im with you on this. how in "raw computing power" can it blow everything else away. even at 2.5ghz on 40nm, it would be only as powerful as 4870s/250s?Originally Posted by trinibwoy
im with you on this. how in "raw computing power" can it blow everything else away. even at 2.5ghz on 40nm, it would be only as powerful as 4870s/250s?
It still doesn't make sense. The raw computing power probably means the number of flops that thing can churn out and to state that it blows the socks off anything else would mean that it should do at least 3-4 TFLOPS, considering the 5800's. But the last heard that thing was going to be around 1 tflops and I can't see how they can quadruple that
INTEL Core i7 920 // ASUS P6T Deluxe V2 // OCZ 3G1600 6GB // POWERCOLOR HD5970 // Cooler Master HAF 932 // Thermalright Ultra 120 Extreme // SAMSUNG T260 26"Has anyone really been far even as decided to use even go want to do look more like?
Well depends on what kind of TFlops of larrabee has. If its 1TF DB (which the rumors from 2008 suggested), it blows everything out of the water, including HD5870 and G300.
DP to SP ration on Ati cards are 1:5 and 1:12 on NV GPUs.
If the ration stay the same Ati delivers with 2,78TFlops SP about 544 GFlops DP and NV would need a lot more then that.
So in terms of raw performance larrabee would be a badass even more if they can reach 2TFlop, but if they can translate that raw performance into actual performance is another thing.
Last edited by Hornet331; 09-21-2009 at 02:04 PM.
Inform yourself more then...really? as far as I'm aware Project Offset is still in the Project phase, and there's no game base on this Project that's announced, or shipped!
yeah it will blow everybody else out of the water in x86 perf... wow really? shame on you ati and nvidia! how come you cant beat them in x86?
but then what is lrb for? is it to offer massive glfops and steal away some of that MASSIVE MULTI BILLION DOLLAR gpgpu market?
thats what lrb seems to be a winner at, but yeah, whats it good for?
heres an interesting thought: intel has problems creating a gpu that is x86 compatible, nvidia cant do it cause they have no x86 license... only amd has both rasterization/gpu experience and expertise as well as x86 experience and expertise AND an x86 license... dont think itll help amd though, they are too slow and unefficient to make use of this advantage...
I think that this is more software problem then the hardware one. LRB has enough power to emulate rasterization and this is only matter of time and good programmers team when LRB will reach "acceptable" graphics performance. I do not know about all the pitfalls faced by developers but "acceptable" may mean performance level of 285GTX or with a big luck even performance level of ATI 5870 (may be I'm a bit optimistic here). Also, I suppose, the LRB competitiveness against AMD/Nvidia depends on the stage at which developing of 45nm LRB is.
how can a general purpose design have enough power to emulate a fixed function design using the same transistor and power budget?
if that would be the case there wouldnt be any fixed function hardware to begin with
and if you look at intels graphic driver history... its not exactly standing out as fast, efficient and boasting good compatibility...
they struggled to make things work at all in the past, and never had much time to tweak for perf...
and yeah, their lrb team might be much better... but i doubt it... they are a new team, never worked together, 0 experience as a team...
I suppose you didn't read the link I provided two pages ago.how can a general purpose design have enough power to emulate a fixed function design using the same transistor and power budget?
if that would be the case there wouldnt be any fixed function hardware to begin with
http://software.intel.com/en-us/arti...n-on-larrabee/
As Tom Forsyth puts it, because the whole chip is programmable, we can effectively bring more square millimeters to bear on any specific task as needed - up to and including the whole chip; in other words, the pipeline can dynamically reconfigure its processing resources as the rendering workload changes. If we get a heavy rasterization load, we can have all the cores working on it, if necessary; it wouldn't be the most efficient rasterizer per square millimeter, but it would be one heck of a lot of square millimeters of rasterizer, all doing what was most important at that moment, in contrast to a traditional graphics chip with a hardware rasterizer, where most of the circuitry would be idle when there was a heavy rasterization load.
This time is another story. Intel has much more experience with x86.and if you look at intels graphic driver history... its not exactly standing out as fast, efficient and boasting good compatibility...
they struggled to make things work at all in the past, and never had much time to tweak for perf...
and yeah, their lrb team might be much better... but i doubt it... they are a new team, never worked together, 0 experience as a team...
I know he's an Intel man and all, but, you know, I'd rather believe the project insiders than rumours from other sources.
