who will win the next round of battles, AMD, Intel or NVIDIA?

[deadrats]

for those of you scratching your heads, i am talking about the new generation of processors set to be released in the first quarter of 2011, particularly as the "battle" relates to the primary focus of this web site; video editing and encoding.

"Sand Bridge":

this is intel's horse in this race, the first cpu to have an integrated gpu on die, as opposed to on the same package. the gpu portion shares the L3 and each gpu core has 6 "EU's" (similar to a cuda core) and some processors will have 2 gpu's. for a max of 12 "EU's".

on the cpu side, intel has been a bit tight lipped about the pipeline but we do know a couple of things:

1)it will support the new AVX instruction set, which is a 256 bit SIMD (single instruction, multiple data) instruction set, these instruction sets are targeted at high performance floating point math, suitable for HPC and many media applications:

http://en.wikipedia.org/wiki/Advanced_Vector_Extensions

2) it will have dedicated video transcoding hardware integrated onto the cpu. this speaks volumes about what a threat intel consider nvidia's cuda to be (more on that later).

3) the gpu and the cpu will be able to "turbo" independently of one another as needed, up to the max thermal ceiling.

4) just like now, some cpu's will have hyper-threading, presumably an "improved" version.

all this sounds great and "sandy bridge" seems to have a ton of potential, but we need to keep the following in mind:

1) no matter how much performance AVX may offer, programmers still need to a) know how to write code that exploits the capabilities and b) actually care enough about optimizing their code to write said code.

when you consider that SSE4 has been out wince the "penryn" days and despite benchmarks that show SSE4 offers substantial performance gains when used, i can count on one hand how many applications are available that actually make use of set instructions. there is no reason to believe that this will change once AVX enabled cpu's become available and a big reason for that is that programmers often will target their code for the lowest common denominator, with the current lowest denominator being either SSE2 or SSE3, with the most glaring exception being the people behind tmpg express, who seem to be one of the few that embrace the latest simd's as soon as they become available.

2) as good as the dedicated hardware transcoder sounds, a lot depends on whether it is transparent to the application or if the program has to be coded to exploit it. if it requires special programming to exploit it, then we are back to the above with AVX.

personally, i have trouble believing intel, considering that they promised a video transcoding driver, that would allow the core i3 to use its on die gpu for transcoding, over a year ago and it was supposed to be here by last february and we still haven't seen it, to me it seems like a politicians promises.

then we have the preliminary benchmarks, from reputable sites, that don't show the new sandy bridge cpu's being all that great:

http://www.anandtech.com/show/3871/the-sandy-bridge-preview-three-wins-in-a-row/8

across all video encoding benchmarks, it certainly doesn't look like anything to get excited about.

now it's true that there seems to be a ton of unexploited potential in the new cpu that may very well change all these benchmark standings, but my feeling is that sandy bridge will be like the really hot girl friend that you only get half way between 2nd and 3rd base with, who's parents and friends don't like you and are constantly trying to break up and who other guys are always trying to steal from you, basically you think she would be a really great lay and would make you a very happy man but you'll never find out because fate is against you.

basically i think it will be intel's next P4.

"Bulldozer":

amd's response to sandy bridge, this cpu's specs have morphed more than once since it was announced, the latest change being that it will not have an integrated gpu, this feature is saved for a future cpu, at present the generation after bulldozer.

what bulldozer does bring to the table is some no nonsense brute force:

1) like sandy bridge, it too will support AVX and it will finally support SSE4.

2) it will support SMT, albeit in a very different way than hyper-threading: each core features 2 integer cores, each capable of handling a single thread, so each core will be able to execute 2 integer threads simultaneously. for those that don't know, integer operations include logic evaluations, such as if/then/else, case/switch and true/false decisions as well as the obvious math calculations. as such, each core will appear as 2 cores to the OS, thus a quad core bulldozer will look like 8 physical cores to whatever Os you run.

