Return to Level1Techs.com

The POWER and PowerPC General Discussion / News Thread

news
mega_thread
architecture
#221

If I bought this, I would use it as daily driver for productivity work plus some vintage gaming through emulation. So what would be very interesting on top of the everyday use element, is some emulation related tests. I would like to propose

  1. The use of qemu or better yet hqemu user mode emulation together with binfmt and chroot to see if simple x86 Linux games run well at all. This was actually demoed by Raptor on a Power8 system before in 2016 when they were still contemplating the Talos 1.

  2. If the above is workable then throw Wine on top of it and see how it does. Given the test done at Talospace on this topic, hqemu seems to emulate x86 at around 1/3 the speed? So let’s call it a 1ghz Pentium. So perhaps try games that date back to early 2000s?

  3. The above at the end is emulation so may not work well. But how about vintage gaming through open source emulators like RPCS3, Dolphin and MAME? I wonder if those will compile and work correctly. If they do then it still provides a decent source of old games for leisure.

  4. Away from emulation, I also wonder how well nVidia GPU works on this system. We know nVidia provides binary drivers for their Tesla cards to work on Power9 systems. I wonder if they would be nice enough to leave the codes in there so either GeForce or Quadro cards would work? The Tesla cards have no video output so “useless” as a display card.

  5. If #4 doesn’t work, would a Tesla card combined with an AMD GPU work, through the Looking Glass project? It’s a big ask though if you have a Tesla card lying around, it might be worth a shot?

Those are my wish list items. Really appreciate your insights here.

Thanks

Michael

#222

I have my powerbook with me everywhere but IDK that you’ll count that lol.

I’d love a talos tho. Even if I could live in @wendell 's basement for a weekend and poke it.

#223

If they were optimized/ported for Power, which they probably aren’t at the moment, RPCS3 and Dolphin could actually end up running faster (potentially), since the PS3, Wii, and GameCube all used PowerPC CPUs. There are potentially some non-standard extensions to implement; for example, the Xbox 360 had increased SIMD with VMX128. Also for the PS3’s Cell, you still have to implement the SPU cores, but again, I’d bet those are more similar to standard Power than x86-64.

An interesting thought, but probably unlikely.
So far only a GeForce 9500 GT has been tested according to the Nvidia section of the Hardware Compatibility List. “No firmware needed” probably means he/she was testing with the Nouveau driver instead.

#224

@wendell I would like to know what is the efficiency of the SMT implementation. I know that they are offering SMT4 and SMT8 solutions. I don’t know which one that you received. I know on the Intel platform, HT basically gives you 1.5 core performance per core. AMD is somewhere around that as well.

I know in the old ppc and power days, the SMT performance was closer to 2 core performance per core before SMT4+. I wonder what the performance is now. One of the reasons that power was “power” hungry was because you had to run at full tilt to get the most performance and efficiency out of the platform. If SMT functions close to the core performance of the thread count, then it is a great time to be alive and we can finally see a shift to alternative architectures becoming mainstream again.

I really plan on picking up a lite to function as a network appliance and build server.

#225

POWER9 Chips

All of Raptor’s boards, Google/Rackspace’s Zaius, and IBM’s AC922 (basically anything intended exclusively for Linux) use the SMT4 versions of POWER9.

The third post in this thread actually has a fairly comprehensive overview, with the same sort of chart that you see on the RCS Wiki, and on Wikipedia’s page:

PowerNV PowerVM
24 × SMT4 Core 12 × SMT8 Core
Scale Out Nimbus codename unknown
Scale Up Cumulus

POWER9 Modules

The “modules”, are more a description of the interposer/socket. AFAIK, all three of the chips described above are the same size silicon, but the module type determines what socket it will use and what IO is available:

Sforza Monza LaGrange
Size 50 mm × 50 mm 68.5 mm × 68.5 mm 68.5 mm × 68.5 mm
Socket LGA 2601 LGA 3899 LGA 3899
Memory 4 × DDR4 8 × DDR4 8 × DDR4
PCIe 4 48 lanes 34 lanes 42 lanes
OpenCAPI None 48 lanes 16 lanes
XBus 4B 1 1 2

XBus is the communication between the two sockets

#226

I’m glad people have minimal interest in this thread again lol.

#227

Looks like I too had posted about this in the thread earlier.
I should note that the RCS Wiki isn’t quite clear on module types being shared across chip types; but I asked Adi Gangidi about it on Twitter. I did link to that tweet earlier, but it is pretty interesting, so I’ll quote it here:

Let me comment on La Grange Power9 . ( Scale-out Datacenter caterer SKU ) . Yes the SMT8 and SMT4 LaGrange chips can physically fit in same socket . However major difference is in management firmware being qualified on systems built with each chip.

SMT8 chip systems qual’ed with FSP for management , SMT4 chip systems qual’ed w/ BMC for management . As far as I know nothing HW wise prevents for adding support for SMT8 chips with BMC but is not done as of today

Vice-versa ( SMT4 chips being qual’ed with FSP/ PowerVM) won’t happen for licensing reasons . Different structure to use 24 core machine for PowerVM stack. I also suspect lots of cores / lesser SMT is not well suited for typical applications that run on that stack .

#228

@Aremis Saw this on Twitter and thought of you; someone is testing Void Linux on their Talos II:

#229

Well theres been work to port to powerpc in my circle. Shouldn’t be hard for him.

#230

Oh damn, I nearly forgot about that! Looks like he’s doing ppc64 big endian too. I should probably get that pull request done soon