Intel: GPUs Only 14x Faster Than CPUs

[sonofliberty08]

"Intel Architecture programming model" = old school programming model
December 2012 = the end of Intel old school architecture

 

[Tamz_msc]

Most people do not use scientific applications, and most programs are not optimized to use more than two cores.Very few people can utilize parallel processing on the GPU.CPUs are way more complex and and most software is coded for utilizing the CPU, and it does a perfect job in handling day to day applications that users NEED.Parallel processing through CUDA is a WANT that is suitable for only a handful of people.NVIDIA has to realize that a person sitting in an office or at home doing his work does not need something that won't benefit him in real world situations.

 

[blackjellognomes]

[citation][nom]ivan_chess[/nom]"-the average CPU only has six cores"
Average? I'm pretty sure most people don't have six cores.[/citation]
[citation][nom]gerand[/nom]"the average CPU only has six cores"
*sighs* Rich, privileged people... surrounded by the latest tech like a fish surrounded by water. Last time I checked, the only commonly available hexa-core processors were AMD's $300 45nm 1090T and Intel's $999 32nm i7 980X.
I have yet to see a system that has either of those in real life, especially since another article on Tom's reported an alleged supply shortage of 1090Ts.[/citation]
Really, guys? I realize most of you only visit this site for the gaming tech, but this is a hardware site including various kinds of tech. You're forgetting that laboratories and such have supercomputers with CPUs containing far more cores than any consumer CPU. In early 2007, Intel announced an 80-core research CPU, and there's no doubt that many similar CPUs have been produced by Intel and other companies since (and probably the past two decades prior). Outliers such as this chip have a strong influence on the somewhat misleading "average."

 

[zodiacfml]

this is just Intel buying time to catch-up (Larrabee) as Nvidia and AMD is well ahead in this area so that Intel can compete once GPU computing becomes more mainstream.

the benefits are undeniable in terms accelerating some functions in a power and cost efficient way. that excludes the scientific apps as my idea is limited to current uses such HD video processing decoding/encoding, Photoshop, and 3D.

read AMD's whitepaper on their Fusion technology.

 

[Guest]

Intel claims of 14x sound perfectly reasonable imo ... Think about it, 240 cores @ ~700 Mhz vs 4 cores @ 3+ Ghz

If a GPU core is as powerful as a CPU core at the same frequency, which I highly doubt, and assuming the CPU's frequency is 4 times higher than the GPU's frequency, we can reach the conclusion that one GPU core is 4 times slower than a CPU core. Since there are 60 times more cores on the GPU, simple math tells us that the extra speed gained would be ~14-15x, but only if there is an application that is perfectly scalable on many cores.

So, even Intel's result is painting a too nice picture for Nvidia's GPU imo, since normally a CPU core is far more complex than a GPU core and faster Mhz wise imo, and also perfect paralelism is an utopia...

 

[rohitbaran]

I can only say, techno-corporate bullshit.

 

[teodoreh]

AMD should exhanchge x86-64 license with CUDA. That would f..ck Intel for good! 😀 😀 😀

 

[zmbcat]

Uh some of you forgot that this article mostly talks about server/workstation computers, that use the multicore cpu's and such...not the average user stuff.

 

[dEAne]

If I were nVidia and intel they work together for much cooler and faster chips.

 

[infodan]

Intel are probably hoping pwople misread that as "Only 14%"

 

[shortbus25]

Where is "ATI" in all of this? The only company I know of that is part of a CPU company "AMD" if anyone had a good chance of making a cpu with a lot of processing units they have a good shot. Now Nvidia is making headway with Cuda but of course as the average gamer/user goes what else do we need Cuda for? We just want to play, and play as fast as possible! Why cant they just make a simple processor with gpu cores in it for lets say a Intel socket or AMD socket and wattage who cares if its 300 watts max if it can handle it push it to it. When they start making a CPU with GPU insides "(GUTS)" then we will see things in a hole new light. CPU that can handle complex instructions and parallel processing Graphics on CPU no need for dedicated graphics unless you want to add lets says 4 GPU's in the box for better and faster gameplay. When they finally realize what they could do then we will all have super computers but for now most every average user with their dual and quad processors will have to just dream of having 16x-100X cores on one CPU.

 

[Guest]

I have programmed with CUDA and achieved a 50x, BUT programming is much harder and limited. Usual speed-up for projects I know is 10x, and only for math-heavy parallelizable applications.

 

[Computer_Lots]

"The average CPU only has six cores"
Seriously..... Maybe at Toms where you guys get the latest of everything but I just upgraded to 2 cores this year.

 

[nottheking]

At least I'm not the FIRST person to get in here to point out the differences in architecture between a GPU and a CPU. GPUs have lots of "Dumb" math power, and very little flexibility, to put it in simple terms. They also only natively do single-precision math, while CPUs can natively handle double-precision. The latter is a major point of difference; major math benchmarks for supercomputers like LINPACK are double-precision, hence modern GPUs can't actually reach the multi-teraflop figures they tout were they to try running them.

