Xeon Phi with 50+ cores in production in 2012

P

pxc

Extremely [H]
Joined
Oct 22, 2000
Messages
33,063
It's the Larrabee/Knights Ferry successor codenamed Knights Corner. Hot Hardware has the presentation slides here http://hothardware.com/Reviews/Intel-Announces-MIC-Xeon-Phi-Aims-For-Exascale-Computing/

The chip can sustain 800 GFLOPS performance with DP, more than 4x what AMD's top (single) GPU is capable of sustaining (190 GFLOPS of DP), and about 8x what an 8 core E5-2600 can sustain. As the Xeon Phi name suggests, the idea of a Larrabee-based GPU is pretty much discarded. Knights Ferry (a non-commercial test chip) was used in content creation and simulation systems, hinting at the uses Intel is now targeting.

Xeon Phi looks to be set to work as both a multi-core CPU and as a high performance GPGPU processor. The advantage it has may be full x86 compatibility, which puts it far ahead of programming environments used by other GPUs (many language restrictions and critical micro-management of data is necessary, and generically written code performance is haphazard between architectures).
 
Could we theoretically toss this in anything and enhance compute power? Like, could I toss it on my Q8200 system and have it offload?
 
Jorona said:
Could we theoretically toss this in anything and enhance compute power? Like, could I toss it on my Q8200 system and have it offload?
Click to expand...

The official slides make it seem like it requires a Xeon to run...also I haven't seen official-looking release as to what instructions it supports...so depending on what you do computationally, you may not be supported.
 
Hmmm. I wonder if there will be something like this for the consumer market in the near future. Perhaps Broadwell or Skylake/Skymont? Then again, I can also see Intel reserving this for business-class computing only for quite a long while in order to get a huge return on investment since it's likely going to be pricey and they'll want to keep that higher pricing model in place for as long as possible. That is, unless nVidia fires back hard and heavy. I'm seriously considering buying a GTX 670 by end of year, but now I may just stick to my typical skip-every-other-generation GPU upgrade scheme, afterall.

Damn, this could be really good for all of us. :cool: :D
 
One thing, it's not a GPU even though that's where the roots are. It was originally meant to be both a GPU and compute much like the direction AMD and Nvidia are headed.

They advertise it as pairing up with a Xeon although from what I read it operates like a separate system running Linux and you're just connected to it. Should be pretty nice for network renders.

I'm curious about how it will pan out since AMD and NV are both approaching 1TF on 7970GHZ/W9000 and GK110.


Edit: By roots I mean that it's original designation. It sprung from a GPU/GPGPU which sprung from a pentium processor.
 
DejaWiz said:
I'm seriously considering buying a GTX 670 by end of year, but now I may just stick to my typical skip-every-other-generation GPU upgrade scheme, afterall.

Damn, this could be really good for all of us. :cool: :D
Click to expand...

this has nothing to do with that, and you and I will not be using this. the reason it is an expensive server product is because it has no other use.

Jorona said:
Could we theoretically toss this in anything and enhance compute power? Like, could I toss it on my Q8200 system and have it offload?
Click to expand...

no. I'd say not in a million years, but maybe in a million years.
 
ImperfectLink said:
I'm curious about how it will pan out since AMD and NV are both approaching 1TF on 7970GHZ/W9000 and GK110.
Click to expand...
Peak theoretical != sustained throughput. The numbers Intel is giving for Knights Ferry are sustained DP throughput in the OpenCL DGEMM benchmark. There are special cases which can be optimized more (AMD has achieved much higher rates that way), but the standard benchmark is more comparable.

The Tesla version of the GK110 may (or may not) achieve higher DP throughput than the 7970, but it tops out with a theoretical peak throughput of 665 GFLOPS on DP workloads. I haven't seen DGEMM scores for the GK110.

If that performance lead is typical, which it likely isn't anyways, nv and AMD would probably be a year or two behind Intel in those types of workloads where KF excels. It still doesn't solve the general programability differences.
 
