I don't know the tech specs of it but it still strikes me as a rather inefficient thing to take a bunch of x86 CPUs and have them do graphics. I mean, that's why the graphics card was created, because it was designed specifically for graphics and did a much better job than a CPU
The Cardboard Master Crunch with us, the XS WCG team
Intel Core i7 2600k @ 4.5GHz, 16GB DDR3-1600, Radeon 7950 @ 1000/1250, Win 10 Pro x64
Soundcards where created for the same purpose, and now every mobo has a chip with the sound codecs on board and the cpu takes over the decoding. Its just a matter of calculation power.Originally Posted by SparkyJJO
I don't know the tech specs of it but it still strikes me as a rather inefficient thing to take a bunch of x86 CPUs and have them do graphics. I mean, that's why the graphics card was created, because it was designed specifically for graphics and did a much better job than a CPU
If your cpu would be fast enough, it wouldn't matter if x86 is inefficent. The problem, right now is, they arn't and thats why specialised hardware is faster (and always will be faster)
Also, intel also is eager to bring x86 to the embeded market. So they arn't focusing only on the high performance market but also on a mass market (set top boxes, VDR etc.).
Intel's Senior Graphics Software Architect Manager...
please, what do you expect him to say?
oh well, now that you ask... lrb is actually quite a weak design and doesnt perform well at all
what what WHAT???? ok... now im surprised... intel doesnt seem to take lrb half as serious as i thought they would...and possibly leading a small team of engineers through any of the above.
its about perf per transistor... intel wont be able to match that of a gpu for rasterization, so they need more transistors, means a bigger gpu... and considering how people crack jokes about gt300 being huge and a failure cause tsmcs 40nm process will probably never have good yields with such a massive chip... well think about a chip that is 20-50% bigger to reach the same performance... at a higher tdp... and now you might see why intel might not be able to capture the highend market...fail post.
if intel says larrabee will compete with radeon/geforce that means it will be low power all the way up to high end.intel designs all of their architectures to be scalable from atom to nehalem-ex so i dont see why they wouldnt do this for their graphics line up.
hahah, wow... intel must be praying that there are enough naive customers like you who buy a product cause they are convinced it must be good at something even if it sucks at what its targeted at
think of phhysix... ageia taught everybody a lesson in how naive many customers are nowadays and how much money they spend on a product that doesnt even have any real use
same as this killer nic card that speeds up your games but you cant meassure it LOL
i think they will wait for 32nm, at least for the highend version, and itll be 64 or even more... and there is no meassured perf so far, its all smart guesses from intel at this point afaik...
heh yeah... and not only that... i think intel is really taking this too light hearted and too arrogantly...I don't know the tech specs of it but it still strikes me as a rather inefficient thing to take a bunch of x86 CPUs and have them do graphics. I mean, that's why the graphics card was created, because it was designed specifically for graphics and did a much better job than a CPU
they want lrb to be:
cost competitive
tdp competitive
rasterization competitive
x86 competitive
anything else?
seriously... how arrogant do you have to be to think you can not only create a product that beats fixed function logic in its home territory, but also delivers outstanding general purpose performance, and all this at the same or worse mfc process, within the same tdp envelope and same price?
its like boing anouncing they will launch a new plane that can transport more people than an A380, fly faster than a concorde, and all that at the same price and fuel consumption as an avergae A330 passanger jet
oh and that not being enough, they make a, using their own words, SMALL, design team work on this... most of them have never worked with each other before and each comes from a diferent background, probably resulting in diferent views on things and conflicts...
i think intels top guys dont seem to realize what an important strategic value lrb has for the future of their company... small team...
ok, lets look at soundcards... while they heavily use the cpu to get stuff done, is there any mainboard, at all, that uses the cpu ONLY?
is there any mainboard that does audio completely in software on a cpu?
thats the same reason why many people doubt doing graphics on x86 cores entirely, makes no sense...
Last edited by saaya; 09-20-2009 at 07:57 PM.
Every Mobo with a Audio codec...?ok, lets look at soundcards... while they heavily use the cpu to get stuff done, is there any mainboard, at all, that uses the cpu ONLY?
is there any mainboard that does audio completely in software on a cpu?