.
oh please... how does that change anything? if resource usage would vary so much, and at one time a part of the gpu is barely used while another is used heavily, then dont you think ati and nvidia would have come up with some way of improving this? ever heard of pixel and vertex shaders getting merged to use the gpus resources more efficiently and assign resources dynamically to whatever needs to be done?
again, this is so arrogant... in essence hes saying ati and nvidia dont use their transistors efficiently and they can use the same transistors a lot more efficiently... and hes making claims saying rasterization performance IS great on lrb which is nothing short of a lie, cause when he said this they didnt have silicon AT ALL or only had broken silicon that definately didnt perform great... so he was probably referring to their simlation model... lol...
so what exactly do they think they can do more eficiently?
what part of the gpu is not being used in some scenarios?
they didnt mention anything...
its probably less work to optimize a game for lrb cause you dont have to take into account how many resources you got for what, you can use all resources for this or for that... but in reality, why would you go for massive geometry and no shading? or massive shading and very little geometry?
i believe lrb will def be very power hungry, will be priced very high to acceptable and will perform "good enough" in rasterization... but i doubt itll beat the topend gpus...
Last edited by saaya; 09-22-2009 at 01:58 AM.
I can't argue about resource usage in games (which, however vary from game to game) since I'm not a game developer. But if in some point of game all ROPs gets saturated it wont matter how much shader processors you have (it is only one of possible bottleneks). I think Michael Abrash (who is a game guru and dosn't work for Intel) makes sence here.
He dosn't say this.again, this is so arrogant... in essence hes saying ati and nvidia dont use their transistors efficiently and they can use the same transistors a lot more efficiently...
The article is from June 3, 2009 while LRB samples are humored from late 2008. But any way he didn't do any estimation of perf. relative to other GPUs.and hes making claims saying rasterization performance IS great on lrb which is nothing short of a lie, cause when he said this they didnt have silicon AT ALL or only had broken silicon that definately didnt perform great... so he was probably referring to their simlation model... lol...
Again - searching for bottleneks in current GPUs wasn't a target of this article.so what exactly do they think they can do more eficiently?
what part of the gpu is not being used in some scenarios?
they didnt mention anything...
With my knowlege in games, for me it is a rethoric question.its probably less work to optimize a game for lrb cause you dont have to take into account how many resources you got for what, you can use all resources for this or for that... but in reality, why would you go for massive geometry and no shading? or massive shading and very little geometry?
well yes, he does say ati and nvidia use their transistors inefficiently, cause they dont load balance like lrb does...
hes basically arguing that the fixed function benefit is lost and actually makes the overall design slower than general purpose logic, cause it cant be balanced dynamically to perform diferent tasks... what tasks? for gpgpu this is definately true, but for games?
he didnt make any perf estimates compared to gpus? then what is he refering to with this statement please if not perf vs gpus?
and that quote is wrong as well cause there is no way you can use the whole chip for one task... and im not even going into picky stuff like saying the crossbar and memory controller and other io... how about the display engines? how about the video encoders and decoders? how about the display engine? how about the texture units?because (LRB) is programmable, we can effectively bring more square millimeters (than a gpu) to bear on any specific tasks as needed - up to and including the whole chip
lrb already went from entirely general purpose to partly fixed function, and nobody is denying that lrb3 will be even more fixed function...
why would they do that if general purpose is sooo awesome and allows so much more flxibility which results in better resource management and hence more perf? intel making lrb more fixed function with every revision is the best way to show how much truth that comment contains and how much is just a bold arrogant claim...
Last edited by saaya; 09-22-2009 at 07:00 AM.
What he is explicitly saying is that current GPU programming is limited to executing on the SIMD Cores in said GPU, which as of current only occupy 62% of the Die, where as Intel's Design has no vestigial parts in regards to graphics processing. [ATi and nVidia don't dare release a graphics card that doesn't support the common 3D primitives]
However one could argue that those 3D primitives are analogous to Intel's x86 legacy instructions; since they both must be supported and as transistor counts grow the cost decreases.
The major difference is that the cost is relatively fixed for the entire GPU regardless the number of SIMD cores, where as Intel has to include legacy support with each and every single x86 core.
Fast computers breed slow, lazy programmers
The price of reliability is the pursuit of the utmost simplicity. It is a price which the very rich find most hard to pay.
http://www.lighterra.com/papers/modernmicroprocessors/
Modern Ram, makes an old overclocker miss BH-5 and the fun it was
and you're... Sh1tyMcGee's lawyer?
please feel free to point out to one shipping title, or announced title with this Project of engine?
