Intel expands Xeon Phi co-processor lineup

At International Supercomputing Conference at Leipzig, Germany, Intel officially introduced Xeon Phi 3100 and 7100 series co-processors, and revealed details of future "Knights Landing" products. Intel Xeon Phi family was announced last November, and so far Intel released three different models from 5100 series, with only 5110P SKU available in retail. New Xeon Phi products, launched this week, include 7120P and 7120X high-performance parts, 5120D for dense servers, along with 3120A and 3120P "value" co-processors.

Xeon Phi 7120P and 7120X are implemented as PCI-Express cards, housing the co-processor chip and on-board GDDR5 memory. The 7120P also includes a passive thermal solution (heatsink). Both parts have 61 cores, operating at 1.238 GHz. Xeon Phi 7100 series co-processors support Turbo Boost feature, that can increase frequency up to 1.33 GHz when needed. The cards have on-chip 30.5 MB L2 cache, and there is also 16 GB of GDDR5 memory soldered on the board. The peak performance of these co-processors exceeds 1.2 TFLOPS. The 7120P and 7120X are rated at 300 Watt TDP, and priced at $4129.

Xeon Phi 3120A and 3120P are positioned as inexpensive co-processors. They incorporate only 6 GB of on-board memory, therefore they are better suited for CPU bound tasks. The 3120A and 3120P have 57 cores and 28.5 MB L2 cache. The co-processors operate at 1.1 GHz, which results in peak performance of 1003 GFLOPS. Thermal Design Power of 3100 series cards is 300 Watt. The 3120A has an active cooling solution, while the 3120P has a passive one, and that is the only difference between them. The official price of these parts is $1695.

Xeon Phi 5120D is produced in a different form factor than 3100 and 7100 series cards. This part is shipped as a PCB with installed co-processor and 8 GB of on-board GDDR5 memory. The PCB has a 230-pin edge connector, that fits into x24 PCI Express slot. The 5120D has 60 cores, and 30 MB of L2 cache. The part runs at 1.053 GHz, does not support Turbo Boost technology, and does not come with a thermal solution. The TDP of this co-processor is 245 Watt, and the peak performance is 1011 GFLOPS:

ModelCoresThreadsFrequency /
Turbo
L2 cacheOn-board
Memory
TDPPrice
Xeon Phi 3120A 57 228 1.1 GHz 28.5 MB 6 GB 300W $1695
Xeon Phi 3120P 57 228 1.1 GHz 28.5 MB 6 GB 300W $1695
Xeon Phi 5120D 60 240 1.05 GHz 30 MB 8 GB 245W $2759
Xeon Phi 7120P 61 244 1.23 / 1.33 GHz 30.5 MB 16 GB 300W $4129
Xeon Phi 7120X 61 244 1.23 / 1.33 GHz 30.5 MB 16 GB 300W $4129

Additionally, Intel disclosed some details on "Knights Landing" architecture, that will be used for the next generation of Xeon Phi offerings. "Knights Landing" products will be fabricated on 14mn process, and will be shipped as co-processors, as well as processors. Future co-processors will be similar to "Knights Corner" parts, that is they will be available as PCI-E cards. On the contrary, "Knights Landing" processors will fit into a motherboard socket, and will work as a standalone CPU and a co-processor at the same time. The processors will also have additional on-package memory, that will considerably increase their memory bandwidth.

Related News (newer articles):

Related News (older articles):

Comments

There are 8 comments posted

"8 MB of on-board GDDR5 memory"?!

2013-06-20 05:56:28
Posted by: cpu0620

What's the use of such a small amount of memory?

 

2013-06-20 17:36:58
Posted by: gshv

It was a mistype. I corrected it. Thank you!

Really?

2013-07-26 21:36:35
Posted by: Robert Edison

Over $4k for only 1.2 teraflops? This is a joke? A misprint? It there something I'm not understanding? If I'm not mistaken, typical $4k workstations these days are between 4 and 6 teraflops already. Why would anyone pay $4k for for just 20 to 25% more performance. And I would assume that performance would be limited in scope even at that?

I'm not sure I understand fully but from what I gather these Phi cards appear to the host operating system over a virtual ethernet connection as a connected "system". I assume that's going to look very much like any other desktop or workstation "system" connected to the host over ethernet. So we're limited to network based controllers like render managers or maybe video application's network assist engines. I mean to say the coprocessor boards won't be available in the native compute stack - applications not designed for network assisted computing or which aren't specifically written to utilize the Phi will not benefit from it's presence.

So it 20 to 25% speed-up only sometimes for about $4,000. Why not just connect up another $4k workstation and get about four times the horsepower in the same way?

Please correct me if this or any of the above is mistaken or in error.

Is it Power Savings?

2013-07-26 22:04:35
Posted by: Robert Edison

In further consideration it doesn't seem to pan out as a power saving device either. Let's assume the 7120 consumes only 275W at full tilt and not actually a full 300W. So a system with four times the horse power (like a 12-core E5 workstation) should generally on average, consume 1100W when he CPUs are above 90% and something like an installed GTX 780 is at around 80% to 90% as well (giving us that 4 to 6 teraflops). But it's not. Typically such a system working that hard consumes somewhere between 600W and 800W.

So assuming I don't have something backwards here the Phi 7120 is electrically a lot less efficient per FLOP than a typical workstation computer system. Wait, what?

I'm confused.

 

2013-10-02 22:55:05
Posted by: Lee J. O'Riordan

That quoted figure for GTX 780 is single precision FLOPS, not double precision, as given by the Phi. Also, the double precision performance of the GTX 780 is quoted as being 1/24 that of the GTX Titan, which has a DP FLOPS value of about 1.5 Teraflops.

there is a lot more you need to know

2013-09-24 09:09:07
Posted by: Number Cruncher

as the subject of my post says, there is a lot more you need to know both hardware architecture and software programming.

you can use 6TFlop just for a game, or you can use that RAW power for processing information. my new PC has a core i7 4770 it can do max 50GFlops for 500$. I'll never consider paying that much of money to xeon phi for home use but wait, if you're building a modern number cruncher, you have to pay that amount, even more.

Animation

2014-01-19 02:43:27
Posted by: Adrian Pérez

I have my Mac Pro 8 core and i can go up to more than 50 core, hig performance can be operated, render engine can be used in less time?

Cost/Performance ???

2014-03-27 12:01:02
Posted by: Steve

For that kind of money one can buy whole new high performance workstation with 2 physical CPU slots ! I think Tesla Cards are more cost effective but also overpriced ! Imagine in theory you could gain more than 100% CPU power just with one card or even more. Every appropriate mobo could do the job and have home supercomputer. Personally for me 100% CPU gain what so ever card needed would be priced for the price cpu and mobo together so cpu 250 eu and 250 eu would be caa 500 eu for the cost of such card. If price would be so low imagine how many people would buy it !

Terms and Conditions · Privacy Policy · Contact Us (c) Copyright 2003 - 2010 Gennadiy Shvets