3) the floating point unit is a beast. some time ago i posted that i wished someone would just make a dual core with lots of L2 cache and a beefy fpu, at least 256 bit but 512 bit would be preferable. it seems amd decided to answer my prayers somewhat, as each bulldozer core will have a single 256 bit fpu consisting of two 128 bit fmac units. these fmac's are backed up by by two MMX units. when you consider that non-AVX simd instructions are 128 bit instructions (a single instruction works on up to four 32 bit pieces of data or two 64 bit pieces of data) and default floating point instructions on most compilers are 32 bit (as is the associated data), it's reasonable to expect that a bulldozer cpu would be faster running code using the floating point code path rather than a simd code path, example software such as tmpg express or avidemux where you can choose what cpu optimizations you wish to enable. i can't see how a quad core bulldozer with four 256 bit fpu's chugging along wouldn't be faster than the same cpu using four 128 bit simds to perform the same task.

unfortunately this is all speculation as there is no bulldozer tests that i am aware of but amd has said that it expects bulldozer to be, on average, 50% faster than a comparably clocked phenom 2.

on the bright side, thanks to the approach taken by amd, with using beefy integer and floating point units, the performance should be evident with all existing software, regardless of optimization levels.

NVIDIA:

the x factor in all this. clearly intel considers them a threat, but there are a number of things to consider:

1) nvidia's cuda technology suffers from the same problems that AVX and other simd's do: it's not easy to write the code, many developers don't want to be bothered and developer's like to target the lowest common denominator from a hardware standpoint.

2) while there are some programs out there that make some use of cuda to provide gpu acceleration, many have suffered from bugs that caused crashes, bugs that caused poor image quality, lack of features and incomplete exploitation of what cuda offers.

on the other hand, nvidia has invested a ton of money in trying to get market penetration with their cuda technology, even going as far as to sponsor programming classes at major universities aimed at teaching programmers how to code for cuda and gpu's. unfortunately these classes are all graduate level meaning they are targeted at those that already have their comp sci degrees. couple that with the problems nvidia has had in the past with overheating gpu's and it's understandable that cuda hasn't reach the level of acceptance nvidia would like to see.

on the bright side, the promise of being able to spend less than $200 on an add-in card, install some drivers and get an incredible performance boost is appealing and cuda has seen good acceptance in hpc applications.

so who do you guys think will win?

[poisondeathray]

For editing, Nvidia has already won - at least in performance terms . Nothing comes close to CS5 + Cuda enabled Mercury Playback Engine and it works NOW. No onboard graphic/GPU solution will come close. Cuda has already been firmly established, and several years ahead in development. Even if some developers focus on onboard solutions or even OpenCL solution, it will be far far behind in performance and stability for years.

"Pure" Cuda encoders have improved, and the quality is acceptable for some circumstances, but for top quality & compression, software still is much better at this stage. I can't forsee that ever changing, but some parts of the software encode process can be accelerated (GPU decode, + some motion estimation)

For encoding, AVX won't help because current DCT based compression schemes do not use floating point . Maybe for some future codec or format

But what really matters, is who wins financially. How many people need to edit 8 or 9 HD streams in realtime ? For most people they don't care. I believe who can deliver the most affordable price points with sufficient performance (at adequate margins and yields) will "win". Hell, most people think "Youtube" delivers excellent quality. "performance" and "quality" are all relative terms. But Intel has the "deepest pockets" and can sustain and win any price wars.

[Alexstarfire]

nvm

[johns0]

Since its competition between the manufacturers thus giving us better products with better prices it's the consumer that wins.

[deadrats]

I assume you mean in GPU terms since nVidia doesn't make CPUs. It'll be like this: nVidia > AMD > Intel.

actually i meant from a video editing/encoding standpoint. while nvidia doesn't make cpu's and amd's bulldozer won't feature an integrated gpu, the fact remains that video is the "killer app" that makes the average do-it yourself'er upgrade. as such, intel, amd and nvidia are in a three way race.

[chowmein]

my feeling is that sandy bridge will be like the really hot girl friend that you only get half way between 2nd and 3rd base with, who's parents and friends don't like you and are constantly trying to break up and who other guys are always trying to steal from you, basically you think she would be a really great lay and would make you a very happy man but you'll never find out because fate is against you.

Love it LOL

[RabidDog]

I think.
Nvidia is set to become a (more) Niche player.. purely of Highend GFX and HPC turbochargers UNLESS they get some ARM A9 love and implement a CPU/GPU package like the other two contenders.
Intel's new "larrabee" is supposed to have adequate performance, equal to low/Mid end gfx cards, so it eats that market up.
AMd is left in the race as a Backstop non-monopoly stig. A Value play with lower perf but also low wattage.

Next wars?? Atom tri-core VS Apple A5 ..... Servers, NAS, Media Appliance, Tablets, Top-end Phones.

Lowest common denom, in future, will include VAX (HA!) so it will be used in software...

longer Term.. Intel will become One among Many, and I STILL wont have enough Fast reliable Storage

The 2012 Apocalypse has been cancelled due to lack of substance