xeon phi Intel begins to ship its KNL Xeon Phi processors

[NumLock21]

At ISC 2016, Intel has announced their newest KNL (Knight's Landing) Xeon Phi processors has started to ship. KNL Xeon Phi, will actually be a host processor (socket) or co-processor (PCIe), not like their older Xeon Phi products where it's only a co-processor. What that means is, the host based version, can actually be used as a stand alone CPU. There are 2 kinds, one with Omni-Path, which provides insane network bandwidth, and the other without. Currently there are 4 SKUs 7210, 7230, 7250, and 7290. 64 cores running at 1.3GHz are for the bottom 2, and other 4 cores and +100MHz are added for the 7250. A additional 4 more cores with another +100MHz, topping out at 72 cores running at 1.5GHz for the 7290.

Quote

The KNL (>8 billion transistor) 14nm processor features an innovative 72-core architecture, which is split into 36 dual-core "Tiles" that are arranged in a mesh configuration. Each quad-threaded Silvermont core features two AVX-512 VPU (Vector Processing Units) per core, for a total of 144 VPU. Each tile houses 1 MB of shared L2 cache, which equates to 36 MB of total L2 cache. The KNL processors feature up to 3x the single-threaded performance relative to the previous-generation Knights Crossing products.

Each CPU has 16GB of onboard MCDRAM, using Micron HBM memory, providing more than 400GB/s of bandwidth, except for 7120, which may run slower, since its HBM transfer rate is capped at 6.4GT/s compare to 7.2GT/s on the other 3 SKUs. On the standard memory side, they runs on 6 channel DDR4 @ 2400MHz and up to 384GB max. The 7210 is clocked slower at DDR4 2133MHz. 36 PCIe 3.0 lanes are provided by the CPU with additional 4 from the Wellsburg PCH.

For the Omni-Path fabric,

Quote

The aforementioned integrated Omni-Path fabric is connected to the die via dual PCIe x16 ports and provides a dual-ported 100 Gbps pipe for networking traffic.

All of them are built on the 14nm processor using the Silvermont architecture, which is what Intel uses to built their Atom based cpu running on mobile phones and tablets. TDP ranges from 215w to 245w for the 7290. Omni-Path version adds a additional +15w to the total TDP, so a 7290 with Omni-Path will have a TDP of 260w. It uses LGA3467 and price starts out at $2,438 for the entry level 7120 and up to $6,254 for top of the line 7290.

 

Product Name: Intel Xeon Phi

Code Name: Knights Landing

Cores: 64-72

Speed: 1.3GHz - 1.5GHz

Turbo: 1.5GHz - 1.7GHz

Cache: 32MB L2

MCDRAM: 16GB Micron HBM

Intel Xeon Phi with Omni-Path fabric

 

http://www.intel.com/content/www/us/en/processors/xeon/xeon-phi-detail.html

http://ark.intel.com/products/family/92650/Intel-Xeon-Phi-Product-Family-x200#@Server

http://www.tomshardware.com/news/intel-xeon-phi-knights-landing,32121.html

 

[TheRandomness]

72 cores and 288 threads... 4 threads per core? Cool.

[sof006]

Jesus... Thats a lot of cores. Wonder how many of them you'd need to make a 100 Petabyte machine...

[Mr.Meerkat]

7 minutes ago, TheRandomness said:

72 cores and 288 threads... 4 threads per core? Cool.

Hmmmm...a situation where 288 threads are all used...25% of the speed (per thread) of an actual core anyone? (I know that's not exactly how HT works but still )

[FloRolf]

4 minutes ago, sof006 said:

Jesus... Thats a lot of cores. Wonder how many of them you'd need to make a 100 Petabyte machine...

Wtf is a 100petabyte machine? 

[Bouzoo]

Those prices are a bit higher than I expected. Didn't know 4C/16T were worth $1000-1600. 

 

 

[zMeul]

wait ... it has 16GB or HBM on die?! holy shit 

[zMeul]

8 minutes ago, Bouzoo said:

Those prices are a bit higher than I expected. Didn't know 4C/16T were worth $1000-1600.

if I recall, nVidia was asking 12k for a P100 - Intel's offering is a shit ton more cost effective

[Bouzoo]

1 minute ago, zMeul said:

if I recall, nVidia was asking 12k for a P100 - Intel's offering is a shit ton more cost effective

That doesn't make it any better. It's just who is the lesser evil out of 2. 

