Intel: CUDA will be just a 'footnote' in computing history

What's going on? I post once and get two copies -whatever. - Rob Lucien

Comment by lucien86 at 10:01pm 5th July 2008

I love this suggestion but let me pour a little cold water on Intel's new chip. Flexibility verses Context (dynamic verses static) is one of the oldest saws in computing, the computer scientists didn't solve it in decades of work and if Intel have really solved it now I'll eat my hat. As for the Cell being dismissed, learning new architectures is difficult but its part of the world. the high performance computing people are taking the Cell up, the fastest computer isn't using Intel any more its using Cells. Everyone knows that multi-core is where its going (4, - 8, -10 -20, etc) and the Cell is currently one of the best. As for CUDA, anything that can extract that enormous processing power from GPU's has got to be good. 256 stream processors is probably unbeatable by any standard CPU we are likely to see in 10 years. The real question is still Microsoft, does Windows continue to be x86 only or do they take a jump. A lot of the work has already been done - the Xbox 360 is running on a version of Direct x on a multi-core Power PC architecture. The Xbox architecture is positioned to switch straight to the Cell, the question we should all ask is why wasn't it Intel based? even Microsoft are wary of sharks.

Comment by lucien86 at 3:10pm 5th July 2008

I love this suggestion but let me pour a little cold water on Intel's new chip. Flexibility verses Context (dynamic verses static) is one of the oldest saws in computing, the computer scientists didn't solve it in decades of work and if Intel have really solved it now I'll eat my hat. As for the Cell being dismissed, learning new architectures is difficult but its part of the world. the high performance computing people are taking the Cell up, the fastest computer isn't using Intel any more its using Cells. Everyone knows that multi-core is where its going (4, - 8, -10 -20, etc) and the Cell is currently one of the best. As for CUDA, anything that can extract that enormous processing power from GPU's has got to be good. 256 stream processors is probably unbeatable by any standard CPU we are likely to see in 10 years. The real question is still Microsoft, does Windows continue to be x86 only or do they take a jump. A lot of the work has already been done - the Xbox 360 is running on a version of Direct x on a multi-core Power PC architecture. The Xbox architecture is positioned to switch straight to the Cell, the question we should all ask is why wasn't it Intel based? even Microsoft are wary of sharks.

Comment by lucien86 at 3:10pm 5th July 2008

I recently ported a financial algorithm on CUDA using 8800 GTX. It ran 125 times faster than AMD Athlon 2.41GHz. Technically speaking: GTX has 128 cores running @ 1.35GHz. If we just equate the cores in terms of the speed the speedup should be around 128*1.35/2.41 ~ 70. But the fact that CUDA got it around 120 proves that there is lot of latency hiding that CUDA provides because of its threading and execution model! Thats the basis of performance in the CUDA world!! I am quite skeptic of whether Larabee - touted as general purpose multi-core - would scale this much hieght!! Anyway, Lets wait and see! Much depends on what applications would use these powerful machines to create splendid effect for users!- Adobe is already rumored to be looking @ CUDA for Fotoshop CS4. If more and more apps get benefited out this, CUDA could eventually become the winner!

Comment by Sarnath at 9:48am 4th July 2008

No matter what parallel processing technology you use -- programmers have to get accustomed with parallel programming! Whether it is CUDA or INTEL -- does not matter!! Mr.Gelsinger talks about usage of Intel cores!! Ha... does it really matter? CUDA is not an assembly language. It is a C extension -- dats all. What does the programmer care what cores runs under? For optimizing your performance -- you need to optimize your algorithm and not machine instructions! -- in 99% cases! I have programmed in CUDA. Its quite cool! Parallel programs are hard to write. I dont see how Larrabee is gonna change it.

Comment by Sarnath at 5:37am 4th July 2008

its funny to see everyone discussing this as though its gona be the next gen stuff... just think... If the developers don't share the technology with other devs/manufacturers, then game developers wouldn't touch it, because not every gamer could play their game then... plus that would meen big losses for the game devs! - so if intel/nvidia dont share then they will have to just wait until it hits them that they had just spent millions researching something that wouldn't even be used! :S

Comment by NikoBellic at 7:56pm 2nd July 2008

I have coded for CUDA and have a good understanding of how it will compare to Larrabee. CUDA is great for very limited set of tasks. Video compression is a good example of one such task. Think of CUDA as an army that marches in lock step. Each soldier doing exactly what every other soldier is doing. If just one soldier in a squad needs to do something different, the entire squad has to go through the motions while it waits for him. In Larrabee, each soldier can do something different. This means that Larrabee is far more flexible and can be used to make almost any computer task faster. An example is ray tracing where every ray can go in a different direction and hit different objects in a scene. In Larrabee, the speed of ray tracing is multiplied by the number of CPU cores on the chip. With CUDA you get very little peformance increase even if you have a GPU with 200 cores. NVidia is absolutely correct when they say that their GPUs will be able to vastly outperform Intel's Larrabee processor for games like Crysis. Intel is absolutely correct when they say that Larrabee will be able to peform real time ray tracing vastly better than NVidia's GPUs can. Larrabees problem is whether it can survive int the market as a both a low performance GPU and a super high performance massively parallel CPU.

