Intel Larrabee finally hits 1TFLOPS - 2.7x faster than nVidia GT200!

[ajaidev]

"During the recently held SC09 conference in Portland, Oregon - Intel finally managed to reach its original performance goal for Larrabee. Back in 2006, when we first got the first details about Larrabee, the performance goal was "1TFLOPS@ 16 cores, 2.0 GHz clock, 150W TDP". During Justin Rattner's keynote, Intel demonstrated the performance of LRB as it stands today.

At SGEMM Performance test [4K by 4K Matrix Multiply, QCD], Intel achieved 417 GFLOPS using half the cores on the prototype card, and reached 825 GFLOPS by enabling all the cores. While looking at the numbers alone, one might think that these scores are below the level of ATI Radeon 4850 and nVidia GeForce GTX 280/GTX 285..."


http://www.brightsideofnews.com/news...ia-gt200!.aspx

So using the assumptions he said 5870 = 2*4870" so 5970 = 2*5870 and so 5970 = 4*300 or 1200 gflops hum. I know i know the clocks and the bandwidth, etc.... anyways someone should try to run this on the 5970 can you pass the 1Tflop mark if yes your thing is as powerful as a LRB

[zalbard]

http://www.xtremesystems.org/forums/...d.php?t=239176
BSN is slow.

[kl0012]

So using the assumptions he said 5870 = 2*4870" so 5970 = 2*5870 and so 5970 = 4*300 or 1200 gflops hum. I know i know the clocks and the bandwidth, etc.... anyways someone should try to run this on the 5970 can you pass the 1Tflop mark if yes your thing is as powerful as a LRB

You're mixing up theoretical FLOP/S and real FLOP/S. In fact theoretical SP perf for LRB should be arround 2 TFLOP/S (32 cores * 32 ips * 2 GHz). But in real life you're some times limited by other factors (such as mem bandwidth) especially in apps such matrix multiplication.

[Hornet331]

It's not larrabee, its some sort of polaris derivative.

Edit, so it is larrabee. Seems theregister has screwed smething up in there article.

[ajaidev]

You're mixing up theoretical FLOP/S and real FLOP/S. In fact theoretical SP perf for LRB should be arround 2 TFLOP/S (32 cores * 32 ips * 2 GHz). But in real life you're some times limited by other factors (such as mem bandwidth) especially in apps such matrix multiplication.

Are you kidding me, i was talking about real flops/s how can a 5970 "2*320 FAT shaders" not out around 1000gflops aka 1Tflops when a card based on RV770 "160 FAT shaders " can out 300gflops.

In theory 5970 is good for 4.6 Tflops, but in reality it has to be good enough for at least 1 Tflop....

[BulldogPO]

This will be mighty WCG cruncher

[Tim]

Seems like Larrabee is more and more pitched to GPGPU. I think nVidia and Intel are on to this massive market, but would very much like to see what this means for us gamers.

[LedHed]

haha you can't base the performance of a card solely off it's teraflops, that is just silly. I mean look at the ATI 2900 series for example, great numbers, crappy performance.

[deeperblue]

We had all this discussion already in the old thread.

4870 gets around 900GFlop/s in Matrix Multiplication with optimized algorithm linked to in the old thread. 4890 breaks the Tflop/s.

So 5870 should be good for a much higher performance. Of the 5 shaders 4 can be used for the matrix multiplication. Theoretically this could give us 2Tflop/s for the 5870.

Let's wait until LRB finally get's released.

(and I also thought this presentation was about the 80core Polaris multicore test chip)

UPDATE: here the link
http://cerberus.fileburst.net/showthread.php?t=54842

[Hornet331]

We had all this discussion already in the old thread.

4870 gets around 900GFlop/s in Matrix Multiplication with optimized algorithm linked to in the old thread. 4890 breaks the Tflop/s.

So 5870 should be good for a much higher performance. Of the 5 shaders 4 can be used for the matrix multiplication. Theoretically this could give us 2Tflop/s for the 5870.

Let's wait until LRB finally get's released.

(and I also thought this presentation was about the 80core Polaris multicore test chip)

You can't compare a optimized algorithm with the algorithm in the benchmark... it would be like pitting superpi against hyperpi...

[deeperblue]

You can't compare a optimized algorithm with the algorithm in the benchmark... it would be like pitting superpi against hyperpi...

But what do you think Intel is running there?
http://www.ats.ucla.edu/clusters/com..._benchmark.htm
It's of course running there optimized Intel MKL library optimized for especially that purpose and that hardware. Nvidia is also using there CUBLAS implementation. AMD only needs to update there ACML library with the algorithm in the link I posted...

[naokaji]

Larrabee? meh... lets see how many years its going to take until software makes proper use of it after it launches at a unknown date in the future first before praising it.

[BulldogPO]

It is X86 so softwares should work pretty soon.

[LedHed]

I think it will be at least 3 years before we see Larrabee being used in anything worthwhile to actually buy the card.

[anubis]

I think it will be at least 3 years before we see Larrabee being used in anything worthwhile to actually buy the card.

I`ll BOINC it, when its available and can do WCG

[Smartidiot89]

haha you can't base the performance of a card solely off it's teraflops, that is just silly. I mean look at the ATI 2900 series for example, great numbers, crappy performance.

haha bother to read the article and edit your post.

This is no theoretical performancebeast like HD2900 or any of AMDs GPU line-up, this is actual, real-world performance where it outperforms GTX285 by 170%.

Intel opens a can of whoop-ass on the GPGPU market?

[LedHed]

haha bother to read the article and edit your post.

This is no theoretical performancebeast like HD2900 or any of AMDs GPU line-up, this is actual, real-world performance where it outperforms GTX285 by 170%.

Intel opens a can of whoop-ass on the GPGPU market?

It's kind of hard to open an article when the site is down.

but please show me anywhere in this article that benches larrabee against a GTX 285 in real world apps. All of the numbers are theoretical based on a synthetic benchmark, not actual real world performance in gaming/apps.

[Helloworld_98]

150w?

not bad considering that's less than a 5850 and it will be in a league of it's own in GPGPU.

[Olivon]

Intel shows off CPU with 48 cores

[ajaidev]

Intel shows off CPU with 48 cores

This is news worthy, you should have posted it in the front page IMO.

ANyway's the 567 mm2 means small cores that means multi threaded performance not single threaded. Great for cloud and other multi thread app's, where the hell is SUN's rock cpu???

[Helloworld_98]

Intel shows off CPU with 48 cores

567mm2?

any chance it's a typo of 576mm2, which is the size of GT200.

[Face]

http://www.youtube.com/watch?v=ynjYu...layer_embedded

[Olivon]

http://www.youtube.com/watch?v=ynjYu...layer_embedded

Thx for the video Face

ajaidev : corrected

[eric66]

whats is impresive in this ati already reached teraflop i think 1.8 or something

[god_43]

so the the 5XXX and nvidia are still way ahead of intel? this just proves that they are not building vaporware. by the time Larrabee comes out.....it will be blown away. i guess intel has to start some where right?