Printable View
Precisely, APi is API nomatter what card. the game just sends DirectX commands to the API. And unlike ATI and nVidia. Update the driver. And your card now supports DX11, DX12, OpenGL 4.0, Something completely new standard etc etc. With the other 2 you need a new card each time.Quote:
Originally Posted by terrace215
Exellent article. Thanks.Quote:
Originally Posted by terrace215
Yeah, but AFAIK Windows isn't designed to use such thing well.Quote:
Originally Posted by FUGGER
It would have to know that CPU cores are fast and GPU - slow. It would need to know more or less how they are linked and how's their access to memory. NUMA support comes only with high end versions of Windows Server.
Also, no Windows supports more than 64 (logical) cores.
Sadly, OS can do nothing unless it's designed for supercomputers or for Larabee, only applications recompiled for it will get performance gain.
There's one more thing - why x86? It will only delay x64 market adoption...
It is X86-64. I don't even know if such a thing as X64 even exists, if it does then it will probably not resemble the x86 64-bit code base we now.
EDIT: I just read some more on this subject and well, the "x64" is also widely accepted so I apparently am bickering about nothing really. But you already have your answer, Larrabee is x86-64 compatible. x86-64 is actually just a superset to the x86 instruction set.
x86 has been here for over 30 years, quite much all existing software should be rewritten if we wanted something new. x86 has been expanding with different extensions like MMX, SSEs and 64-bit support etc.Quote:
Originally Posted by m^2
Everything looks good on papers... I want to see how this turns out.
Can this be counted as "revolutionary" if Intel succeeds in this ?
Well, MS uses "x64". Ubuntu uses "AMD64", and Fedora uses "x86_64". The exact term is x86_64 of course.Quote:
Originally Posted by Helmore
Let me guess, your main OS is Fedora? :p:Quote:
Originally Posted by alucasa
The exact term is indeed "x86-64", this is also the name that AMD used for it when it announced it and they are the creators after all. It also is, as far as I know, the only major addition to the x86 instruction set that was developed by AMD.
Nah, my main OS is Vista ultimate x64 as seen from my signature.Quote:
Originally Posted by Helmore
Anyway, x64 is now widely accepted as the term, there is no need to be picky about it.
i think nvidia is starting to sweat a bit now without having an x86 license. if larabee is competitive, intel and amd will have the graphics segment sewed up (as soon as amd makes an x86 compatable gpu).
this will put supercomputers in every household... well,at least what we consider supercomputers now anyways.
im pretty excited, but its still a bit off until we see some performance numbers.. i remeber before reading something about larabee performance vs. the g80 gpu... can't seem to find it though.
May someone more knowledgeable throw some light on this super larrabee chip for me. I think the chip is more of cpu than gpu and when and how does cpu beat gpu in graphic processing and physics rendering. If at the end it excels on certain cpu based task does that mean that we should end up shutting our eyes to an inferior graphic processor.
It is a multi-core CPU with a very wide vector unit. The OpenGL and DirectX support is added via a software renderer (which is rendered using the GPU of course, not touching the CPU).Quote:
Originally Posted by Shocker003
I find this solution very interesting, as DX and OGL are done in software, support for new versions of both can be easily added by a driver\library update without the need to replace the whole card!
Its a CPU alright. But its greatly enhanced and boosted to handle FP ops. 512Bit SIMD units, 4 threads per core etc. Each core is alot faster than even a Core 2 Duo at FP.Quote:
Originally Posted by Shocker003
Plus its 24-32cores or so and at 2Ghz. Quite far from even the fastest GPUs.
Thats how.
Raytracing also got alot closer...very close.
And no, it wont replace a CPU either. Its specialzied at FP mainly. Abit like the T1/T2 and Cell design ideas.
No, it's not like a CPU, but rather like a 64 processor supercomputer. And it's not meant to have inferior graphics performance.Quote:
May someone more knowledgeable throw some light on this super larrabee chip for me. I think the chip is more of cpu than gpu and when and how does cpu beat gpu in graphic processing and physics rendering. If at the end it excels on certain cpu based task does that mean that we should end up shutting our eyes to an inferior graphic processor.
Looks to be about right in line with rumored next gen GPU specs. Should be interesting.
It's not meant to be inferior of course, but there is still too little information available to make any performance estimates and where this would position it in the market against its rivals. Intel has the power to make a couple of different SKUs pretty easily as they can increase the core count (and some fixed functionality it has) to make a bigger part. The low end part would be 8 cores, mainstream 16, high-end 24 and enthusiast 32 for example, as long as it is dividable by 8. It's just that, there is no way for us to know where that 32-core part would stand against NVIDIA's competing chip, for now at least.Quote:
Originally Posted by m^2
Even Intel has not said anything about final core count for it's top performing chip, but 64-cores is probably a little too high for their first generation Larrabee and it will probably be something like 32 or 40 cores.
32 sounds low to me. Consider that each core might be slightly bigger than an Atom. Consider that Nehalem will be almost 800 million transistors on a single die, and that Larrabee will have to be much bigger than that to offer an okay compute capability as a GPU (also that they can fuse off underperforming cores/bits of cache and sell them off as lower-end SKUs). ATI has almost 1 billion today on TSMC's 55nm process and can't count the cash fast enough. Intel can easily afford a much more complex die on its 45nm process.Quote:
Originally Posted by Helmore
I think this '32 core' talk is just a bait and switch, like the supposed '480SP RV770'.