Comment by GrangerFX at 5:07pm 2nd July 2008

The 4870 does seem to perform better (marginally) in some games, however you have to look at what these are and why. It tends to be in older games where the balance of information is more suited to its architecture. The 280 (and 260) have been developed to provide processing muscle where it is needed for the latest and future iteration of games. In these games it tends to perform better (unsuprisingly). More importantly, is you look at the tets, the 280 and 260 show much closer values for the average and lowest framerates, providing a smoother gameing experience... personally I'd quite happily swap higher frame rates (which you dont notice in game anyway, unless you always play with FRAPS running) and have a more consistent performance. I'm not currently buying either at the moment, I'll wait and see what the future tells, but my money, if I had it to burn, would be on the 280 architecture being more future proof for the next wave of games.

Comment by NewParadigm at 4:56pm 2nd July 2008

It's funny when you don't get to see these mysterious labtests where the new 4XXX series cards managed to cure cancer and end world poverty. I do wish people would think before they opened their mouths.

Comment by crazyceo at 2:41pm 2nd July 2008

If Intel need to focus on what it does best. CPU's. It's good they are expanding but could this spell the end for AMD and ATI.

Comment by tunedude at 2:21pm 2nd July 2008

Except when they dont... the 4870 has shown comparable, or higher, performance to the GTX260 in a number of reviews. In a couple of tests it has even out performed the GTX280 (this is rare, but does happen in some games), whilst costing half as much!

Comment by NotFred at 12:51pm 2nd July 2008

I think you've become confused with the comparisons of the 4870 and the 9800 GTX+ and not the GTX 200 series you mentioned. The GTX 200 series wipes the floor with the 4XXX series cards on every level. This is because it's aimed at the high end of the market and not the mid range the 4XXX sets itself in. It's like matching a 350Z with a Micra.

Comment by crazyceo at 9:40am 2nd July 2008

From Tomshardware... "With the preview version, unfortunately compatible only with the GT 200, we were able to compress our test video (400 MB) in iPhone format (640*365) at maximum quality in 56.5 seconds on the GTX 260 and 49 seconds on the GTX 280 (15% faster). For comparison purposes, the iTunes H.264 encoder took eight minutes using the CPU". I think intel are a little worried more than anything.

Comment by gavomatic57 at 8:30am 2nd July 2008

I thinnk Intel Do take CUDA seriously as CUDA is useful in renderfarms, for high-end 3D & video editing in the movie industry etc. And CUDA is basically an array of GPU's acting as mini CPU's. LarryLadybird is kinda the same in reverse, 16 mini Intel CPU's pretending to be GPU's. However CUDA can be adopted professionally, as nVidia profesional cards do kick ass and even using consumer cards gives good results. But it's wierd...CUDA isn't designed for gaming or any other general consumer use, and here we have Intel using their (assumingly) consumer graphics card, and comparing it to something non consumer. Plus by the time LarryLadybird gets released, more people will've adopted CUDA and put it to use, and be very happy with it, so they won't be in a rush to run out and spend more money on a new mini-supercomputer. I think what I-Robot was trying to say was: "Blah blah blah....blah...blaaahhh...CUDA...blah blah blah...blaahh, here's what our architecture will look like...blah blah

Comment by EdArch at 12:10am 2nd July 2008

larrabee 16 cores in 2009. Nvidia 256 cores in 2008 and more in 2009 guess what will be faster and what developers are developing for now. Intel, just too little and just too late. Better start getting a license from nvidia...

Comment by anhe64 at 11:45pm 1st July 2008

Just 2 months ago one of the representives in nvidia said AMD was in big trouble and presents no threat to nvidia boarding collapse, then we have 4870 totally whipping nvidia's new 200 series both in value and performance. Its common in the tech market place to bad mouth and down play rivial's tech, I could imagin intel take CUDA very seriously behind its walls.

Comment by Wolfy at 11:30pm 1st July 2008

moving on... I personally would hate to see Intel dominate the GPU world too... Just look at how long we've been putting up with similar speed CPUs, I just really wanna see techs performance going faster than a buggati veyron can get from 0-5Mph!! lol

Comment by NikoBellic at 11:14pm 1st July 2008

I honestly can't think of a single historical precedent for programmers picking up a new language for any reason... honestly its true... honest... ok I'm lying

Comment by NewParadigm at 11:18pm 1st July 2008

wasn't the larabee demo recently featured on here running on 8 or 16 cores or something like that? Exactly how expensive is this going to be!

