Larrabee delayed: anyone know what's happening?
|
Prev: PEEEEEEP
Next: Texture units as a general function
From: Paul Wallich on 9 Dec 2009 13:10 Robert Myers wrote: > On Dec 7, 9:51 am, "Andy \"Krazy\" Glew" <ag-n...(a)patten-glew.net> > wrote: >> n...(a)cam.ac.uk wrote: >>> Now, there I beg to disagree. I have never seen anything reliable >>> indicating that Larrabee has ever been intended for consumers, >>> EXCEPT as a 'black-box' GPU programmed by 'Intel partners'. And >>> some of that information came from semi-authoritative sources in >>> Intel. Do you have a reference to an conflicting statement from >>> someone in Intel? >> http://software.intel.com/en-us/blogs/2008/08/11/siggraph-larrabee-an... >> >> Just a blog, not official, although of course anything blogged at Intel >> is semi-blest (believe me, I know the flip side.) > > The blog post reminded me. I have assumed, for years, that Intel > planned on putting many (>>4) x86 cores on a single-die. I'm sure I > can find Intel presentations from the nineties that seem to make that > clear if I dig hard enough. > > From the very beginning, Larrabee seemed to be a technology of destiny > in search of a mission, and the first, most obvious mission for any > kind of massive parallelism is graphics. Thus, Intel explaining why > it would introduce Larrabee at Siggraph always seemed a case of > offering an explanation where none would be needed if the explanation > weren't something they weren't sure they believed themselves (or that > anyone else would). It just seemed like the least implausible mission > for hardware that had been designed to a concept rather than to a > mission. A more plausible claim that they were aiming at HPC probably > wouldn't have seemed like a very attractive business proposition for a > company the size of Intel. > > Also from the beginning, I wondered if Intel seriously expected to be > able to compete at the high end with dedicated graphics engines using > x86 cores. Either there was something about the technology I was > missing completely, it was just another Intel bluff, or the "x86" > cores that ultimately appeared on a graphics chips for market would be > to an x86 as we know it as, say, a, lady bug is to a dalmatian. From an outside perspective, this sounds a lot like the Itanic roadmap: announce something brilliant and so far out there that your competitors believe you must have solutions to all the showstoppers up your sleeve. Major difference being that Larrabee's potential/probable competitors didn't fold. paul
From: Robert Myers on 9 Dec 2009 15:25 On Dec 9, 3:47 am, torb...(a)diku.dk (Torben Ægidius Mogensen) wrote: > > Libraries are, of course, important to supercomputer users. But if they > are written in a high-level language and the new CPU uses the same > representation of floating-point numbers as the old (e.g., IEEE), they > should compile to the new platform. Sure, some low-level optimisations > may not apply, but if the new platform is a lot faster than the old, > that may not matter. And you can always address the optimisation issue > later. > But if some clever c programmer or committee of c programmers has made a convoluted and idiosyncratic change to a definition in a header file, you may have to unscramble all kinds of stuff hidden under macros just to get it to compile and link, and that effort can't be deferred until later. Robert.
From: Robert Myers on 9 Dec 2009 16:49 On Dec 9, 1:10 pm, Paul Wallich <p...(a)panix.com> wrote: > From an outside perspective, this sounds a lot like the Itanic roadmap: > announce something brilliant and so far out there that your competitors > believe you must have solutions to all the showstoppers up your sleeve. > Major difference being that Larrabee's potential/probable competitors > didn't fold. In American football, "A good quarterback can freeze the oppositions defensive secondary with a play-action move, a pump fake or even his eyes." http://www.dentonrc.com/sharedcontent/dws/drc/opinion/editorials/stories/DRC_Editorial_1123.2e4a496a2.html where the analogy is used in a political context. If I were *any* of the players in this game, I'd be studying the tactics of quarterbacks who need time to find an open receiver, since *no one* appears to have the right product ready for prime time. If I were Intel, I'd be nervous, but if I were any of the other players, I'd be nervous, too. Nvidia stock has drooped a bit after the *big* bounce it took on the Larrabee announcement, but I'm not sure why everyone is so negative on Nvidia (especially Andy). They don't appear to be in much more parlous a position than anyone else. If Fermi is a real product, even if only at a ruinous price, there will be buyers. N.B. I follow the financial markets for information only. I am not an active investor. Robert.
