Intel Xeon PHI?

[Xyzzy]

On Windows, there is _OX in the upper right of each GUI app window (minimize, maximize, terminate respectively). On Gnome _ (minimize) is missing.

The "gnome-tweaks" package will fix this.

[ewmayer]

Originally Posted by Xyzzy

The "gnome-tweaks" package will fix this.

Lack of keyboard-based minimize on my Ubuntu desktop (my main Linux work-managment machine - all others are remote-access) also bugged me - at least under Ubuntu/Gnome, "option-h" (where 'option' is the key between Ctrl and Alt at lower left - on my USB keyboard it has no lettering, just a Windows logo) works. Another hugely useful key combo, which appears to work under both Gnome and MacOS, is Alt/[Apple logo]-Tab, which allows cycling among open applications - hold down Alt with left thumb and repeatedly hit Tab with left index finger to cycle, can also hold Alt and use right/left-arrow keys to go in either direction. Under Gnome (but not MacOS) it will also bring forward applications which are currently minimized in the sidebar.

The one downside of Alt-Tab is that Tab is right next to Q, very easy to inadvertly hit Alt-Q and quit the currently topmost application instead of cycling to a different one.

[kachianind]

Hi folks,

Hope it is ok for me to post here like this.

My company used to sell a lot of surplus Xeon Phi hardware, and it is likely that a lot of the systems you all are using passed through our hands on their way to the gray market. I'm very glad to see this hardware has found a new life.

I wanted to mention that I still have a number of 4-node KNL 7250 servers languishing in storage. They are (or were, at last inspection) in perfect working order, but most or all of them have imperfections (e.g. one of two redundant power supplies missing, missing rack ear, dent, deep scratch) that I deemed unsuitable for an unsuspecting anonymous buyer but which needn't be a deal breaker for a bargain shopper.

These are fairly bulky to ship (I'm in California), and I realize four nodes might be overkill for a hobbyist, but I'm willing to let them go at a price significantly less per node than the K1SPE boards go at. Please let me know if that's of interest.

[ewmayer]

After crunching hard for 18 months, my bargain-refurb KNL workstation went "boink" at beginning of May. Suspect the MoBo - possibly the CPU - but after a fair bit of fruitless attempted fault-tracing (many thanks to our own Paul Underwood for much guidance there) it's all become too great a distraction from my coding work, so am flogging it off for parts. First up is the 192GB ECC server-RAM bundle I bought for it.

[paulunderwood]

Originally Posted by ewmayer

After crunching hard for 18 months, my bargain-refurb KNL workstation went "boink" at beginning of May. Suspect the MoBo - possibly the CPU - but after a fair bit of fruitless attempted fault-tracing (many thanks to our own Paul Underwood for much guidance there) it's all become too great a distraction from my coding work, so am flogging it off for parts. First up is the 192GB ECC server-RAM bundle I bought for it.

I understand the distraction to your key skills of coding but there are a few more things you should try with the PSU+Motherboard+CPU:

Step 0. Make sure the pump and radiator fans are plugged into the motherboard.

Step 1. Connect you router to the IPMI socket of your motherboard with a LAN cable.

Step 2. Power on the system.

Step 3. Look at your router's DHCP settings (or assigned IP address) through its browser interface, to see what IP address has been assigned to the motherboard,

Step 4. In another browser tab enter that IP address and login in with user ADMIN, password ADMIN.

Step 5. Poke around the interface.

Step 6. See if the processor is there.

Step 7 Turn off watchdog in the BIOS inside the IPMI interface.

Step 8? Does anyone have any ideas?

[diep]

Use it as a heat toaster this winter?

[diep]

Originally Posted by ewmayer

After crunching hard for 18 months, my bargain-refurb KNL workstation went "boink" at beginning of May. Suspect the MoBo - possibly the CPU - but after a fair bit of fruitless attempted fault-tracing (many thanks to our own Paul Underwood for much guidance there) it's all become too great a distraction from my coding work, so am flogging it off for parts. First up is the 192GB ECC server-RAM bundle I bought for it.

Sounds familiar - when lots of heat wasn't removed properly - slowly it eats up copper tracks - scrapheap remains. Typically such 'eating up' occurs when a chip has been clocked bit higher by intel than they should - and usually that happens bit after a year.

Intels hybrid approach with larrabee incarnations has miserably failed regrettably and AMD's trick (though not a new idea) works really well. So the big issue some years ago putting more cores onto cpu's is that they grow too huge and some time in the past above 300 mm^2 a die really would be too huge to get good yields with it (that might be smaller die area today). Idea now with the threadripper generations is to produce bunch of small chips
and then glue them together to get on a single package what in fact is a 8 socket machine with 64 real cores and probably more in the future. A killer approach.

As each individual chip is relative small and can get good yields. You combine good chips together to get a 8 socket machine in a single package. Genius approach to keep production cost down!