Intel Kills Consumer Larrabee Plans 166
Posted
by
Soulskill
from the vapor-what dept.
from the vapor-what dept.
An anonymous reader tips news that Intel has canceled plans for a consumer version of their long-awaited and oft-delayed Larrabee chip, opting instead to use it as a development platform product. From VentureBeat:
"'Larrabee silicon and software development are behind where we had hoped to be at this point in the project,' said Nick Knuppfler, a spokesman for Intel in Santa Clara, Calif. 'Larrabee will not be a consumer product.' In other words, it’s not entirely dead. It’s mostly dead. Instead of launching the chip in the consumer market, it will make it available as a software development platform for both internal and external developers. Those developers can use it to develop software that can run in high-performance computers. But Knuppfler said that Intel will continue to work on stand-alone graphics chip designs. He said the company would have more to say about that in 2010."
the performance is there (Score:3, Interesting)
http://www.brightsideofnews.com/news/2009/12/2/intel-larrabee-finally-hits-1tflops---27x-faster-than-nvidia-gt200!.aspx [brightsideofnews.com]
way faster than amd's or nvid's hottest....
So the next mini, low end imac and 13" macbook's w (Score:2, Interesting)
So the next mini, low end imac and 13" macbook's will be stuck with shit video and the mac pro will start at $3000 with 6 core cpus.
Will apple move to amd just to get better video in low end systems?
Heterogeneous Processors Are Doomed (Score:0, Interesting)
The idea that the future of parallel processing somehow rests on the use of a bunch of hybrid cores built on the same die was wrong right out of the gate. If parallel CPU cores are a pain in the ass to program, what makes them think that it will be easier by combining them with a non-compatible type of parallel hardware? The CPU/GPU marriage is a match made in hell and, deep down, Intel knows it. Larrabee was just so much puffery and chest beating, king of the jungle and all that jazz.
The way to solve the parallel programming crisis is by first acknowledging that last century's computing paradigms are completely inadequate in the age of massive parallelism. It is time to change to the true computing religion and abandon the outmoded worship of the hopelessly flawed Turing Machine.
Next in line for destruction: AMD's Fusion. You read it here first.
How to Solve the Parallel Programming Crisis [blogspot.com]
Re:Oh rats (Score:5, Interesting)
NVidia hasn't let ATI do anything. Actually, NVidia is dealing with a series of problems - from serious packaging problems last year to TSMC yield issues now. ATI/AMD has been really effective lately; NVidia historically had a dominant position, but definitely not a monopoly, and I'll say that they have slipped a lot recently. Things change fast in the GPU race, so NVidia may recover quickly. But ATI/AMD have a solid amount of momentum, and the only real execution problem I've seen them make in the last few months in GPUs has been to rely on TSMC.
Take a look at the Dell Zino HD - it combines AMD's 'just enough CPU' with top end GPU to make a very compelling system. Intel has cut NVidia out of the chipsets, so they don't get the synergy that AMD has with ATI.
AMD is definitely better situated for the long haul than NVidia, and actually may be better off than Intel for complete systems.
Re:Oh rats (Score:3, Interesting)
last i checked their flash disks were pretty kickass
Re:Oh rats (Score:1, Interesting)
Intel delivered the first sub-40nm flash memory [intel.com] and has delivered two generations of top-flight solid state drives [anandtech.com]. Intel has [highbeam.com] always [hothardware.com] been strong [encyclopedia.com] in flash memory.
I wonder if Bangalore has anything to do with it. (Score:5, Interesting)
I think the announcement of the 48-core Intel 'Bangalore' chip [slashdot.org] just recently is not a coincidence.
When I first read about the Larrabee chip, I thought the decision to make it a cache coherent SMP chip to be simply insane - architectures like that are very difficult to scale, as the inter-core chatter scales roughly as the factorial of the number of cores. Remember how Larrabee was designed around a really wide 1024-bit ring bus? I bet that's required because otherwise the cores would spend all of their time trying to synchronize between each other.
So, Larrabee is effectively cancelled, but only a day or two before Intel announced an almost identical sounding part without cache-coherence! It sounds to me like they've given up on the 100% x86 compatibility, and realised that a chip with some extra instructions around explicit software controlled memory synchronization and message passing would scale way better. Without cache coherence, a "many core" chip is basically just an independent unit repeated over and over, so scalability should be almost infinite, and wouldn't require design changes for different sizes. That sounds like a much better match for a graphics processor.