The "Audio Codec" in hardware on the mobo is nothing more then a A&D/D&A converter.
The real decoding is done through software aka the CPU.
your right, bad example...Every Mobo with a Audio codec...?
The "Audio Codec" in hardware on the mobo is nothing more then a A&D/D&A converter.
The real decoding is done through software aka the CPU.
I know he's an Intel man and all, but, you know, I'd rather believe the project insiders than rumours from other sources.
.
here is a better idea of what larrabee is/is not.I don't know the tech specs of it but it still strikes me as a rather inefficient thing to take a bunch of x86 CPUs and have them do graphics. I mean, that's why the graphics card was created, because it was designed specifically for graphics and did a much better job than a CPU
nvidia has 240SP,AMD 160SP's, and intel 32.
you are on the right track ... at the end, when all DirectX 11 stuffs are adopted, it will end up to a race to instruction per clock ... because you will start having "special" processing for each pixels , and when doing so, you ll have to start reusing all the tricks of the CPU ... load units that don't have issue with alignment, fast branching ... SIMD and all the usual gadgets.here is a better idea of what larrabee is/is not.
nvidia has 240SP,AMD 160SP's, and intel 32.
by 2015 to 2020, you ll be back to CPU cores, that is my prediction.
(It is ok to dissagree, don't have to beat me up verbally :-P)
DrWho, The last of the time lords, setting up the Clock.
that picture is very misleading...here is a better idea of what larrabee is/is not.
nvidia has 240SP,AMD 160SP's, and intel 32.
here, this is much better:
Larrabee
GT200
RV770
somebody correct me if im wrong, but this is how i understood it from the siggraph papers:
LRB = 32+ "16way" processors, 2flops/clock
RV770 = 10 "16way" processors, 10flops/clock
GT200 = 30 "8way" processors, 3flops/clock
RV870 = 20 "16 way" processors, 20flops/clock
GT300 = 60(?) "8way"(?) processors, 6flops/clock
EDIT: corrected the infos
http://graphics.stanford.edu/~kayvon...g/diagrams.pdf
Last edited by saaya; 09-25-2009 at 08:41 AM.
You're mixing different things. 4way for LRB means 4 threads per core. It has 512-bit vector FPU, so each core can execute 16 32-bit ops per cycle. Considering FMAC capability it has in total 32 FLOP/cycle.
There are many more parameters to concidere, like the speed of branching, the speed of loading aligned and unaligned ... you are back to the instruction per clock race ... hehehehe ...that picture is very misleading...
here, this is much better:
Larrabee
GT200
RV770
somebody correct me if im wrong, but this is how i understood it from the siggraph papers:
LRB = 32+ "16way" processors
RV770 = 16 blocks of 10 "5way" processors = 160 "5way" processors
GT200 = 8 blocks of 3 8way processors = 24 "8way" processors
RV870 = 32 blocks of 10 "5 way" processors = 320 "5way" processors
GT300 = 16(?) blocks of 3(?) 8way(?) processors = 48 "8way" processors
so if you just look at processors, afaik lrb is 32+ (probably 40-48), gt300 probably 40-48 too and rv870 is a whooping 320. lrb and gt300 have fast and very beefy processors while rv870 has rather simple processors, clocked at only half of lrb and gt300, but makes up for it with the sheer number of them.
DX11 and other OpenCL will open the door to this again, IPC is the key ... actually IPC/watts.
check out here: http://www.realworldtech.com/page.cf...WT090909050230
you ll figure out that Silverthome (Atom) is as efficent as the RV700 , and it is without the 512 bits execution units of Lrb , and withoout its texture sampler...
Engineers understood where it is going, I wish that you guys put your fanboys hate in your pocket, and look at the technology itself.
Being able to compensiate for the overhead of x86 was the challenge, this graph shows that it is done.
now, the next challenge is to make Many of those cores x86 to work together ...
Then , it comes down to who is able to make the large transistor dices, and we all know who has the best fabs on the planete.
May the Core be with you!
Francois
Last edited by Drwho?; 09-25-2009 at 08:46 AM.
DrWho, The last of the time lords, setting up the Clock.
Posting Permissions
Reply With Quote
Bookmarks