Last edited by Nedjo; 09-22-2009 at 03:23 PM.
Adobe is working on Flash Player support for 64-bit platforms as part of our ongoing commitment to the cross-platform compatibility of Flash Player. We expect to provide native support for 64-bit platforms in an upcoming release of Flash Player following the release of Flash Player 10.1.
Ressource usage can change very extremely from game to game, look at af in hawx. Ati looses a lot with af because the tmus are limiting. But the problem is it isn't easy to improve this thing. But there are people sayin, we'll see improvements in the not so far away future
saaya, maybe the difference will be made up by better IPC and something about LRB being 2ghz? instead of 850mhz like todays gpus
Intel 2600K @ 4.8ghz 1.31v on Water.
ASROCK Z68 Ex4 Gen 3, 16GB G.skill pc1600
MSI GTX 680 1200/6800mhz
2x Vertex LE 60GB Raid 0
unoid, clockspeed and ipc are just other words for transistor performance...
i believe intel can get better perf per transistor than tsmc, probably, but not that much more...
i just read some 5870/5850 reviews...
they doubled the resources of their previous chip and only got a 20-40% performance boost on average... ouch... what happened?
sounds like they are hitting a perf wall somehow?
if you can get 75% the perf with half the transistors of todays highend cards... then lrb really shouldnt have a problem performing well in rasterization...
is this some sort of rasterization limitation? whats holding the perf back?
i thought graphics was THE example of multithreaded performance scaling where doubling the resources always gets you close to double the performance?
you didn't actually read the reviews, thats what happened...i just read some 5870/5850 reviews...
they doubled the resources of their previous chip and only got a 20-40% performance boost on average... ouch... what happened?
Intel demo'd Larrabee at IDF for those who missed it.
Last edited by zalbard; 09-23-2009 at 05:03 AM.
Donate to XS forums
IDF 2009 - World's First Larrabee Demo, More Clarkdale, Gulftown:
http://www.anandtech.com/cpuchipsets...oc.aspx?i=3645
Photos from the demo:
Larrabee die:
j/k about the die
Last edited by Piotrsama; 09-23-2009 at 03:37 AM.
A pretty poor demo really. Not exactly a photorealistic scene, and probably rendered with a progressive ray tracer so that it doesn't have to complete the whole frame before moving to the next. I would have been more impressed if they moved the camera, as that would clearly show if a fairly limited number of rays were being traced at any one time. Throw some supersampling in there and it would have crapped itself too.
Larrabee die:
j/k about the die
Did you read the text below the picture? It is not Larrabee die. That's what a single Core i7 would look like on the 386's manufacturing technology.
Intel QX9650 @ 3.6GHz 1.25v in bios
Asus P5E64 WS Evolution
Crucial Ballistix DDR3-1600 7-7-7-20-1T 4X1Gb 1.76v
EVGA 8800GTS 512Mb
Antec P180b with Corsair 620HX
Did you read what I wrote under the picture?Did you read the text below the picture? It is not Larrabee die. That's what a single Core i7 would look like on the 386's manufacturing technology.
It was meant as a joke.
Larrabee Fails To Impress at Tomshardware.
Was Gulftown doing all the heavy lifting here?The platform was a 32nm Gulftown (six-core) CPU. The demo was an almost real-time ray tracing demo based on Enemy Territory: Quake Wars. The demo has been shown in the past, although this particular iteration was (supposedly) running on actual Larrabee hardware. It looked to be running at maybe ten frames per second.
Drawing any conclusions would be premature, though. We really need to see performance on DirectX or OpenGL games. But it’s certainly an inauspicious demo, and if anything was clear from the demo, Larrabee still has a long ways to go.
yeah i read several now and got a better idea of perf... slightly slower than 4870x2 for a slightly higher price... makes sense for people who have 30" displays...you didn't actually read the reviews, thats what happened...
5850 is great for people who dont have a 4890 or gtx285 now... i think its a good upgrade from a 4850 and gts250 and below...
hahahah good one!Larrabee die:
j/k about the die
yes thats what i wondered as well... why did they stick lrb in a box with a 6core 12 thread cpu? could be that it was partially done by the cpu actually...
and seriously, even if its done on lrb only... this is raytracing, lrbs home territory... and it can only pull around 10fps? and rasterization perf is supposed to be on par with gtx285 and more?
Posting Permissions
Reply With Quote
Bookmarks