Comment by NewParadigm at 11:14pm 1st July 2008

Just look at this guys photo attached to the article........I bet he's still a virgin! Hang on, I'm still a virgin too.......at 35! Damn you PC Geeks! I'm off to see what the lovely sweet internet can satisfy my virgin urges with...........

Comment by crazyceo at 10:53pm 1st July 2008

Nvidia is a huge monster inc. that has it's fingers in lots of pies. I don't think for one minute they aren't prepared for statements from Intel like this. Do you really think that Intel have all the answers to GPU's? If that were true then why would they have held it back until now? Nvidia or ATi shouldn't be worried by these comments as it's just the marketing machine of Intel running wild again. They should stick to the CPU market which they pretty much dominate hands down at the moment. Larrabee is a great idea but let's wait and see the benchmarks and reviews before we start planning are Intel only systems!

Comment by crazyceo at 10:49pm 1st July 2008

Why is this "Genus Epiphany" happening now in the year 2008? How about Intel having thought about this a decade ago or more...

Comment by toycityworld at 10:36pm 1st July 2008

...Larrabee isn't going to come out until late 2009/early 2010 at the earliest. That's way too far away to seriously consider, who knows what will have happened on the nvidia/ati side by then?

Comment by aceizace at 10:45pm 1st July 2008

well,we all know what happened with those dinosaurs... http://fastra.ua.ac.be/en/technicalfaq.html

Comment by dreams at 9:36pm 1st July 2008

larrabee wont replace graphics cards, because I think games producers will be unwilling to switch to a new programming language, and there will be no guarantee of compatibility with old games, and new games wont work at all on old computers. developers wont make games that use larrabee until there is a large amount of people with a larrabee pc, and the majority of people wont buy larrabee until there are games which can make use of all that power. Also, CUDA wont replace the CPU for the reasons that intel say, and it will also be difficult to move certain tasks from the CPU to the GPU without major redesigns of the GPUs architecture. But CUDA will still play a very important role in many applications requiring a very large amount of processing power, such as folding or video encoding.

Comment by Cogwulf at 8:27pm 1st July 2008

The other thing is... x86 is universal. Write for x86 and every desktop can run it. Write for CUDA and about 30% of PCs can run it. (once you remember most PCs have Intel integrated graphics)

Comment by NotFred at 6:56pm 1st July 2008

It does make perfect sense if it is true that CUDA adds more effort then what the output is, or is just too complicated to work with. I guess we can either take Intel's word for it or nVidia both are likely to slightly bend the truth in their favour. I would have thought nVidia knew the software houses pretty well and as long they make CUDA accessible enough, it should stand a chance. I guess it is like all warfare, you will always have to take both sides propaganda with a shovel of salt

Comment by jonisaksson at 6:49pm 1st July 2008

In some ways its sad, but they are right. Also, why would Intel want to work with Nvidia? Nvidia get a FSB liscence, then once they have it remove SLI from Intel chipsets, they take every opportunity to claim the CPU is redundant (and make themselves look rather stupid in the process). If I were Intel I would hate Nvidia.

Comment by NotFred at 6:50pm 1st July 2008

take what you know, make it more efficent and better, instead of starting all over again.

Comment by WoodSpoon at 6:38pm 1st July 2008

Unfortunately I have to agree with Intel, few people don't care if the new architecture is 10 better if it requires 100x more effort.

Comment by DudQuitter at 6:29pm 1st July 2008

Yeah I have known about the goings on for a while lol... just summin it up in a short sentence!

Comment by NikoBellic at 5:55pm 1st July 2008

I would of thought that intel would be helping Nvidia to take out ATI not vice versa... Shoking!!

Comment by NikoBellic at 5:51pm 1st July 2008

• One top four motherboard maker will die, says MSI
• AMD plans supercomputer with 1,000 GPUs
• Nvidia announces GeForce GTX 285
• Creative takes wraps off Zii processor architecture
• Nvidia announces stereoscopic 3D glasses
• BFG unveils hand built Phobos gaming PCs
• UPDATED: LiveJournal announces mass layoffs
• Top 10 PC games chart
• Seagate Barracuda 7200.12 gets 500GB on single platter
• Creative reveals new Zii processor

These are the most popular news articles over the past seven days.

• Phenom II overclocked to 6GHz
• GeIL unveils Cyclone RAM cooler
• Rumour control: GeForce GTX 260 GX2 in January
• Nvidia cuts price of 216 SP GeForce GTX 260
• VIDEO: PC version of Mirror’s Edge features hardware PhysX
• Get Half-Life for less than a dollar on Steam
• IBM plans to build virtual human brain
• RivaTuner 2.20 adds SLI and CrossFire support
• VIDEO: ArmA II trailer released