Exclusive: Intel Larrabee 5 times faster than Nvidia, AMD

[003]

http://softwarecommunity.intel.com/U...e_manycore.pdf

The theoretical throughput for a single larrabee core at 1 GHz is 32 GFLOPs, at 16 cores 1 GHz this comes in at 512 GFLOPs, running at 2 GHz would be 1 TFLOP, 64 running at 2 GHz would be 4 TFLOP, 64 core running at 2.5 GHz would be 5 TFLOP.

Your error is assuming Larrabee is simply a single Pentium MMX core ... this is too simplified of an assumption, Intel has constructed the core to process 4 threads with up to 16 SP vector operations per clock per core.

Jack

Ok then, regardless. 64 cores at 2GHz is only 4TFlops, where as current generation quad 4870s or dual 4870x2s put out 4.8TFlops of processing power.

By the time Larrabee ships, we will be at R900 or R1000, plus whatever nvidia has at that time. Right now, all intel has going is it's brand name and widespread corporate acceptance. Unless they step up their game, Larrabee will be blown out of the water before it sees the light of day.

[T_Flight]

^^^

That stuff that the wh00pass master said is the "MOTHER of all Screwups!" Just like in war, you NEVER (NEVER EVER EVER) understimate your competition. Especially competition that has the power to assimilate you just like they did AMD. I guess he didn't learn anything at all about Intel and has no clue at all, and that may very well be his undoing. I'm an nVidia user, but that's as big of a blunder as you can make to think that way. If he rests on his laurels and continues his slow progress he may just find himself surpassed so far by advanced and superior technology he may never again be able to catch up.

people that doubt intel forget that this is a completely vertically integrated company that basically turns sand into money as Nvidia well knows, with the best manufacturing tech in the industry and the abilty to hire/buy experts just as well as anyone else.

Intel is literally the GOD of computing right now, and like I said above, those who underestimate them do so to their complete failure as a buisness. We've already seen the results of that from AMD. I hope Mr. Wh00pass seriously gets a grip and realizes the reality he faces. Intel has more disposable income than Nvidia currently has in total net worth! What Mr Wh00pass is talking is commercial suicide if he doesn't snap out of it. That's no joke. He may not like his competition, but he better start respecting them, before his company too ends up becoming assimilitated. Resistence is futile right now, and they most certainly have the power.

[JumpingJack]

Ok then, regardless. 64 cores at 2GHz is only 4TFlops, where as current generation quad 4870s or dual 4870x2s put out 4.8TFlops of processing power.

By the time Larrabee ships, we will be at R900 or R1000, plus whatever nvidia has at that time. Right now, all intel has going is it's brand name and widespread corporate acceptance. Unless they step up their game, Larrabee will be blown out of the water before it sees the light of day.

I am simply arguing the information that is out there and pointing out your error ... however, is it reasonable to assume Larrabee would be competitive if it actually delivered 4-5 TFLOPs within the next year or two? Prehaps. Do you think ATI will make a single chip = 4 of their current gen chips in a year or two? Not likely. Will Intel come out with Larrabee and immediately be on top? Not possible in my opinion.

Jack

[saaya]

if they really place the memory BELOW the gpu then that would be pretty stupid

they have 75W from the pciE slot, plus 150W plug, plus 75W plug... so that adds up to 300W... i guess the max we can expect will be 225W though...

depending on how well itll do against amds and nvidias solutions intel will probabaly reduce the power and itll only run at around 175W, the same as current high end cards.

now regarding the bold claims from intel...

