Knights Landing announced

[mikk]

More concrete infos from here: http://www.heise.de/newsticker/meld...em-InfiniBand-und-3-TFlops-DGEMM-2049668.html


- ~3 Tflop DGEMM
- 14-16 Gflops/W
- up to 72 modified Silvermont cores with quadruple Hyperthreading
- bigger Caches
- Out-of-Order FPU (Atom Silvermont only does In-Order)
- AVX512
- up to 384 GByte DDR4/up to 100 GByte/s
- 8 or 16 GB stacked memory with up to 500 GByte/s
- PCIe 3.0
- 14nm

 

[ShintaiDK]

More concrete infos from here: http://www.heise.de/newsticker/meld...em-InfiniBand-und-3-TFlops-DGEMM-2049668.html


- ~3 Tflop DGEMM
- 14-16 Gflops/W
- up to 72 modified Silvermont cores with quadruple Hyperthreading
- bigger Caches
- Out-of-Order FPU (Atom Silvermont only does In-Order)
- AVX512
- up to 384 GByte DDR4/up to 100 GByte/s
- 8 or 16 GB stacked memory with up to 500 GByte/s
- PCIe 3.0
- 14nm

I would be very, very worried if I was in nVidias Tesla department.

 

[Khato]

Well they don't need to be too worried quite yet considering that Knight's Landing isn't going to show up until late next year at the earliest. Which means it should be going up against Maxwell rather than Kepler... but NVIDIA's roadmap leaves a lot of interpretation as to what their projected performance per watt is - Kepler is between 4 and 8 while Maxwell is between 8 and 16. We know that Kepler is slightly on the low side of its range at 5-5.5, so it really depends on where Maxwell lands.

Regardless, the kind of numbers that Knight's Landing is supposed to provide according to that are simply crazy. And even better it'll be going to an instruction set that'll eventually be in Intel CPUs as well.

 

[Ajay]

I would be very, very worried if I was in nVidias Tesla department.

Well, it sure is clear why NV is partnering with IBM. KL will catch up to Maxwell in performance and perf/watt, but Intel has NV beat on system design and integration by a mile this go around. Hard to see how KL will compete on price though, unless Intel treats it like a loss leader to snuff out competition from NV (by taking SC wins, and hence headlines, away).

 

[NTMBK]

Well they don't need to be too worried quite yet considering that Knight's Landing isn't going to show up until late next year at the earliest. Which means it should be going up against Maxwell rather than Kepler... but NVIDIA's roadmap leaves a lot of interpretation as to what their projected performance per watt is - Kepler is between 4 and 8 while Maxwell is between 8 and 16. We know that Kepler is slightly on the low side of its range at 5-5.5, so it really depends on where Maxwell lands.

Regardless, the kind of numbers that Knight's Landing is supposed to provide according to that are simply crazy. And even better it'll be going to an instruction set that'll eventually be in Intel CPUs as well.

When do you think big Maxwell is going to turn up? :hmm:

 

[Homeles]

Here comes the stacked memory... honestly, it'll be really nice to see that hit the market.

 

[ShintaiDK]

Here comes the stacked memory... honestly, it'll be really nice to see that hit the market.

I wonder when we will see regular CPUs with stacked memory.

On the GPU side we know nVidia will do it with Volta in something like a 2016-2017 timeline.

 

[NTMBK]

Interesting that this claims Silvermont cores- David Kanter basically dismissed that prospect out of hand. They'd have to be pretty heavily modified! http://www.realworldtech.com/knights-landing-cpu-speculation/

 

[Ajay]

Interesting that this claims Silvermont cores- David Kanter basically dismissed that prospect out of hand. They'd have to be pretty heavily modified! http://www.realworldtech.com/knights-landing-cpu-speculation/

Thanks for the link - I thought that was an odd jump from a modified P54C to a modified Silvermont.

 

[mikk]

Interesting that this claims Silvermont cores- David Kanter basically dismissed that prospect out of hand. They'd have to be pretty heavily modified! http://www.realworldtech.com/knights-landing-cpu-speculation/

Yes he was wrong.

