Larrabee delayed: anyone know what's happening?
|
Prev: PEEEEEEP
Next: Texture units as a general function
From: Bernd Paysan on 10 Dec 2009 09:28 Michael S wrote: > I they are doing all that I simply can't see how one of existing GPUs > (i.e. not Fermi) could possibly beat 3 GHz Nehalem by factor of >10. > Nehalem is rated at ~100 SP GFLOPs. Are there GPU chips that are > significant above 1 SP TFLOPs? According to Wikipedia there are not. AFAIK the ATI 5870 can achieve up to 3.04 SP TFLOPS at 950MHz. That's a single chip. And the data is on Wikipedia - where did you look? Look here: http://en.wikipedia.org/wiki/FLOPS -- Bernd Paysan "If you want it done right, you have to do it yourself" http://www.jwdt.com/~paysan/
From: Michael S on 10 Dec 2009 12:11 On Dec 10, 4:28 pm, Bernd Paysan <bernd.pay...(a)gmx.de> wrote: > Michael S wrote: > > I they are doing all that I simply can't see how one of existing GPUs > > (i.e. not Fermi) could possibly beat 3 GHz Nehalem by factor of >10. > > Nehalem is rated at ~100 SP GFLOPs. