Messages

256

GPU Compute / Accelerators / Radeon Pro vs regular Radeon cards - worth the price premium for POWER users?

« on: June 05, 2020, 01:05:50 pm »

The RX 5700 came out last year and it was followed 6 months later by the Pro version, specifically the Radeon Pro W5700.

The W5700 is basically double the price of the RX 5700.  From the perspective of a POWER user, is this worthwhile?

Summarizing some of the key differences in the Pro version:

- AMD is testing the hardware and drivers more thoroughly: but do they test on any POWER9 systems?

- AMD is releasing driver updates for the Pro cards on a regular schedule: do these bug fixes appear in the amdgpu release for Linux users just as quickly?

- the marketing material describes various features, such as the AMD Remote Workstation (use your GPU remotely from a laptop) but is that relevant for a Linux user?  The software they offer is proprietary, so there are a large percentage of people in this space who would not use it anyway and we also have free software alternatives

- the last significant benefit I could see: the overall design is less aggressive, slightly less power consumption and lower clockrates than other cards so even ignoring the questions about drivers, maybe it will last longer and be more stable

- some people justified the purchase of Radeon Pro products when they included ECC RAM but in the W5700, it is not ECC, it is the same as the RX 5700

I've got an open mind about this: for example, an OEM built RX 5700 that has liquid cooling and isn't overclocked may be more relevant to some people than the W5700.  But if AMD is regularly testing amdgpu with W5700 on POWER9 then that alone would make me feel they are investing in this architecture.

257

Applications and Porting / Re: VP9 benchmarks: have they improved?

« on: June 05, 2020, 12:18:35 pm »

This blog post gives some detailed discussion about VPX / VP9 / VP8 as it relates to POWER9 and other platforms.

258

Applications and Porting / VP9 benchmarks: have they improved?

« on: June 05, 2020, 12:12:38 pm »

One of the top search results for VP9 benchmarks is this OpenBenchmarking.org site where POWER9 is the slowest with 9.37 frames per second (less than real time).  Even the Intel i3 achieves 28.38 fps (better than real time).

Has this code been fixed and does anybody know how to get fresh results in that site?  I think it is very unfair to the platform when search engines show something like that if it is no longer valid.

On the other hand, if it is not fixed, has anybody proposed a bounty for working on it?  IBM is asking people to suggest issues that they will fund

259

Talos II / CPU hotplug / completely power down one CPU in a multi-CPU system

« on: June 05, 2020, 05:42:47 am »

Linux provides a mechanism to disable individual cores.  This can be useful to reduce peak power consumption or to simulate a smaller environment, for example, if a developer with a Talos II wants to know how their application would perform on a Blackbird with a 4-core CPU, they can turn off all but 4 cores.

echo 0 | sudo tee /sys/devices/system/cpu/cpu1/online

Spoiler: If you put all the cores of one CPU offline with that command then you won't be able to access the RAM and PCI slots connected to that CPU and you might observe strange behaviour.

Is it possible to go one step further and completely power down a CPU socket and maybe the associated RAM banks too, almost as if they were removed from the board?

There is some documentation about Linux kernel hotplug and it suggests x86 only.  Maybe this would be good for another bounty but first it is important to understand whether the Raptor and POWER9 hardware supports this and whether it would lead to energy savings or other benefits.

Problems that would be solved with this:

- reducing heat output from Talos II workstations during summer heatwaves

- extending runtime for a system on UPS batteries

260

Applications and Porting / Re: gstreamer performance, lessons from ffmpeg efforts?

« on: June 05, 2020, 02:10:26 am »

For anything like the problem you describe, the best thing to do is to open a bug report and give them a sample input file, output file and command line

It is also important to give them output of these commands:

ffmpeg -v

ffmpeg -buildconf

This particular thread was about performance but obviously correct functionality is even more important than speed.

261

General CPU Discussion / Re: equivalent to a HP Z cooler?

« on: June 05, 2020, 02:06:49 am »

The HP Z cooler is not simply fanless, it is based on 3D vapor chamber phase-change cooling with staggered hex-fin heat exchangers

HP published some research about this technology

262

General OpenPOWER Discussion / Re: Price getting higher and higher

« on: June 04, 2020, 12:50:49 pm »

Did you see the bounties?

https://www.bountysource.com/teams/ibm/bounties

263

Applications and Porting / Re: gstreamer performance, lessons from ffmpeg efforts?

