Intel Xeon Phi Coprocessor 31S1P Special Price Promotion of $195*

Boddy

Active Member
Oct 25, 2014
780
145
43
i'm down for 1 if someone is willing to help me out with intl. ship

edit: even 2 if the 100 unit price is reached
Have you considered using a 3rd shipping service in the USA, like: 'Shipito' or 'MyUSA' where you can get a virtual USA address and have stuff shipped to you, often at cheaper prices than EBay.
 

Blinky 42

Active Member
Aug 6, 2015
615
230
43
46
PA, USA
i'm down for 1 if someone is willing to help me out with intl. ship

edit: even 2 if the 100 unit price is reached
Sure - Just PM me and I can help you with intl. re-shipping if needed (and we get enough folks together to place an order). Whoever places the order can probably handle it as well since they would be re-shipping the individual cards out anyway unless we can get colfax to direct-ship to multiple places.
 

Cesar L

New Member
Feb 1, 2016
2
1
3
34
Have you considered using a 3rd shipping service in the USA, like: 'Shipito' or 'MyUSA' where you can get a virtual USA address and have stuff shipped to you, often at cheaper prices than EBay.
I did not know of those services, thanks for the info
 

pc-tecky

Active Member
May 1, 2013
202
26
28
So, what's the status? Still working on details, or did it fizzle out? Wouldn't mind 1 or 2 myself.
 

Xamayon

New Member
Jan 7, 2016
25
14
3
Fizzled out or not, the Phis are very interesting. The 10x Phis for 1k that Colfax does was too good a deal to pass on when I heard about it mid last year. One interesting way I have found to house and power them is to get an old Tesla S1070, remove the ancient Teslas and put in the Phis. The S1070 is more or less just a plain old external PCI-e expander chassis with a huge power supply.

Due to some limitations which I have not bothered to track down yet, one Tesla needs to remain plugged in or the s1070 won't stay online. Still though, 3 Phis in 1U is pretty decent. The S1070 systems are pretty cheap these days too.


The S1070s fans can be a bit intense (I have never heard fans soooo loud) but that can be fixed by intercepting and replacing the PWM signal used to control them. Attached are a few pics of what it looks like.
 

Attachments

Patrick

Administrator
Staff member
Dec 21, 2010
12,341
5,469
113
@Xamayon what have you found them to be useful for?

BTW - I was looking at those chassis earlier. They look like really interesting solutions.
 

Xamayon

New Member
Jan 7, 2016
25
14
3
I've been doing some image processing experiments with them. I run a reverse image search engine, and pretty much all aspects of that are high intensity and highly parallelizable/vectorizable. Nothing phi related in production yet though, I like playing around too much... ;)

One of the really interesting things about them is that they are their own completely separate nodes in a way. They have their own OS, filesystem, and virtual network interface. If server applications are compiled natively for them, they can be bridged to the real lan and used without any host system involvement. They can also be used as offload targets, but the native execution in particular is very different than what CUDA/etc allows.
 
  • Like
Reactions: Patrick

Patrick

Administrator
Staff member
Dec 21, 2010
12,341
5,469
113
Xamayon - any chance you can share some tips on which HIC/ cables you are using?

@Blinky 42 I think they are PCIe 2.0?
 

Blinky 42

Active Member
Aug 6, 2015
615
230
43
46
PA, USA
Yes I am pretty sure the S1070 chassis is PCIe2.0 their PDF brochure for it http://www.nvidia.com/docs/IO/43395/NV_DS_Tesla_S1070_US_Jun08_NV_LR_Final.pdf has
>>Host system with PCI Express x8 or PCI Express x16
>> PCIe Gen2 for best results

But depending on how much GPU <-> Host memory transfer your workload has vs on-GPU it may not be a huge problem for the older Teslas or Phi's

The expressbox 3450 is PCIe 3 x16, but you are paying $5k+++ for it from a little google searching.
 

Xamayon

New Member
Jan 7, 2016
25
14
3
The S1070 is pci-e 2.0, which is not limiting since the Phis are also 2.0. Intel decided not to use 3.0, as it (at the time) had a higher latency than 2.0. The S10xx and S20xx seem to be the same chassis with different teslas. There are also Quadro versions. I would expect all to work more or less the same.

I have several HIC Cards, which seem to work fine interchangeably. Low profile single port x16 P797 HIC, and the Full height two port x16 P894 HIC. I typically use the two low profile ones, which gives each side of the S1070 a full x16.

As far as cables go, the P797s came with .5 meter cables of unknown part number, which are the perfect length for use in a rack, but I also have much longer cables which I use on my test bench.
The longer cables were listed as "Dell H6GPT NVIDIA 030-0238-000 External PCIe Cable Leoni Molex 111008-0113"

Here are two auctions which should work together:
S1070 and HIC: NVIDIA TESLA S1070 GPU COMPUTING SYSTEM WITH 4 X TESLA M1060
Long Cable: NEW Dell H6GPT NVIDIA 030-0238-000 External PCIe Cable Leoni Molex 111008-0113
 

brodonalds

New Member
Jan 22, 2018
9
4
3
45
Sorry to have my first post be a revive of an old thread but is anyone still interested in a bulk deal here. I want to grab some and wanted to see if this was still of any interest to anyone here. Also @Xamayon thank you for all the info about the tesla sleds. Can someone point me to the documentation about presenting the cards to an os from the tesla platform or is it all self contained in the 1u?
 

Xamayon

New Member
Jan 7, 2016
25
14
3
Sorry to have my first post be a revive of an old thread but is anyone still interested in a bulk deal here. I want to grab some and wanted to see if this was still of any interest to anyone here. Also @Xamayon thank you for all the info about the tesla sleds. Can someone point me to the documentation about presenting the cards to an os from the tesla platform or is it all self contained in the 1u?
The 1U Tesla things are basically just two PCI-e expanders, both accepting two 16X cards. There is no special Tesla specific configuration as far as interfacing with the host goes. If everything is hooked up properly, the unit turns on and the cards just show up. Unless you modify it to never turn off though, one old Tesla card will still be needed to make the unit power on.
 

Joel

Active Member
Jan 30, 2015
833
179
43
40

Robert Fontaine

Active Member
Jan 9, 2018
113
28
28
55
Closer to 300 watts per card. I'm actually just about to put 2 of mine up onto flea-bay with alphacool liquid cooling plates
 
  • Like
Reactions: Joel