I mean I’m not denying that, OEM laptops exist under that principle, I’m saying that if this comes out I’ll be the first to go in debt. If it doesn’t, then I’ll stick with PPC macs (which apple released to the open front and theres now EFI images for) and just enjoy what I have already.
Different for the sake of different is fine, but that doesn’t sell systems.
OpenPOWER is going to have to provide significantly better performance and capabilities at a lower price just to get people to even consider spending money on it. And frankly that’s going to be nigh impossible. And when you factor in the inconvenience of having to use different software, learning, maintenance, and peripheral hardware, the deal gets progressively worse.
x86 is the most open system. As much as I’d like to see Intel dethroned, I don’t see it happening until all sorts of factors change to favor a competitor. Intel wasted more money on the iTanic boondoggle than what AMD is worth. Intel couldn’t even dethrone their own technology.
Well, for the most part, the open systems are the old ones. You know, the HP/Compaq NW8000’s of the world where hacking the BIOS and replacing it with Openboot isn’t even illegal anymore because the OEM’s don’t give a shit. Its all very… Smooth. However you have to be willing to learn how the old systems work. Quite often its something like a core duo macbook 1,1 but everyone just ree’s and you get the lounge having a hissy fit about you not buying ryzen to do email.
The problem is less about adoption. Amiga’s are still around and getting new users and a good amiga setup nowadays is 4000 bucks. It’ll have fans, that isn’t a problem. Hell I know the university I went to will buy at least one and replace a lot of servers with it that just take up space.
Well Intel and AMD… Ultimately they are binary twins one cant exist without the other not only due to the sherman anti trust act but because they do provide half the instrunction set of each others processors to each other. AMD providing the 64bit extensions and intel providing the IA-32 bit instructions from way back on the 386 (yes they have improved and changed since then but this is essentially what makes up x86-64)
but yeah as much as id like to see them dethroned… their system works… it would take a very long time to switch over and the benefits would need to be significant
You know I’ve had an interesting thought brewing for a while… We have all these Intel chips that are holy shit balls amazing but they still run windows like trash… Why?
You know what? Blog thread. Intel is getting dethroned by AMD right now, but I have other questions on top of all of it.
I finished time marking a POWER9 chip webinar from January if anyone is interested, it has a fair amount of detailed information:
The YouTube video appears to be down at the moment, apparently unintentionally, but the slides and video can be downloaded at the AIX Virtual User Group’s webpage, the slides can be found under the heading “January 26, 2017 - POWER9 - Jeff Stuecheli”, while the video files are listed by a bare index page at public.dhe.ibm.com.
Time markers
0:00 intro/VUG announcements - Jill Armstrong
2:59 presentation start - Jeff Stuecheli (chip development)
3:54 POWER roadmap
7:22 POWER9 markets
11:15 core design - SMT8 vs SMT4
14:20 core execution slice microarchitecture
17:43 pipeline improvements
19:30 core diagram slide - SMT4 core
21:00 memory attachment - SO vs SU
25:08 common features
32:12 SO/SU SMT4/SMT8 matrix
34:15 performance vs POWER8
35:53 POWER ISA 3 instructions
38:42 POWER 9 logical view - interconnects/throughput
39:57 25 Gb/s signaling for SMP, Open CAPI, NVLink 2
41:10 16 socket topology - probably called 980
42:23 interrupt design
44:18 accelerators
47:09 accelerator connection types
49:38 progression - PCIe3 → PCIe4 → NVLink → Open CAPI
50:42 POWER9 ecosystem
51:30 Open CAPI
55:01 members
56:12 Open CAPI 3 features
57:24 virtual addressing
57:55 attached memory et al.
59:21 end of slides
1:00:01 Q&A start
1:00:39 Q: performance vs POWER8 - graph same for SMT4 compared as for SMT8 compared?