 

[Arachnotronic]

Yes he was wrong.

Yep. They don't have the budget to do a full on custom core, but they've pulled a number of the sharpest folks from the Core team to help with the Atom modification efforts. I know some geniuses on that team, so it'll probably be very good.

 

[podspi]

More concrete infos from here: http://www.heise.de/newsticker/meld...em-InfiniBand-und-3-TFlops-DGEMM-2049668.html


- ~3 Tflop DGEMM
- 14-16 Gflops/W
- up to 72 modified Silvermont cores with quadruple Hyperthreading
- bigger Caches
- Out-of-Order FPU (Atom Silvermont only does In-Order)
- AVX512
- up to 384 GByte DDR4/up to 100 GByte/s
- 8 or 16 GB stacked memory with up to 500 GByte/s
- PCIe 3.0
- 14nm

Hopefully some of that comes back into the Atom line. I wouldn't mind (normal) hyperthreading. At this rate, we're going to be rendering on our tablets...

 

[Ajay]

Hopefully some of that comes back into the Atom line. I wouldn't mind (normal) hyperthreading. At this rate, we're going to be rendering on our tablets...

Probably will be able to after a couple of more shrinks.

 

[NTMBK]

Yes he was wrong.

Apparently so- unfortunate bit of timing for him, he only published that article yesterday

 

[NTMBK]

Another write up over on El Reg: http://www.theregister.co.uk/2013/1...e_knights_landing_cpu_will_be_a_game_changer/

 

[Ajay]

Apparently so- unfortunate bit of timing for him, he only published that article yesterday

His analysis was good, right up to the end.

 

[Hans de Vries]

I would be very, very worried.

Yes, times have changed.


No politics in the technical forums, go to the politics subforum if you want.
Markfw900
Anandtech Moderator


......

 

[Ajay]

Yes, but some how the Nation University of Defense Technology is reported as the "National Super Computer Center in Guangzhou" on the Top500 SC's website. Can't let the world know that we are helping the Chinese in a way which we never would have done with the Soviet Union. Strange days indeed...

 

[Borealis7]

at first i read that as King's Landing and i was like "huh?"

 

[Leyawiin]

http://en.wikipedia.org/wiki/Knights_Landing,_California

Little farm town on the Sacramento River. I drive through it once in a while.

 

[iMacmatician]

From VR-Zone: "Intel's Xeon Phi Knights series expands in 2015" contains slides that I'm assuming are from Intel and some additional information.

A lot of the information was already seen in the heise online report, but some info that's new to me is as follows:

• ≤ 3x single-threaded performance compared to Knights Corner
• 4 threads/core, 2 AVX-512 units/core
• Some parts (KNL-F) will have an integrated fabric
• 160-200 W TDP for KNL, 165-215 W TDP for KNL-F
• Apparently mid 2015 for first variants
• PCIe coprocessors coming in late 2015

The release timeframes of Knights Landing are somewhat later than what I thought (early 2015), and so I think it's likely that NVIDIA and AMD can release 20 nm "big" chips some time beforehand. There might even be 16FF GPUs released around the same time as the PCIe Knights Landing parts.

 

[mikk]

16FF is more likely a 2016 part. Big Maswell in 20nm is expected for 2015 as well. KNL TDP looks good. 3x the performance and lower TDP.

 

[NTMBK]

The release timeframes of Knights Landing are somewhat later than what I thought (early 2015), and so I think it's likely that NVIDIA and AMD can release 20 nm "big" chips some time beforehand. There might even be 16FF GPUs released around the same time as the PCIe Knights Landing parts.

Don't be so sure about that. 20nm is seriously delayed, and GPUs don't get wafer priority any more.

 

[iMacmatician]

Don't be so sure about that. 20nm is seriously delayed, and GPUs don't get wafer priority any more.

Oh, I wasn't aware it was delayed so much. So would you say it's likely there will be another year of 28 nm in late 2014 through a significant chunk of 2015?

 

[erunion]

Don't be so sure about that. 20nm is seriously delayed, and GPUs don't get wafer priority any more.

It is? Any links you could share?