« on: June 04, 2020, 03:44:50 am »

Mozilla's bug tracker lists a range of PPC related issues, including problems with the WebRTC stack and Youtube:
https://bugzilla.mozilla.org/buglist.cgi?quicksearch=ppc

I made a post about the bounties on the gstreamer-devel mailing list, some suggestions about future bounty projects have already appeared:
https://lists.freedesktop.org/archives/gstreamer-devel/2020-June/075144.html

264

Blackbird / Re: power consumption figures?

« on: June 03, 2020, 08:20:16 am »

I feel it might be a good idea to split the tables in both the Talos II and Blackbird pages.  One table would only have wall socket measurements and the other table would have per-component measurements.

265

Applications and Porting / gstreamer performance, lessons from ffmpeg efforts?

« on: June 02, 2020, 03:01:38 pm »

A bounty recently provided fixes for ffmpeg

Are similar fixes needed for gstreamer or any of its dependencies?

Could lessons from the ffmpeg porting be applied to any other similar multimedia frameworks, for example, gstreamer, the WebRTC media stacks in Firefox and Chrome?

266

Blackbird / Re: power consumption figures?

« on: June 02, 2020, 02:48:26 pm »

On the fourth row, for Blackbird bundle, do those figures include the CPU power usage or it is necessary to add the CPU figure from above?

The paragraph above the table suggests that power was measured at the wall, that implies the bundle figures include the CPU power.

It also seems very odd that the top row, a 4-core CPU running at 2.15GHz is using more power (54W) than the 8-core CPU at 3.8GHz (34W).  Is the figure for a 4-core system based on the complete bundle, motherboard, RAM, etc?

267

General CPU Discussion / equivalent to a HP Z cooler?

« on: June 01, 2020, 04:29:53 pm »

Is there any possibility of somebody marketing a fanless POWER9 cooler similar to the HP Z coolers?

Maybe this would not be feasible for the 22 core chips but it would be very interesting for 4-core.  Not only is it silent but it also eliminates one possible type of mechanical failure.

268

Blackbird / power consumption figures?

« on: June 01, 2020, 04:25:21 pm »

The wiki has some rough power figures for Talos II

Could anybody provide similar figures for basic Blackbird builds, with bare minimum spec, no GPU?

E.g.
Blackbird + 32GB + basic SSD + CPU (4 core v2 idle) = ? Watts
Blackbird + 32GB + basic SSD + CPU (4 core v2 100% utilization) = ? Watts
Blackbird + 32GB + basic SSD + CPU (8 core v2 idle) = ? Watts
Blackbird + 32GB + basic SSD + CPU (8 core v2 100% utilization) = ? Watts

This type of detail is really useful for anybody trying to estimate PSU requirements, runtime on a UPS and heat output of the box

In comparison, HP provides a couple of example builds for each of their Z workstations in quickspecs documents like this (see page 26).  Most people can extrapolate from one of the examples to their own use case.

269

General Discussion / LSI HBAs on ordering page - blobs?

« on: June 01, 2020, 04:03:11 pm »

The Raptor ordering page includes some LSI HBAs for SATA / SAS:

• LSI 9300-8i 8-port Internal SAS 3.0 HBA
• LSI 9300-8e 8-port External SAS 3.0 HBA
• MegaRAID 9361 8-port Internal SAS HBA w/ BBU
• MegaRAID 9361 16-port Internal SAS HBA w/ BBU

The non-RAID cards have a PowerPC 476 running firmware from LSI.  The RAID cards have some kind of RAID SoC, I haven't looked at the details but it also includes firmware.

Is there any particular reason these are recommended in the ordering process or is source code available now?

Do these cards allow people to replace the firmware or do they expect cryptographically signed blobs from the vendor?

270

Talos II / OCuLink ambiguity

« on: June 01, 2020, 02:05:08 am »

When I look at the ordering pages on the Raptor site, it offers various OCuLink products:

• 4x OCuLink PCIe Expander (Gen 3)
• OCuLink to U.2 Cable

The manual:
https://wiki.raptorcs.com/w/images/e/e3/T2P9D01_users_guide_version_1_0.pdf

includes errata stating that

ER0001
μPCIe connector (J10108) is incorrectly identified as OCuLink (SFF-8621) due to errors that were discovered post-production in IBM OpenPOWER reference designs.  While the physical connector is the same as OCuLink ports, the wiring is different, notably in the presence detect / clocking pins.  As such, while an OCuLink peripheral is unlikely to be damaged by being plugged in to the μPCIe port, it will not function while attached.


If that is the case, buyers selecting the OCuLink products will not be able to use them - or has the board design been changed somehow to make OCuLink work?