F yeah brother. I am more than keen if and when something of this calibre comes down to reality where most people live – it needs to happen at some point.
Well, given the first generation, publically available hardware is within the reach of the high-end enthusiast rather than just the enterprise I’d expect hardware within the the reach of the HEDT market within the next few years.
Who knows, maybe there’ll be mainstream Power processors within the next decade.
My guess is many people don’t want to get their hopes up.
Since there is a Talos II, even the after Talos I crowdfunding was unsuccessful, I’m not too worried about Raptor’s dedication to the project; I am more concerned about it bankrupting them if there are insufficient sales.
My bet is that this is why there are two separate companies; Raptor Computing Systems is the seller, to isolate sales from harming Raptor Engineering in a worst case scenario.
I really hope so, but true mainstream will probably be chromebooks/tablets/phones only in a few years. Maybe workstation/gaming though; Unreal Engine 4 already runs on OpenPOWER:
|AmigaONE X5000||Cyrus||e5500-core*||Power v2.06**|
|AmigaONE X1000||Nemo||PA6T-1682M||Power v2.04|
*no specific part listed
**incomplete support, AltiVec is missing for example
See, I know next to nothing about the architecture, so I’m trying to learn before I start contributing to the conversation, considering I try not to speak about things of which I’m uninformed.
Expression is not without appreciation.
I know the feeling, I made this account specifically to braindump what I’ve been able to figure out so far. If you find something interesting or informing just be sure to add it to the
borg collective I mean thread.
The most helpful thing I’ve read/watched is the POWER9 presentation in January 2017 by Jeff Stuecheli given to the AIX Virtual User Group, the video is available for download (something went wrong with their YouTube uploads), and the slides are available as a PDF. If you want to skim it, I have timemarkers in a post from the previous thread.
The image about POWER9 Processor Family that I posted earlier is a slide from that presentation.
Something I noticed when trying to export the slide as an image, the PDF is horribly screwed up in places; some of the images (the Power Systems logo for example) were exported as horizontal strips. So instead of one image, the PDF stores 100+ images that are 1px tall. If you are wondering what I mean, try opening up PDF in Inkscape and you’ll see.
That reminds me, I timemarked another presentation, unlike the Jeff Stuecheli presentation, this one is more about the IBM product than the processor itself. I would highly recommend skimming to what looks interesting to you, if anything.
First Power 9 System and PowerAI
The slides and video download links are on the AIX VUG wiki, if you are reading this in the future, you’ll have to scroll down or
F to it; its full heading is December 12, 2017 - First Power 9 System and PowerAI - with Chris Mann, even though the first two slide decks are by Joel Dodd.
Disclaimers: Slide 19 is shown while they continue talking about CORAL contracts, the time shown is for the start of 6-GPU design discussion. “Nutanix” is a guess, its a bit hard to hear them. “Dodd” as Joel’s last name is also a guess.
00:00:00 setting up webinar
00:00:47 Intro - Joe Armstrong
00:04:06 PowerAI presentation - Joel Dodd
00:04:20 (slide 1-2)
00:04:56 (slide 1-3) IBM AI framework packaging
00:06:44 (slide 1-4)
00:07:41 (slide 1-5) car analogy
00:08:53 (slide 1-6) PowerAI Vision
00:10:19 (slide 1-7)
00:12:17 (slide 1-8)
00:14:35 (slide 1-9) Distributed deep learning
00:15:56 (slide 1-10) Large model support
00:17:19 (slide 1-11)
00:18:01 (slide 2-2) Lab services - cognitive workshops
00:19:40 (slide 2-6) Lab services
00:20:20 intermission - Joe Armstrong
00:21:05 Power Systems AC922 - Chris Mann
00:21:36 (slide 1)
00:22:02 (slide 2) IBM strategy
00:23:20 (slide 3) OpenPOWER HPC family
00:25:10 (slide 4) PowerAccel
00:28:52 (slide 5) AC922 overview
00:30:18 (slide 7) POWER9 processor
00:31:37 (slide 8) AC922 4-GPU design
00:33:55 (slide 9) Volta specs
00:35:10 (slide 10) NVLink changes
00:38:23 (slide 11) GPU bandwidth comparison
00:40:07 (slide 12) I/O attach evolution in POWER
00:43:12 (slide 13) IB-EDR PCIe Gen 3 vs Gen 4
00:44:20 (slide 14) Front + rear views
00:45:16 (slide 15) AC input (Rong Feng 203P-HP)
00:46:55 (slide 16) Memory options
00:48:06 (slide 17) CORAL
00:49:53 (slide 18) delivery/contract discussion
00:53:16 (slide 19) AC922 6-GPU design
00:55:26 (slide 20) CORAL install at LLNL
00:57:08 (slide 21) CORAL install at ORNL
00:59:05 (slide 22) closing
00:59:11 Questions intro - Joe Armstrong
00:59:27 Q: OS for AC922
01:00:25 Q: Fan loss tolerance
01:00:59 Q: Use cases for CORAL labs
01:02:33 Q: Will it run Crysis 3?