Thanks for the info pxc. I hope the price doesn't drown out the performance on the KF. It could present a very attractive rendering solution.
 
bAMtan2 said:
this has nothing to do with that, and you and I will not be using this. the reason it is an expensive server product is because it has no other use.
Click to expand...

Eventually we will be. Same thing was said when 3D capable graphics processors were first talked about.

"They're just an expensive niche product that won't ever catch on."

"3D processing is nice for businesses involved with things like CAD or movie CGI, but the technology will never become viable for the consumer in regards to wide-spread availability or affordable pricing."

How did that turn out, again?
 
16 GFLOPS per core? That's the best we can do in 2012? Where are the flying cars, the holodecks and 10 TP per core chips?

This is supposed to be the future, and the best Intel can do is a sub 1TF chip with 50 cores?

We should have 10TF chips. I know Intel has 14nm running in the lab. They've probably already created a die-shrink of Haswell (Broadwell). Let's put 500 cores on a chip. Let's crank these babies out and get an Exaflop supercomputer by Christmas.

Come on!
 
aphexcoil said:
16 GFLOPS per core? That's the best we can do in 2012? Where are the flying cars, the holodecks and 10 TP per core chips?

This is supposed to be the future, and the best Intel can do is a sub 1TF chip with 50 cores?

We should have 10TF chips. I know Intel has 14nm running in the lab. They've probably already created a die-shrink of Haswell (Broadwell). Let's put 500 cores on a chip. Let's crank these babies out and get an Exaflop supercomputer by Christmas.

Come on!
Click to expand...

It's the best product for the price Intel believes the market is willing to pay.

Welcome to capitalism.
 
DejaWiz said:
Eventually we will be. Same thing was said when 3D capable graphics processors were first talked about.

"They're just an expensive niche product that won't ever catch on."

"3D processing is nice for businesses involved with things like CAD or movie CGI, but the technology will never become viable for the consumer in regards to wide-spread availability or affordable pricing."

How did that turn out, again?
Click to expand...

we will not be using this eventually. this will be on ebay for $8.00 before we need or want 50 x86 cores.
 
aphexcoil said:
16 GFLOPS per core? That's the best we can do in 2012? Where are the flying cars, the holodecks and 10 TP per core chips?

This is supposed to be the future, and the best Intel can do is a sub 1TF chip with 50 cores?

We should have 10TF chips. I know Intel has 14nm running in the lab. They've probably already created a die-shrink of Haswell (Broadwell). Let's put 500 cores on a chip. Let's crank these babies out and get an Exaflop supercomputer by Christmas.

Come on!
Click to expand...
Where have you been ?????
http://www.flixxy.com/volkswagen-levitating-car.htm
 
pxc said:
Peak theoretical != sustained throughput. The numbers Intel is giving for Knights Ferry are sustained DP throughput in the OpenCL DGEMM benchmark. There are special cases which can be optimized more (AMD has achieved much higher rates that way), but the standard benchmark is more comparable.

The Tesla version of the GK110 may (or may not) achieve higher DP throughput than the 7970, but it tops out with a theoretical peak throughput of 665 GFLOPS on DP workloads. I haven't seen DGEMM scores for the GK110.

If that performance lead is typical, which it likely isn't anyways, nv and AMD would probably be a year or two behind Intel in those types of workloads where KF excels. It still doesn't solve the general programability differences.
Click to expand...

wow. Are you Intel's hired gun?? GK110 is spec to have SP performance of 5.184TFLOPS and DP performance of 1.728TFLOPS. This is double the performance of Xeon Phi. Unless Intel sells it cheap, it is not going to work.
 
Intel getting into this market is a very good thing for the world IMO, now it's not just AMD & nVidia we are relying on to push the || computing envelope...
 
ymc said:
wow. Are you Intel's hired gun?? GK110 is spec to have SP performance of 5.184TFLOPS and DP performance of 1.728TFLOPS. This is double the performance of Xeon Phi. Unless Intel sells it cheap, it is not going to work.
Click to expand...
Try reading the post before dismissing it. Thx.
 
You must log in or register to reply here.
Back
Top