XBitlabs: Intel Shows Off "Knights Corner" MIC Compute Accelerator, Beats Nvidia's Fe

[Idontcare]

Intel Shows Off "Knights Corner" MIC Compute Accelerator, Beats Nvidia's Fermi.

Intel's Knights Corner accelerator has over 50 cores and delivers 1TFLOPS of double precision floating point performance, as measured by the double-precision, general matrix-matrix multiplication benchmark (DGEMM). Currently the most powerful special purpose highly-parallel accelerator is Nvidia Tesla 2090, which boasts with 665TFLOPS (sic) of peak performance, which is considerably below peak performance of Intel's KNC.

Source

Someone is going to have to explain this to me because as I read it Intel basically just said it takes a near 2-node performance advantage (in Intel's favor) in order for their approach to be competitive.

Intel needs their super-duper 3D xtor 22nm process tech in order for Knights Corner to beat Nvidia's 40nm (not even HKMG let alone 3D xtor, and 40nm to boot!) Fermi chip?

Am I reading that right? If I am, this is hardly flattering, why would Intel even make such a comparison?

Nvidia's 28nm GPU's are likely to blow these numbers out of the water, let alone whatever they release on 20nm.

 

[RavenSEAL]

I'm trying to read that and all I hear is BLAHHHHHHH. My mind capacitors are blown from helping Mac users over the phone...

So now i am intrigued as hell...can someone translate it to derp dialect so i can try to figure it out before I pass out?

 

[Silverforce11]

HPC is all about perf/watt, if intel can beat the competitors there, they have an awesome shot.

 

[Rifter]

That makes no sense to me either. Obviosly a brand new cutting edge chip on a brand new process will beat a 2+ year old tech chip made with a even older process, i mean no crap it beats it why announce it like its a good thing.

Thats like saying a 2012 BMW M3 beats a 2009 BMW M3, well no crap i would hope in 3 years you could improve tech.

 

[Silverforce11]

Does it matter what process it's on, because Intel is always ahead with their process tech.

Even if Kepler doubled Fermi performance (unlikely, an 80&#37; increase is probably the max), it will match KC MIC in dp performance. At what power use though? <-- key point here.

 

[Fox5]

A 6990 breaks 1TF of double precision floating point performance.

 

[boxleitnerb]

Still, I think the process evolution will slow down in the next couple of years and then Intels advantage will shrink. It is rumored that Kepler will more than double DP performance compared to Fermi (1.4-2 teraflop are the current rumors). This will be an interesting battle next year and in the future.

 

[Borealis7]

A 6990 breaks 1TF of double precision floating point performance.

yes but how many Watts it takes to do that? i imagine KNC offers that 1TFlop at low power or else this product isnt viable.

 

[Will Robinson]

Intel's experience is heavily biased towards out of order CPU architecture as against NVDA's parallel GPU processing.
It's not unreasonable to expect Intel's (early) solution(s) to be less elegant than NVDA's, given that.
Also,aren't you a microchip process engineer?...thought this would be right up your alley.

 

[SunnyD]

Is this what became of Larrabee?

 

[Arkadrel]

Intel's Knights Corner accelerator has over 50 cores and delivers 1TFLOPS of double precision floating point performance

Currently the most powerful special purpose highly-parallel accelerator is Nvidia Tesla 2090, which boasts with 665GFLOPS (0.665TFLOPS)

Intels KC's 1.0 TFLOPs > Nvidia Tesla 2090 0.665 TFLOPs (~50&#37; more performance or so)

But more than performance, is performance/watt.
Intels KC is gonna use ALOT less power than Nvidia's GPUs do, to reach those levels of performance.


Intel wants to "steal" nvidia's lucrative server GPU bussiness.

down the line this is gonna hurt amd/nvidia, because all that GPGPU stuff might be for nothing
(if they cant sell cards for it, because Intel is sitting on the market).


This might mean, AMD and Nvidia either try to do less GPGPU stuff in the future, or they go more into it (meaning we ll all get less performance/watt for gameing purposes).

A 6990 breaks 1TF of double precision floating point performance.

Yes but at what wattage? 400watts?
Intels KC might be like 40watts or something

in short, if Intel has like 10x performance/watt of Nvidia/AMD, and cheaper chips......
Intel is gonna take over intire GPGPU market with its KC cards, in the server market.

 

[Lonyo]

Intels KC's 1.0 TFLOPs > Nvidia Tesla 2090 0.665 TFLOPs (~50% more performance or so)

But more than performance, is performance/watt.
Intels KC is gonna use ALOT less power than Nvidia's GPUs do, to reach those levels of performance.


Intel wants to "steal" nvidia's lucrative server GPU bussiness.

down the line this is gonna hurt amd/nvidia, because all that GPGPU stuff might be for nothing
(if they cant sell cards for it, because Intel is sitting on the market).


This might mean, AMD and Nvidia either try to do less GPGPU stuff in the future, or they go more into it (meaning we ll all get less performance/watt for gameing purposes).

The fact that it's compatible with existing x86 code means that it's already a different beast to Fermi.
And even if it's more than 20w of power when the final product comes out, it will still be significantly less than Fermi without a doubt.

 

[Arkadrel]

@Lonyo

Yeah... this isnt like a small victory. This is a HUGE win for Intel.

Like x10 times less power use than Nvidia/AMD, and reaching same levels of performance.
Compatibility with x86 code has to be a bonus too.

Also this being the first of its kinde by Intel, Im sure the future products in this line are just gonna get better with time. This is gonna push nvidia out of its GPGPU bussiness, for servers.