01:03:08 Q: DCM (Dual Chip Module) or SCM (Single Chip Module)?
01:03:25 Q: CAPI 2 vs OpenCAPI?
01:05:35 Q: Mixing DIMM sizes?
01:06:28 Crysis 3 explained
01:06:43 Q: AC not LC, will it run AIX?
01:07:55 Q: Mellanox adapters for storage?
01:10:34 Q: Manufacture location?
01:10:55 Q: AIX general questions
01:11:31 Q: NVLink configurations elaborate?
01:12:31 Q: AC922 model numbers (8335GTG as public model)
01:13:55 Q: VM is not PowerVM? KVM instead.
01:14:58 Q: Leak detection? No.
01:15:28 Q: Fans in water cooled systems?
01:16:00 Q: Hardware clustering
01:17:32 Q: Clock speed?
01:19:06 Q: Different model numbers for air or water cooled?
01:19:40 Q: PowerAI on AC922?
01:20:07 ESP version of PowerAI for AC922
01:20:35 Q: Does AC922 run in the Nutanix cluster?
01:21:01 Q: Water cooling in normal datacentre?
01:22:21 closing - Joe Armstrong
01:22:38 closing - Chris Mann
01:23:07 closing - Joel
01:23:24 closing - Joe Armstrong
Summit/Sierra supercomputer delivery completion is expected in June 2018. 1000 nodes delivered so far to each. IBM is shipping 100 nodes per day.
Aurora supercomputer using x86 was was delayed.
Crysis 3 is not a POWER9 benchmark
Does anyone have access to hardware to test meltdown/spectre against?
Probably everything I own.
I was talking specifically PPC hardware.
give me a minute. gears are clunking.
Ah, got it.
I have a few bits and bobs I could test, but I think only G5 stuff will be testable for me. I can say that the new freescale chips coming out soon will be patched, as well as, well at least probably, the SAMS440 and 460EX chips.
Patched or not vulnerable?
Patched could mean they’re leaving performance on the table. (and driving down perf/watt)
Patched in this case would mean a change in chip design, which more or less in the SAMS chips is nothing new since they change every few years (where AMD chips are relatively similar for like 6 years straight, SAMS has a redesign completely when they want to change the scale of the platform). I imagine they would be vulnerable at some time seeing as how 2003-2007 PPC was around in laptops, used by a fuck load of people until 2013, and is still in HPC and server markets all over the place.
I mean we can check, I just think Spectre and Meltdown are pretty common bugs seeing as how those attacks are at an angle not many people thought about.
That’s my thought. It’s an interesting situation to be sure. It is what it is, huh.
Same with Rhino in the 90’s. MMX was new! But chips could be fucked anyways! They never really fixed it, they just added Exec Disable bit to hope it’d work. It didn’t, and theres still Atoms that are based on Pentium M’s today that are vulnerable.
Study proc architectures for a few years its fun.