1:02:34 Q: can different chip types be mixed
1:03:26 Q: chip speeds
1:04:33 Q: ISA affecting AIX version compatibility
1:05:51 Q: live partition mobility
1:06:36 Q: number of 25 Gb/s links
1:08:10 Q: SMT4 vs SMT8, Linux vs AIX
1:10:29 Q: POWER9 vs x86
1:12:39 Q: DW on chart - looking for slide
1:14:04 – relevant slide for above found
1:14:36 Q: are charts/slides confidential
1:15:17 announce NDA session in Orlando
1:16:45 Q: Java improvements
1:17:43 Q: Open CAPI attached memory uses
1:19:20 Q: PCIe4 adapters/uses
1:21:34 Q: accelerator effect on rPerf
1:23:17 Q: will it run Windows
1:23:36 Q: SAP HANA offerings, upgrade path
1:24:03 Q: dynamic SMT
1:24:56 threading changes from POWER8
1:25:45 end remarks
Interesting Notes
Intended Use of Chips (slide 10)
| SMT4 | SMT8 | |
|---|---|---|
| SO | OpenPOWER | dual-socket PowerVM |
| SU | multi-socket PowerVM |
The SMT4 column is labelled “Linux ecosystem”, which is what OpenPOWER is about; saying Linux is actually a bit confusing since you can run Linux under PowerVM.
I wonder why PowerVM benefits from SMT8 where Linux does not?
25 Gb/s optical connexion (slide 14)
in addition to PCIe4 (which carries CAPI 2) there is also optical 25 Gb/s connection used by Open CAPI, NVLink 2, and SMP interconnect; cool.
LOL whut
Yeah… especially seeing as this was a webinar with the AIX users group.
Still, they did answer the question so I marked it down as such.
1 Like
I never got the impression that the Xeon Phi coprocessor sold particularly well, and it appears that Intel would agree.
OpenPOWER is a general-purpose CPU, and I think the ability to run a normal OS and applications on it is a point in its favor. There would have to be a compelling use case to justify having to learn specialized tools to develop for it, like you do with Xeon Phi add-in cards, GPUs, and FPGAs. As a standalone system, the only requirement is a compiler target.
Well it can pipe a fuckload more data than the current 4 core xeons. Plus you can get 4 core 32 thread P9’s because fuck it why not?
POWER is cool, but people have to work with it. I’m still getting one of these if for nothing else as a stupid fucking amiga.
STOP GETTING MY HYPED FOR THE TALOS PRINCIPLE 2.
I just want to play it.
GRRRR!
Me slappy, you crappy.
I had not noticed that they finished the FAQ page, but what is interesting, is that the “Why Raptor Computing Systems? …” question mentions:
… we are not just shipping a stock reference design, like other vendors did with POWER8; rather, Talos™ II contains numerous unique features that increase usability, promote openness, and boost system security. Talos™ II is truly one of a kind and is additionally protected against unauthorized hardware clones by patents and/or patents pending …
which makes me wonder what these non-reference features are.
The only other POWER9 board I know about is the Zaius mainboard that Google and Rackspace designed.(google blog, rackspace blog) Keep in mind though, it is using LaGrange CPUs, not the Sforza chips that Talos II will.
wasn’t this game just a bone stock survival puzzler type thing that everyone just bought as a low level API benchmark back when pickings were slim?
ALSO: Anyone got SQL performance or video encoding benchmarks for this POWER9 cpu?
ALSO ALSO: What’s the point of nvlink I/O if there’s no POWER9 nvidia driver
SQL and encoding performance should be insane on these. Also, I believe the NV drivers are available on these sorts of systems as they tend to have the compute and server GPU’s in the rack setups.
I wanted benchmarks, not speculation from our resident RISC fetishist. I already know what your opinion of these CPUs is.
Also, there’s definitely no public POWER unix drivers for their nvlink devices. I’m in the cuda/cudnn developer program and there’s no easy access there either. Chances are average Joes will never be able to use that feature of this Mobo.
That’s literally the only thing that makes this box interesting to me, as it might’ve been a cheaper way to accelerate ML work and development. If it can’t encode or do db faster than the equivalent 5k USD x86 server, then it’s pretty useless tbh.
Well this is also like an 18000 dollar machine. This isn’t exactly Greasemonkey Bob’s home computer.
it’s 5-7k, not 18k.
Also, it’s a general purpose workstation machine. If it performs worse at everything than other computers at the same price, and all of it’s keystone features have zero real world applications, It doesn’t matter what you buy it for. It’ll be a very bad value.
Even if I was an enthusiast or hobbyist, I’d want to see benchmarks for my application before even considering one of these.
My point is that people are paying out the ass for the tech. at hand, I assume you get a lisence for some cool shit in there as well.
That point makes very little sense.
Even if it did, that tech can’t be used for anything, and accelerated computing is off the table entirely because nothing that pushes decent OCL/Cuda performance is supported on POWER by their manufacturers.
Best case it’s better at like, one thing that’s CPU-bound that isn’t impossible to compile for on the platform.