Intel has revealed that each Larrabee core will pack a supercharged ALU capable of 16 vector operations per cycle. With 64 cores that works out to no less than 1,024 vector ops per cycle. By comparison, AMD's top GPU, the Radeon HD 4870, can do 800 per cycle while NVIDIA's GeForce GTX 280 knocks out a relatively modest 240.

gtx280 is faster than 4870, i dont think anybody disagrees with that, right?
so comparing those raw numbers 4870 is almost 3x faster than the 4870... which makes it pretty obvious how completely useless those numbers are to compare actual graphics performance...

not to mention that even IF larabees effective gpu power is 5x as much as a current high end card... well even then thats nothing special and will only mean intel will be able to catch up with amd and nvidia.

by the end of 2009 we should have 2x as powerful gpus as we have now, if not more, and we will be able to use 2, 3 or even 4 of them, which means intel will need at least 2 larabee cards/gpus to compete with the high end gpus at the end of 2009... more likely they will need 3 laranee gpus to capture the performance crown. so they need something like sli or xfire... something you cant just build up over night as both nvidia and ati have shown with crossfire.

so yeah, intel is on track to deliver a powerful gpu, but it doesnt sound like it will outperform amd and nvidia by a huge margin and will be really revolutionary... depending on how well they can use the raw power they might be able to outperform amd and nvidia, but not by a huge margin... time will tell...

but in the end, im pretty sure intel wont have any multi gpu tech or not a sophisticated one that works as good as xfire and sli (which dont really work that good even after years of tweaking). so intel will be stuck with one chip...

now, regardless of how powerfull that one chip is, they are limited by the power, which is 200W max for one chip... more can simply not be cooled on air in the tight space between the pciE slots. but amd and nvidia have multi gpu solutions, allowing them to use 2, 3 and 4 gpus.

what does this mean?
amd and nvidia can break their processing power and tdp to 2 3 or 4 chips, which means better yields and cheaper mfc costs and higher overall tdp for the entire solution. so in the end it all boilds down to this:

without something like sli or xfire intel wont be able to capture the performance crown! its impossible...

as i said, time will tell

[003]

gtx280 is faster than 4870, i dont think anybody disagrees with that, right?

In todays game engines, yes. But when it eventually comes to ray tracing and the like, raw TFlops will start to come into much greater play. The 280 is 0.933 TFlops, where as the 4870 is 1.2TFlops. And with ray tracing, multi GPU configurations will scale linearly.

[Vozer]

without something like sli or xfire intel wont be able to capture the performance crown! its impossible...

For that, Intel will rely on LucidLogix's HYDRA Technology.

[3Design]

without something like sli or xfire intel wont be able to capture the performance crown! its impossible...

as i said, time will tell

Is'nt there more to be gained from the mid range maket ??

[Epsilon84]

I know its totally worthless basing performance estimates off theoretical numbers, but *IF* it really performs as well as 4870X2 CF or better, even by 2010 standards that is pretty good performance for a first up attempt. It'd be equivalent to a new company coming into the GPU market today with a product equal to say, a 8800GTX SLI setup (top end from 2 years ago). That's nothing to sniff at, even if it doesn't take the outright performance crown.

saaya, in regards to multi-GPU, read up on Hydra:
http://www.anandtech.com/tradeshows/showdoc.aspx?i=3379
http://www.anandtech.com/video/showdoc.aspx?i=3385

[Harshal]

50X50mm now dats a big bad chip!!

About the power... I say with NV and AMD knowing more how things will be they would be in same performance range if not high by Larrabee is out!!

[therightway]

if they really place the memory BELOW the gpu then that would be pretty stupid

they have 75W from the pciE slot, plus 150W plug, plus 75W plug... so that adds up to 300W... i guess the max we can expect will be 225W though...

depending on how well itll do against amds and nvidias solutions intel will probabaly reduce the power and itll only run at around 175W, the same as current high end cards.

now regarding the bold claims from intel...



gtx280 is faster than 4870, i dont think anybody disagrees with that, right?
so comparing those raw numbers 4870 is almost 3x faster than the 4870... which makes it pretty obvious how completely useless those numbers are to compare actual graphics performance...

not to mention that even IF larabees effective gpu power is 5x as much as a current high end card... well even then thats nothing special and will only mean intel will be able to catch up with amd and nvidia.