Each core is somewhat modified Pentium. It has increased caches (L1 is now 32K/32K instead of 8K/8K on the origina and it has a 256KB L2 partition in the L2 cache pool). Each core can do 4 threads at the same time. It's like SMT in the Nehalem, which emulates two cores, this thing emulates 4. The FP unit is of course a "super-FP" unit, it's very wide issue and can perform many ops at the same time.Quote:
Originally Posted by bowman
This architecture is radically different from nVidia and AMD\ATI, so you can't really compare them on "core count" basis.
indeed, anand got that pretty good:Quote:
Originally Posted by The Coolest
http://anandtech.com/cpuchipsets/int...spx?i=3367&p=4
If you count each ALU (as NV/ATI does) you get 512 "SPs" with the 32core version of larrabee.
-From Anantech's article
Sounds very nice.Quote:
No longer will we need new hardware to handle a new DX version with new techniques and effects: we would only need a driver update to add support for the new API. The only obstacle to running games using future APIs will be performance. The only reason to upgrade in the future will be speed. It will be a different world, altogether different than anything we've known or experienced before yet incredibly similar to the roots from which the industry was born.
From Anandtech
The vector ALU can behave as a 16-wide single precision ALU or an 8-wide double precision, although that doesn't necessarily translate into equivalent throughput (which Intel would not at this point clarify). Compared to ATI and NVIDIA, here's how Larrabee looks at a basic execution unit level:
http://images.anandtech.com/reviews/...comparison.jpg
NVIDIA's SPs work on a single operation, AMD's can work on five, and Larrabee's vector unit can work on sixteen. NVIDIA has a couple hundred of these SPs in its high end GPUs, AMD has 160 and Intel is expected to have anywhere from 16 - 32 of these cores in Larrabee. If NVIDIA is on the tons-of-simple-hardware end of the spectrum, Intel is on the exact opposite end of the scale.
We've already shown that AMD's architecture requires a lot of help from the compiler to properly schedule and maximize the utilization of its execution resources within one of its 5-wide SPs, with Larrabee the importance of the compiler is tremendous. Luckily for Larrabee, some of the best (if not the best) compilers are made by Intel. If anyone could get away with this sort of an architecture, it's Intel.
At the same time, while we don't have a full understanding of the details yet, we get the idea that Larrabee's vector unit is sort of a chameleon. From the information we have, these vector units could exectue atomic 16-wide ops for a single thread of a running program and can handle register swizzling across all 16 exectution units. This implies something very AMD like and wide. But it also looks like each of the 16 vector execution units, using the mask registers can branch independently (looking very much more like NVIDIA's solution).
We've already seen how AMD and NVIDIA architectural differences show distinct advantages and disadvantages against eachother in different games. If Intel is able to adapt the way the vector unit is used to suit specific situations, they could have something huge on their hands. Again, we don't have enough detail to tell what's going to happen, but things do look very interesting.
-------------------------
I'm taking this to mean each pentium core is equivalent to a stream processor to Nvidia and ATi. Anand states that Intel could have a 64 core 45nm chip the same size as the GT200. That's damn impressive. Add to that the native x86 core, gives game programmers flexibility and power like never before. They just have to find a way to harness all those cores. Of course there will be a layer to simulate directx, but in the future getting directly at those cores could potentially release a heck of a lot more power.
I'm hoping this will usher in a new age of gaming, since graphics haven't really moved on since dx9. What I've seen is more effects, motion blur, crappy shaders, film grain, etc. What gaming needs is for gpus power to grow geometrically. With the extensibility x86 cores give this could happen in the next few years.
That's not what I meant, I'm not completely dense. 32 of these cores would be very good today but it won't be here for over a year and by then RV970 and GT300 (400?) have rolled out and who knows what they're capable of in regards of programmability, 64-bit FP.. I'm just saying, 2TFLOPS 32-bit float sounds good today, but in over a year it'll be midrange. It won't be competitive in the high-end, they know this and surely they must be aiming higher.Quote:
Originally Posted by The Coolest
thats not the case for all things, u can program in drivers any shader operation that u want its just a matter of optimizing for it and some updates like dx10.1 use hardware and not the shader for capturing streams and soing AA in the frame buffer so u wont always beable to update to the new DX or OGL, u cant use drivers to gain an SSE set for a cpu but u may be able to emulate it with code even though it would be inefficientQuote:
Originally Posted by Shintai
August 12th - what a nice b-day present from Intel... looking forward to it:D
Its pure software renderer. So for Larabee thats the case.Quote:
Originally Posted by zanzabar
There is a few things. Larabee is basicly 512SPs at 2Ghz. Even RV870 if we assume 960SP would most likely be about 1Ghz. Ofcourse there is the X2 part here. Then there is efficiency. We already seen theoretical 300Gflops beat theoretical 500Gflops. Its simply not about that alone either.Quote:
Originally Posted by bowman
I would say Larabee in its first version will compete against RV870 and GT300 or whatever its named. with a max of 24 or 32 cores and a small die on Intels 45nm metal gate process.