From: "Andy "Krazy" Glew" on 9 Dec 2009 23:12 Robert Myers wrote: > Nvidia stock has drooped a bit after the *big* bounce it took on the > Larrabee announcement, but I'm not sure why everyone is so negative on > Nvidia (especially Andy). They don't appear to be in much more > parlous a position than anyone else. If Fermi is a real product, even > if only at a ruinous price, there will be buyers. Let me be clear: I'm not negative on Nvidia. I think their GPUs are the most elegant of the lot. If anything, I am overcompensating: within Intel, I was probably the biggest advocate of Nvidia style microarchitecture, arguing against a lot of guys who came to Intel from ATI. Also on this newsgroup. However, I don't think that anyone can deny that Nvidia had some execution problems recently. For their sake, I hope that they have overcome them. Also, AMD/ATI definitely overtook Nvidia. I think that Nvidia emphasized elegance, and GP GPU futures stuff, whereas ATI went the slightly inelegant way of combining SIMT Coherent Threading with VLIW. It sounds more elegant when you phrase it my way, "combining SIMT Coherent Threading with VLIW", than when you have to describe it without my terminology. Anyway, ATI definitely had a performance per transistor advantage. I suspect they will continue to have such an advantage over Fermi, because, after all, VLIW works to some limited extent. I think Fermi is more programmable and more general purpose, while ATI's VLIW approach has efficiencies in some areas. I think that Nvidia absolutely has to have a CPU to have a chance of competing. One measly ARM chip or Power PC on an Nvidia die. Or maybe one CPU chip, one GPU chip, and a stack of memory in a package; or a GPU plus a memory interface with a lousy CPU. Or, heck, a reasonably efficient way of decoupling one of Nvidia's processors and running 1 thread, non-SIMT, of scalar code. SIMT is great, but there is important non-SIMT scalar code. Ultimately, the CPU vendors will squeeze GPU-only vendors out of the market. AMD & ATI are already combined. If Intel's Larrabee is stalled, it gives Nvidia some breathing room, bit not much. Even if Larrabee is completely cancelled, which I doubt, Intel would eventually squeeze Nvidia out with its evolving integrated graphics. Which, although widely dissed, really has a lot of potential. Nvidia's best chance is if Intel thrashes, dithering between Larrabee and Intel's integrated graphics and ... isn't Intel using PowerVR in some Atom chips? I.e. Intel currently has at least 3 GPU solutions in flight. *This* sounds like the sort of thrash Intel had - x86/i960/i860 ... I personally think that Intel's best path to success would be to go with a big core + the Intel integrated graphics GPU, evolved, and then jump to Larrabee. But if they focus on Larrabee, or an array of Atoms + a big core, their success will just be delayed. Intel is its own biggest problem, with thrashing. Meanwhile, AMD/ATI are in the best position. I don't necessarily like Fusion CPU/GPU, but they have all the pieces. But it's not clear they know how to use it. And Nvidia needs to get out of the discrete graphics board market niche as soon as possible. If they can do so, I bet on Nvidia.
From: Robert Myers on 9 Dec 2009 23:33
On Dec 9, 11:12 pm, "Andy \"Krazy\" Glew" <ag-n...(a)patten-glew.net> wrote: > And Nvidia needs to get out of the discrete graphics board market niche > as soon as possible. If they can do so, I bet on Nvidia. Cringely thinks, well, the link says it all: http://www.cringely.com/2009/12/intel-will-buy-nvidia/ Robert. |