by the end of 2009 we should have 2x as powerful gpus as we have now, if not more, and we will be able to use 2, 3 or even 4 of them, which means intel will need at least 2 larabee cards/gpus to compete with the high end gpus at the end of 2009... more likely they will need 3 laranee gpus to capture the performance crown. so they need something like sli or xfire... something you cant just build up over night as both nvidia and ati have shown with crossfire.

so yeah, intel is on track to deliver a powerful gpu, but it doesnt sound like it will outperform amd and nvidia by a huge margin and will be really revolutionary... depending on how well they can use the raw power they might be able to outperform amd and nvidia, but not by a huge margin... time will tell...

but in the end, im pretty sure intel wont have any multi gpu tech or not a sophisticated one that works as good as xfire and sli (which dont really work that good even after years of tweaking). so intel will be stuck with one chip...

now, regardless of how powerfull that one chip is, they are limited by the power, which is 200W max for one chip... more can simply not be cooled on air in the tight space between the pciE slots. but amd and nvidia have multi gpu solutions, allowing them to use 2, 3 and 4 gpus.

what does this mean?
amd and nvidia can break their processing power and tdp to 2 3 or 4 chips, which means better yields and cheaper mfc costs and higher overall tdp for the entire solution. so in the end it all boilds down to this:

without something like sli or xfire intel wont be able to capture the performance crown! its impossible...

as i said, time will tell

I don't think Larrabee will be faster than Nvidia's or Amd's solution in terms of raw output. But I don't think that will be larrabee's selling points. I think it's selling points will be software based rendering. Games specifically designed for Larrabee will look a lot better than Nvidia or AMD because it's x86 code. There are so many possibilities with x86 code. There will be a whole new generation of computer programs designed to take advantage of Larrabee. Can you say folding monster?

[naokaji]

Is'nt there more to be gained from the mid range maket ??

Yes, but you need the performance leadership for marketing the mid range stuff.

Larrabee clearly faster than current gpu's? It better be if it wants to compete with radeon 6870 quadfire / gtx580 tri sli (or whatever the 2010 cards will be called).

[savantu]

Quad 4870s are capable of just under 5TFlops. 64 pentium MMX based cores at 2GHz, I doubt would even break 1TFlop (a heavily overclocked yorkfield doesn't even break 100Gflops). The only thing intel has going is the brand name.

Is that SP or DP ? I'm pretty sure , Intel will push with DP , a reminiscence of CPU expertise.

how was the atom a huge success? man what a joke

Umhh...revising upward demand every 40 days and the original estimates from 1-2m to 25m for 2008 qualifies for a huge success.We're talking close to a third of AMD's output for a newly introduced chip.

I don't think Larrabee will be faster than Nvidia's or Amd's solution in terms of raw output. But I don't think that will be larrabee's selling points. I think it's selling points will be software based rendering. Games specifically designed for Larrabee will look a lot better than Nvidia or AMD because it's x86 code. There are so many possibilities with x86 code. There will be a whole new generation of computer programs designed to take advantage of Larrabee. Can you say folding monster?

Larrabee is due to tape out this year , release is targeted for H2 2009 , 2010 is for backup.

However , what you say is very true.Really smart people have gathered around Intel drawn by the prospects of a new era in computing , a new paradigm that Larrabee introduces.

Senior Program Manager Phil Taylor announced his departure from Microsoft today on his blog. Taylor, one of Microsoft's top PC graphics programmers, will join chip-maker Intel to help launch a new wave in graphics processing, code-named "Larrabee."

In his message to fellow team members in Microsoft's Flight Simulator community, Taylor wrote, "Sometimes you run across an opportunity that is so big, so cool, that you just have to take it."
Taylor alluded to having served on teams that sprung some important firsts in 3D hardware/software developments. With Larrabee, Taylor expects to join a team preparing to launch what Taylor sees as the "third sea change in PC graphics."