While Intel kept their cards relatively close to their chest, from all of the presentations I've seen, no first-gen Larrabee chip could scale beyond 24 cores even with a 1024 bit bus, while the new Bangalore chip starts at 48 cores. There's no public info on how many lanes Bangalore has in its on-chip bus but based on the bandwidth of its 80 core experimental predecessor, I'm guessing it's either 32-bit or 64-bit (per core).
Re:the performance is there (Score:3, Interesting)
So what we have here is Itanium- look good on paper but impossible to be fully utilized.
That constitutes a failure if you ask me.
Actually I hold the exact opposite view. The hardware isn't ready, and by not ready I mean the performance isn't as high as expected due to design issues.
If I am correct Intel doesn't want a repeat of the 1st gen Itanium where on release the brand name is blemished by the less than expected performance. This perception that IA64 is slow continues to haunt Intel up to this day. So by delaying Larrabee, Intel will have time to improve the cpu to the point where on release it will be a killer product (ie. hyped).
It's not as if Intel needs Larrabee in the near future anyway- AMD doesn't have anything significant in the near future as well; even if they do, with Intel's brute engineering capability, they will just pull a Core2 again.
Another possibility is that no game company is able to support Larrabee's architecture. Rather than releasing a product that 1. nothing old can run efficiently on 2. nothing new is designed for, Intel is delaying the release until more developers hop on the gravy train. When that happen, Intel can release the chip and immediately, consumers will be awe by the chip's performance in the newest games.
Re:Oh rats (Score:3, Interesting)
AMD doesn't look to have anything great coming out until late 2010 or early 2011 based on their roadmap [anandtech.com]. It helps that ATI is kicking ass in the graphics space. Right now they're winning on price and power. If they can get more of their 5800 series out in the market and release the mobile versions of those cards sooner rather than later, they'll be able to push a lot of hardware that way. However, they're not a real threat to Intel until they can get their SOC products out the door and offer a really compelling reason to go with their products.
Settling their legal issues with Intel will also help them a lot in the long run, but they're not out of the woods yet. They're still having financial problems, but if they can get through the next 18 months they'll be in great shape. The fact that they've been ahead of schedule on a lot of their new chips in the last year has probably helped substantially as well. AMD is in good position for the long term, but they need to decent sales in the coming quarters, which may be difficult to do with Intel releasing a lot of great new chips, especially in the mobile market where AMD hasn't been particularly strong recently.
Re:So the next mini, low end imac and 13" macbook' (Score:3, Interesting)
but at least they are dedicated graphics solutions
Actually, the 9400m is not. It uses system memory but does a much better job then Intel. It also acts as the memory controller and does system IO. The reason for the parent's comments is that all future Intel CPUs will have integrated memory controllers (like the i7 and i5) and an integrated GPU. Performance will suck but it will make for cheap systems. This will make it difficult for system builders to make a low end system with good graphic performance as the market for such systems will be small. The smaller market will reduce the quality/performance of available parts for those system builders - one of which is Apple.
Re:In other words... (Score:5, Interesting)
Re:I wonder if Bangalore has anything to do with i (Score:3, Interesting)
The problem is, a many-core non cache-coherent x86-like system isn't particularly interesting. The big advantage of Larrabee was that you could treat it like a normal SMP system, including (presumably) running standard multithreaded C code on it. Once you have to deal with memory synchronization explicitly, Larrabee starts to look a lot more (from a programming standpoint) like Fermi, Cypress or whatever other Nvidia/ATI GPUs are out at the time.
There's nothing magic about x86/AMD64 in the HPC world. It's attractive because it is cheap and has good performance. Clusters can, have been, and still are built using POWER and other architectures.
But for "embarrassingly parallel" problems, which are the target application for these chips, cache coherence is often not necessary, and simply imposes a design burden. There are lots of problems where it's better to have 1000x the performance than 1/2 the developer time.
It may not even involve less development time: Others have pointed out that the Unix "fork" mechanism combined with "copy-on-write" at the memory page level would also work, and wouldn't require cache coherency. Similarly, any existing code designed for message-passing supercomputers would work out of the box, with only a recompile using a new library. Developers just have to start thinking in terms of "many processes" instead of "many threads".
I suspect that in the long term (decades), cache coherency will simply not scale, and most computers will use explicit message-passing internally, even at the single processor level. The transition has already started: most new servers are NUMA systems, where there's a concept of "near" and "far" memory visible to the software, and most of the real heavy lifting in PCs are done by the GPUs, some of which do not have complete cache coherency across all cores.