[zMeul]

1 minute ago, Bouzoo said:

That doesn't make it any better. It's just who is the lesser evil out of 2. 

now's not better since Intel's offering is a CPU with a co-processor on the same package?!

P100 is just the co-processor

 

and if I'm looking the prices, you could buy 2 KNL for a P100 

[Electronics Wizardy]

11 minutes ago, Bouzoo said:

Those prices are a bit higher than I expected. Didn't know 4C/16T were worth $1000-1600. 

 

 

These are competiting with nvdiia tesla's that cost about the same.

[Bouzoo]

1 minute ago, zMeul said:

now's not better since Intel's offering is a CPU with a co-processor on the same package?!

P100 is just the co-processor

 

and if I'm looking the prices, you could buy 2 KNL for a P100 

I meant as in that doesn't make the price right for neither of them. I meant as in jump of 8C/32T for a small price of $2500. NV is just even worse.

[sof006]

29 minutes ago, FloRolf said:

Wtf is a 100petabyte machine? 

Super computer

[FloRolf]

4 minutes ago, sof006 said:

Super computer

No not really... 

[patrickjp93]

38 minutes ago, sof006 said:

Jesus... Thats a lot of cores. Wonder how many of them you'd need to make a 100 Petabyte machine...

At 6.912TFlops per chip, do the math. You can fit up to 8 in a 1U rack (4 CPUs, 4 accelerators)

 

1 minute ago, FloRolf said:

No not really... 

100 PetaFLOPS (not petabytes :P) would be the most powerful supercomputer on the planet right now.

[patrickjp93]

34 minutes ago, Bouzoo said:

Those prices are a bit higher than I expected. Didn't know 4C/16T were worth $1000-1600. 

 

 

The lowest SKU is 62 cores!

[patrickjp93]

4 minutes ago, valdyrgramr said:

I feel like they did this because of the new opteron.

No, it's been planned for a long time. And the Knight's Corner Xeon Phi have been here since 2012 iirc. The Tianhe-II, the former #1 supercomputer, was built out of the predecessor of this product.

 

And the new 32-core Opteron will be up against 24-core Broadwell Xeons and eventually 28-core Skylake Xeons which also contain HMC, HBM2, FPGA, iGPU, DSP, and other acceleration options as outlined in the Purley Platform notes outlined a year ago. The 48-core Opteron will be up against Cannonlake which will probably break 36 cores or more.

[ThinkWithPortals]

There might be a typo in the title. Should it be "Intel begins"?

[patrickjp93]

42 minutes ago, Mr.Meerkat said:

Hmmmm...a situation where 288 threads are all used...25% of the speed (per thread) of an actual core anyone? (I know that's not exactly how HT works but still )

Hey, Power 8 and Sparc 7 have had 8 threads per core since 2013. Knight's Corner had 4 threads per core since 2012.

[themaniac]

if i did the math right it should be about 1500 KNL 7290 correct?

[Bouzoo]

22 minutes ago, patrickjp93 said:

The lowest SKU is 62 cores!

I was talking about 3 more expensive ones since the cheapest one has lower bandwidth. I even stated $2500 difference for 8C/32T.

[patrickjp93]

1 minute ago, Bouzoo said:

I was talking about 3 more expensive ones since the cheapest one has lower bandwidth. I even stated $2500 difference for 8C/32T.

Oh, lost in translation, my bad. Still, that's yields for you.

[sof006]

50 minutes ago, patrickjp93 said:

At 6.912TFlops per chip, do the math. You can fit up to 8 in a 1U rack (4 CPUs, 4 accelerators)

 

100 PetaFLOPS (not petabytes :P) would be the most powerful supercomputer on the planet right now.

Thank you, someone was nice enough to understand what I mean and be nice enough to not be a dick. I obviously meant flops and not bytes, i'm tired. This forum is becoming a cancerous place to be.

[NumLock21]

Supermicro already have servers for it.

https://www.supermicro.com/products/system/2U/5028/SYS-5028TK-HTR.cfm

[BuckGup]

5 hours ago, sof006 said:

Jesus... Thats a lot of cores. Wonder how many of them you'd need to make a 100 Petabyte machine...

Ask china