"Larrabee has the potential to not only change the face of PC graphics, but ultimately the shape of the PC architecture as well," Taylor wrote.

http://redmondmag.com/news/article.a...orialsID=10243

When people like this jump ship for the above reasons , Intel's competitors should be afraid,very afraid.

[The Warlock]

Nvidia will have plenty off time to rename their cards until 2010, so no problem

[3Design]

However , what you say is very true.Really smart people have gathered around Intel drawn by the prospects of a new era in computing , a new paradigm that Larrabee introduces.


http://redmondmag.com/news/article.a...orialsID=10243

When people like this jump ship for the above reasons , Intel's competitors should be afraid,very afraid.

T'is true .. Intel have recruited some very talented people ( Phil among them ) .. and why I'm excited at what may come.

[Sampsa]

exclusive? wtf?

I visited Intel in San Francisco 1,5 months ago, no samples were there and I got impression that there aren't even final silicons done. Just some debug-platforms, simulations, etc. They couldn't even tell how many cores will final products have.

Sorry mate, but I think this is completely bull, see you in 1,5 weeks

[duploxxx]

the first picture is actually the best one of all i just found a new sig

leadership in volume because there are so many idiots that actually buy an igp from intel... and so many are stuck with it because they are not aware...


for the larrabee part it will be a nice co-processor for some rendering parts but then they will have to add few gb of ram on the board, all will depend on implementation for sure if it want to take some gaming marketshare.

and it might be that it has 5 times the raw calculation power of a 4870, but x2 version already exists so then its only 2,5 times the power and if r870 brings double computing power then r770 did over r670 then there is nothing left for lara ....

[r4gm4n]

no-one has asked if an extreme version will be released

[Decami]

Regardless of the size of the company or the money. Theres no way Intel will bust into the market and provide something that much faster than AMD/Nvidia, maybe over time, but def not right away. But by some crazy chance out of nowhere they do actually release a gpu that much faster than current companies, the driver support is good, and quality is as well. I have no problem with leaving Nvidia for such a card. But I very much dought this. There is a bigger chance of the military having a working "invisible" paint by 2010 than Intel jumping into the GPU market with a "god" like chip by then.

[kromosto]

totally agree with decami.

but intel gets a top seat on gpu market by starting hype's about a gpu today that will come 2 years later. what was the last record 8 months.

[Shintai]

Larrabee will be around its competitors when time comes in terms of max FPS performance. However Larrabee offers alot of bonuses and better minimum FPS performance due to flexibility.

[Decami]

I don't think Larrabee will be faster than Nvidia's or Amd's solution in terms of raw output. But I don't think that will be larrabee's selling points. I think it's selling points will be software based rendering. Games specifically designed for Larrabee will look a lot better than Nvidia or AMD because it's x86 code. There are so many possibilities with x86 code. There will be a whole new generation of computer programs designed to take advantage of Larrabee. Can you say folding monster?

Well then Intel needs to shut their mouths. Cause they have done a little bit of talking to its future competitors about how they will be faster than them. Not saying Nvidia hasn't contributed, or even that they didn't start this dispute, but none the less, they participated.

[saaya]

In todays game engines, yes. But when it eventually comes to ray tracing and the like, raw TFlops will start to come into much greater play. The 280 is 0.933 TFlops, where as the 4870 is 1.2TFlops. And with ray tracing, multi GPU configurations will scale linearly.

ok, but whats the point?
so in 2 years the 4870 will be faster than the gtx280 in the latest games, with the 4870 giving us 20fps and the gtx280 18fps... whats the point? :P

same with dx and framebuffer size...
who actually buys a low end dx10 card?
you can play dx10 games! NOT! :P
who actually buys a low end card with 1gb memory?
you can play future games and get better AA scaling!
yeah, 6fps instead of 5fps maybe :P

we still have to see how important those vector raw power is really going to be...
itll be nice for raytracing and gpgpu, but the amount of people who buy a videocard for folding at home is very limited, its a niche market and if a product performs well there it doesnt make it a successful videocard!