[citation][nom]shortbus25[/nom]When they start making a CPU with GPU insides "(GUTS)" then we will see things in a hole new light. CPU that can handle complex instructions and parallel processing Graphics on CPU no need for dedicated graphics[/citation]
Some problems with that: CPUs can handle more complexity, as you put it, due to their much greater physical complexity; branch predictors, schedulers, virtualizers, etc. are all built-in features. By contrast, a GPU can have so many sub-units because they're so tiny and simplified; they do NOT have all those features. So unless the GPU sub-units are handling "embarassingly parallel" problems, the bulk of them will sit idle while a handful of them will actually be working.

And of course, to GET all that ability to efficiently use all the cores for non-parallel tasks, you need all that extra, complex circuitry that, coincidentally, bulks your processing units up to the size you see with CPUs. If there was some "magical secret" that allowed Intel to make a CPU that had countless cores with all those features on each, you can rest assured they'd have made it, and crushed all competition.

 

[shortbus25]

Yeah but they would charge a $10000000000000 bucks for it and no average person like the ones here could have it. Intel has more up there sleeve then we know about, they know what there doin so does AMD but can we advance that fast no for we can not afford to much change at one time. Nor do I want to spend $1000's on processors that in 3 years will be discontinued and replace by newer better model's. Besides the technology keeps getting smaller so before long even putting a CPU and GPU in the same chip will be possible,despite heating issues they will figure it out.

 

[kronos_cornelius]

I think the research is too close to home for Intel for them not to be accused of conflicts of interest. Say the discovered in their tests that Nvidia's chip was in fact 100 times faster. What would the do then.

 

[pyr8t]

[citation][nom]sonofliberty08[/nom]"Intel Architecture programming model" = old school programming model December 2012 = the end of Intel old school architecture[/citation]

The Mayan's say it will be the end of new school architecture as well...

My thought, until CUDA runs without a CPU present, they cannot be compared.

I would like to see more development along the lines of computing done with 2 processors, 1 serial-optimized, 1 parallel optimized. You will still need the serial-optimized CPU to provide guidance to the parallel processor in any case.

 

[Guest]

If there's one thing that makes me near auto-reject a research paper it's a 100x speedup on the GPU. It'd better have some very very good explanation for why they do that. If it's at all unclear I assume it's based on a poor CPU implementation. That is the way I've been approaching GPU paper reviews for a while now.

This 240 core claim is an outright lie, too. I was amazed when I heard the following at an academic conference: "How do you expect Larrabee to compete with nvidia's cards when it only has 32 cores compared with hundreds?"

 

[nottheking]

[citation][nom]pyr8t[/nom]I would like to see more development along the lines of computing done with 2 processors, 1 serial-optimized, 1 parallel optimized. You will still need the serial-optimized CPU to provide guidance to the parallel processor in any case.[/citation]
If I understand what you're trying to say, AMD's actually pursuing something along those lines with their "Bulldozer" core. They've been a bit silent on the exact specifics of it, but it is known that it will pair Phenom II-based cores with stream processors similar to what're seen in the AMD Evergreen GPUs. Such a design would give a selection of both conventional CPU processing resources, along with a lot of power that could be used for more strongly parallel tasks. (i.e, tasks where the CPU's extensive complexity would be wasted)

[citation][nom]ALongerUsernameThatProbablyIsntTake[/nom]This 240 core claim is an outright lie, too. I was amazed when I heard the following at an academic conference: "How do you expect Larrabee to compete with nvidia's cards when it only has 32 cores compared with hundreds?"[/citation]
Well, the use of "cores" can be a bit hazy for a term, as the architectures can sometimes blur these distinctions. Normally, when it comes to parallelism there's a distinct hierarchy, from lowest to highest, in terms of solutions:
-Word length (bit-level parallelism)
-Vector (SIMD)
-Superscalar (multi-pipeline; instruction-level parallelism)
-"multi-core"
-SMP (multi-CPU)
The problem is that the definition of "multi-core" is often a bit arbitrary, as shown by the evolution of CPU designs. This solution can be achieved by simply using multiple CPU dies, or alternatively, through a design that removes many components (like the memory interface) that might otherwise be redundant. In the end, the line between "superscalar" designs and "multi-core" designs can be a bit blurred, to where some reference of "cores" can be at the very least forgiven. This would hold especially true in Larrabee's case, since its "cores" aren't exactly fully-fledged; a lot of the control circuitry only exists on the total chip level, that is further passed down into the individual elements that Intel refers to as "cores."

At any rate, that last question you quote still raises a valid point, that does have a valid answer; as the processing elements CAN be compared and contrasted, it's a fair statement. Of course, the counter-point is that the SIMD units in each "CUDA core" is only a total of 128 bits (4x32) wide, contrast to 512 bits for Larrabee. But in terms of design describing both GPUs' elements as "cores" isn't really more incorrect for one than the other.