Sunday, July 6th 2008
Intel Larrabee Capable of 2 TFLOPs
German tech-journal Heise caught up with Intel's Pat Gelsinger for an article discussing the company's past and future as the silicon giant heads towards 40 years of service this 18th of July.
Among several topics, came up the most interesting one, visual computing and Intel's plans on it. 'Larrabee' strikes as a buzzword. It is the codename of Intel's upcoming graphics processor (GPU) architecture with which it plans to take on established players such as NVIDIA and AMD among others.
What's unique (so far) about Larrabee is that it's entirely made up of x86 processing cores. The Larrabee is likely to have 32 x86 processing cores. Here's a surprise: These processing cores are based on the design of Pentuim P54C, a 13+ year old x86 processor. This processor will be miniaturised to the 45nm fabrication process, they will be assisted by a 512-bit SIMD unit and these cores will support 64-bit address. Gelsinger says that 32 of these cores clocked at 2.00 GHz could belt out 2 TFLOPs of raw computational power. That's close to that of the upcoming AMD R700. Heise also reports that this GPU could have a TDP of as much as 300W (peak).With inputs from Heise
Among several topics, came up the most interesting one, visual computing and Intel's plans on it. 'Larrabee' strikes as a buzzword. It is the codename of Intel's upcoming graphics processor (GPU) architecture with which it plans to take on established players such as NVIDIA and AMD among others.
What's unique (so far) about Larrabee is that it's entirely made up of x86 processing cores. The Larrabee is likely to have 32 x86 processing cores. Here's a surprise: These processing cores are based on the design of Pentuim P54C, a 13+ year old x86 processor. This processor will be miniaturised to the 45nm fabrication process, they will be assisted by a 512-bit SIMD unit and these cores will support 64-bit address. Gelsinger says that 32 of these cores clocked at 2.00 GHz could belt out 2 TFLOPs of raw computational power. That's close to that of the upcoming AMD R700. Heise also reports that this GPU could have a TDP of as much as 300W (peak).With inputs from Heise
77 Comments on Intel Larrabee Capable of 2 TFLOPs
486, Pentium, P4 and possibly Nehalem Rely heavily on Clock speed for performance.
So many chipset driver problems!
Just like you have those 128 / 320 SP's, here, the SP is a x86 processor. It does better out-of-the-order execution,.... crunches numbers better.
Even if Larrabee fails as a GPU, it will most certainly be ported as a CPU, it will be the most powerful CPU ever made.
There's more:
If this thing is sold as a full card by Intel for say $600 (to remain competitive),
as a CPU (when ported to a central processing), even if it costs the same (sans any board, just the chip), you have the most powerful CPU for $600.....2000 TFLOPs on a desktop processor by 2009/10 howzzat?
Unbelieveable power/price.
What is also in the pipeline is a mainboard with an empty socket. And you just plug in a Larrabee for extra zmog horsepower, just like those old x87 chips of yesteryear. Communication with main CPU is via Quickpath.
en.wikipedia.org/wiki/Larrabee_(GPU)
In fact, this is interesting. Based on the spec of each of those in-oder processors, with extra SIMD instructions, they look awfully like Intel 'Atoms'.
Perhaps this is how Intel will scale Westmere/Sandy Bridge. Rather than producing multi-versions of the CPU with core and various atom combinations, it will have just the cores. You then have an add-in socket where you can choose to add 8, 16 or 32 (or whatever) atoms as a Larrabee add-in chip.
Nice.
Pentium P54C is the Pentium 75-200 MHz. Pentium MMX (133-233 MHz) was P55C. Obviously this Larrabee chip isn't made out of 1996 Pentium CPUs. It would never clock above ~300 MHz if that were the case. They just used them as a architectural hint. Actually, the Atom CPU is based on a core from Larrabee, I think. Atom is similar to P54C too.
Atom's lack of speed vs. power use can be an indication of the potential of each Larrabee core. A Core 2 core is dramatically more powerful for most applications. Larrabee will only be fast for apps that can spread across its many cores.
Larrabee will likely still be almost powerless in games, compared to even last generation GPU's. It may have 30 cores, but guess how many ALU's each core has - that's right 1, just like any other CPU. Considering the 2 TFLOP computational power assesment, it is likely a very powerful ALU, but it would still only amount to the same amount on a GPU, which puts the Larrabee at a huge disadvantage against identically-architectured GPU's, such as the G92. It would be a lot more powerful, naturally, but just as the 800 ALU's running under the "R700"' core, it will fail at performing gaming-specific operations. That being said, it could still be great for CUDA, physics or.. just general computing. Because that's what scalar-based "CPU ALU's" are good for - everything, but they perform at it much less proficiantly.
That being said, this is fascinating to even draw up. I would want to see it in action.
I dont think that intel is expecting a whole lot from their first dedicated graphics card.
People need top stop complaining about how the intel graphics card will be using old architecture. If the old architecture works then whats to complain about? I dont have a beef as long as its not overpriced and gives a bad framerate for games. None of which we are sure on yet.
arstechnica.com/articles/paedia/hardware/clearing-up-the-confusion-over-intels-larrabee.ars
"the cores will also have a super-wide 512-bit vector FPU that's capable of processing sixteen-element floating-point vectors (single precision), along with support for control flow instructions (loops and branches) and some scalar computations."
That sounds interesting ^