I don't think Larrabee will be faster than Nvidia's or Amd's solution in terms of raw output. But I don't think that will be larrabee's selling points. I think it's selling points will be software based rendering. Games specifically designed for Larrabee will look a lot better than Nvidia or AMD because it's x86 code. There are so many possibilities with x86 code. There will be a whole new generation of computer programs designed to take advantage of Larrabee. Can you say folding monster?

i still have to see raytracing becoming the next big thing...
there is a lot of cool stuff you can do with raytracing, yes, if so, how come NOBODY has shown ANY engine demo showing those awesome and beautiful rendered ratraced grapics? all mayor game studios show off their new engines every now and then. the first time we saw doom3 engine shots was 2 or 3 years before it came out, the first time we saw far cry engine shots was 3 years before it came out, same for UT3 etc etc...

so i ask you, where are those raytracing demos showing the raytracing infrastructure larabee will be so great in and outperform amd and nvidia?

there will be maybe 1-3 titles that support raytracing as an add on, thats it...
if you look at how it went with ageia you know thats not enough...
intel is not stupid, they know that larabee has to perform well with RASTERIZING, and then they can slowly push the game developers to use raytracing which gives them an advantage. but they have to start with rasterizing!

For that, Intel will rely on LucidLogix's HYDRA Technology.

i dont really know what to think of this tbh...
of course its possible the way they explain it, but it DOES make you wonder why nobody else has done it so far.
cause the principles of how it works are known to everybody for years already...
the thing is that you need quite some computational resources to process all the data, and you add a delay that can cause one gpu to wait for the other which makes the whole streamlined process stutter and efficiency drops to the basement...

so until i see it running and reviewed and examined by independant sources i dont really trust in what they claim...

[Vozer]

If Intel wants to improve Larrabee's throughput against Nvidia and ATI's cards, why don't they just implement an even wider VPU, say, 32-wide VPU, instead of increasing the number of cores (64 cores! ) and with it, the die size, manufacturing cost and power consumption?

[Decami]

ok, but whats the point?
so in 2 years the 4870 will be faster than the gtx280 in the latest games, with the 4870 giving us 20fps and the gtx280 18fps... whats the point? :P

same with dx and framebuffer size...
who actually buys a low end dx10 card?
you can play dx10 games! NOT! :P
who actually buys a low end card with 1gb memory?
you can play future games and get better AA scaling!
yeah, 6fps instead of 5fps maybe :P

exactly, not even that, your being kind. We will NOT see Ray Tracing by then for shure, no way in hell. But we will have (hopefully) lucid hydra, which will (supposedly) have nearly 100% linear scaling for multi GPU configurations, and not using SLI or CF, and being able to handle any 2 or 3 or more cards of any type from one company so long as both are supported by the same driver. If indeed hydra makes its entrance into the market here soon, and AMD and Nvidia do everything to support it, (why wouldnt they?!) it might be a little tougher for Intel than they thought jumping in 2009-2010.

You may think "yeah, but Intel can make drivers to support hydra too" But keep this in mind and its an important one. There are not "old" Intel graphics cards that people have to run with Hydra. Most of the market buying into multi GPU already have a card and need another. Would you rather buy 2 cards for hydra and trash your old Nvidia/ATI card? Or would you keep that card and buy another for Hydra? get what im saying?

If indeed Intel comes in 2009 and hydra makes its predicted appearance of early 2009, Intel is gonna have major troubles their first year, major! Even if they wait it will be the same, they are gonna have it bad for a little while either way, if Hydra does for shure hit the market soon, and is everything its cracked up to be.

[shiznit93]

The first Larrabee might not be faster than NV/ATI gpu's of the time (it will be damn close though and offer additional advantages) but by Larrabee2 Intel will be on top. Their manufacturing tech and deep pockets are too much to handle.