I know, it is normal on Tyan Habenero platforms to see:

  • ERRL errors
  • failure of ISTEP 14.4 the first time
  • a duplicate sensor in the DeviceTree
  • Centaur errors for MFSI & FSISCOM
  • NVRAM partition 0xffff0 errors
  • PHBs 0-2 failing on the first try
  • a missing FLASH ROOTFS partition
  • OPAL re-initialising the CPU with flag 0x2
  • an empty kernel command line
  • PCI Bus 0000-0002 needing to “reconfigure”
  • an IPMI IRQ mapping error
  • log file ending with udevd after network logging started
  • as well as two “BT seq“ messages needing to be re-sent.
2 Likes
884

Everday integer functions ran quite fast, but floating point suffered. Javascript was the biggest killer of Bulldozer based parts in terms of web performance.

2 Likes
885

I am still on Poor-dozer for my desktop. IT was a design ahead of it’s time. And with all of the mitigations in place on the Intel hardware due to exploits of their [Intel’s] Hyper-threading implementation, Bulldozer has proven to be a competitive platform, especially now that software is being programmed to be multi-threaded and task schedulers are actually aware of hardware differences. Bulldozer shined best on not MS Windows systems.

2 Likes
886

True, I built a few linux servers based on the Bulldozer Era Opterons and they worked quite well. Very efficient for the smaller server tasks or parallel server tasks since most of it was integer workloads vs FP like games on desktop.

2 Likes
887
3 Likes
888

It is excellent to hear that, though somewhat disheartening to see how long such work takes on a donated budget.

If there is a sequel up project in the future, I wonder if we would see it as a drop in replacement for a frame.work chassis,
as I wrote this I decided to check the comments on the blog post:

Roberto Innocenti
2023-07-13 09:48
… in fact we are in contact with Frameworks, we are interested in future collaborations with them, …

Not that it should really be a concern at the moment, changing chassis now unnecessarily would be a grievous waste of funds.

2 Likes
889

what is the state of POWER these days? It seems like a lot has happened since this thread was first started.

I touched on some POWER servers at work a number of years back, actually about the same time as this thread was started, and it looked promising but we had issues with the software ecosystem. IBM sold us on the idea that we could use GitHub - biobuilds/biobuilds: Curated collection of open-source bioinformatics tools for all of our software needs, but when we started trying to port over some workflows, we immediately hit issues with the POWER built versions of the software not working. Looking at that git repo it has not been touched in 5 years, and https://biobuilds.org/ now points to some other non-IBM company, does not give me much hope. Are things still progressing in other fields for POWER?

890

One of the reasons my enthusiasm about power isa dropped, is because voidlinux-ppc creator and maintainer dropped it and created his own chimera os: A modern, general-purpose non-GNU Linux distribution. It uses llvm clang, uses apkv3 from alpine, does a lot of things right, but I just can’t get excited about it. IDK what’s the status on packages support, but I doubt it’s any better than void or nix (alpine in general seems to be lacking on that side even with the community repo, I can’t imagine a lesser known distro would fare any better, particularly since some programs build templates need to be adapter for ppc).

I still hope the guys at libre-soc manage to get something workable, but I’ve been focusing on really low-power stuff, like ARM and which will eventually lead to me hopping over to RISC-V when it gets decently powerful and gathers more software support.

Unfortunately, for whatever reason, power stuff is very pricey to get into (look at raptor computers). Or if you want to dip your toes, try finding old macs (of which perf. and perf/watt will be atrocious).

And with what IBM did to Red Hat (not that RH was a really trustworthy company either, but seemed to get worse under IBM), I don’t think I trust big blue to keep the ISA open. Once there’s more adoption, they’ll quickly switch around. That is also potentially true of RISC-V, but it’d be way harder for them to pull it off (because their entire support was built around openness, while IBM’s power didn’t start out as open and was more of a “this thing is backed by a big corporation that could probably never go under”).

891

The open ISA to me is a very minor thing; the openness advantage of POWER chips is the lack of proprietary blobs needed to operate them. RISC-V chips can (and some do) require proprietary blobs of code to turn on. Besides, the ISA for Power has been (as I understand) the property of the Linux Foundation since OpenPOWER Foundation became its subsidiary in 2019.

As for IBM’s trustworthiness, yes it can and has wavered on the chip openness front (POWER10 for example has required some proprietary blobs for its memory controllers) but to my knowledge POWER8/POWER9 is still best situation for an open, serious CPU since OpenSPARC and its UltraSPARC T1 and T2.

Apparently though, we will be soon seeing non-IBM alternatives,

Also, while much less powerful and (I think) requiring some sort of blobs its NXP-made CPU, the PowerPC Notebook project is still slowly and steadily making progress,
https://www.powerpc-notebook.org/en/
Note that its CPU lacks AltiVec (PowerPC SIMD, akin to SSE or AVX on x86), so some software may be unexpectedly slower if it depends on SIMD.