 

[Genx87]

Does it matter what process it's on, because Intel is always ahead with their process tech.

Even if Kepler doubled Fermi performance (unlikely, an 80&#37; increase is probably the max), it will match KC MIC in dp performance. At what power use though? <-- key point here.

Kepler is expected to quadruple the performance\watt of Fermi. So either we can expect a Kepler GPU to provide 2.7TFLOPS of performance at the same level of power consumption of Fermi, or provide 655GFLOPS at 1/4th the power.

But remember, this is on a process one node larger than Intel. Meaning, it is possible this thing isnt impressive at all.

A more appropriate comparison would be with Maxwell as it will be on the same node. That is expected to deliver 14x performance\watt of Fermi. An estimated 9.2TFLOPS at the same power.
What does Tesla consume? 225Watts? If Intel hits their power threshold it would take over 9 of these to equal one Tesla or about 180-200 watts. So it is possible with the infrastructure required to hook 9 of these up that 25-45 watt difference will be nullified.

 

[happy medium]

Kepler will improve the DP performance at least two folds and will be available for purchase in H1 2012.
KNC is no where to be seen.

22nm vs 40 nm? get back to this in 2 months when AMD and Nvidia are at 28nm.

 

[SirPauly]

That 22nm tri-gate process is some differentiation for them.

 

[Dribble]

The 20W figure is the aim for the 2018 excascale chip, not KC. It's also for the whole subsytem, the cpu of the 2018 chip is meant to be 5W.

 

[sm625]

Just wait until Intel takes a half a dozen KC cores and ties their schedulers together with a sb/ib core. In essence, KC becomes a part of the FPU. Imagine 16 or 32 SSE4 instructions per clock cycle. Bye bye AMD. I tried to warn you but you stood there with your thumbs up your butts.

 

[happy medium]

I'm guessing these chips would require a whole new system/board rather than a pci-e swap?

Would this be expensive for customers?

 

[Nemesis 1]

Someone is going to have to explain this to me because as I read it Intel basically just said it takes a near 2-node performance advantage (in Intel's favor) in order for their approach to be competitive.

Intel needs their super-duper 3D xtor 22nm process tech in order for Knights Corner to beat Nvidia's 40nm (not even HKMG let alone 3D xtor, and 40nm to boot!) Fermi chip?

Am I reading that right? If I am, this is hardly flattering, why would Intel even make such a comparison?

Nvidia's 28nm GPU's are likely to blow these numbers out of the water, let alone whatever they release on 20nm.

Intel MIC products will offer both high performance from an architecture specifically designed to process highly parallel workloads, and compatibility with existing x86 programming model and tools. One of the benefits of Intel MIC architecture is the ability to run existing applications without the need to port the code to a new programming environment. This will allow scientists to use both CPU and co-processor performance simultaneously with existing x86 based applications, dramatically saving time, cost and resources that would otherwise be needed to rewrite them to alternative proprietary languages

 

[ocre]

Is this what became of Larrabee?

I would say, Yes!

Intels KC's 1.0 TFLOPs > Nvidia Tesla 2090 0.665 TFLOPs (~50% more performance or so)

But more than performance, is performance/watt.
Intels KC is gonna use ALOT less power than Nvidia's GPUs do, to reach those levels of performance.


Intel wants to "steal" nvidia's lucrative server GPU bussiness.

down the line this is gonna hurt amd/nvidia, because all that GPGPU stuff might be for nothing
(if they cant sell cards for it, because Intel is sitting on the market).


This might mean, AMD and Nvidia either try to do less GPGPU stuff in the future, or they go more into it (meaning we ll all get less performance/watt for gameing purposes).



Yes but at what wattage? 400watts?
Intels KC might be like 40watts or something

in short, if Intel has like 10x performance/watt of Nvidia/AMD, and cheaper chips......
Intel is gonna take over intire GPGPU market with its KC cards, in the server market.

No way! Its not going to blow away nvidia cards in massively parallel workloads. Especially in the same power envelope. Think about it. If they could do this then they wouldve had a GPU out by now.

Everyone think, Mayfield! Intel hyped it to the moon.

I wouldnt be surprised if a chip like that would excel in a few cases but there is no way what so ever it can be more efficient at extreme parallel workloads.

where is this chip at? Its just crazy to believe that parallel processing is best with x86 cores. Why did we ever make GPUs in the first place? There is a reason, GPUs are much better at processing boatloads of extremely parallel codes. This is what they do.

While we all wait for Mayfield an this Knights Corner........

Have you guess seen nvidias projections? Look at the jump just for 2013-

http://www.engadget.com/2010/09/21/nvidia-reveals-fermis-successor-kepler-at-28nm-in-2011-maxwel/


Maxwell will be a beast.

 

[Concillian]

Where I work, we paid tens of thousands of dollars and many months to get a software version made that could work on a CUDA solution so our physics modelers could have something to put on their desk instead of sharing time on a cluster.

Do not underestimate this being an x86 solution that wouldn't need new software.

I'm sure performance per watt is better than Fermi, but I think that's relatively minor compared to being x86.

 

[wlee15]

AMD's GCN will bring 1/2 rate DP to their high end GPU compared to 1/4 rate in the current Cayman. So you probably get an 3 to 4 times the DP FLOPS when compared to current AMD GPU. Sure Knights Corner might be competitive at 35W, but what if it's 65W? 95W? 130W? 300W?

 

[grimpr]

This propably works on Intels software tools and librarys with 0 open source and microsoft support, yet and most propably will require major code rewrite to take advantage of the beefy vector cores in KC, Intel just blows PR on ease of development on this thing.

 

[IlllI]

intel's drivers will still suck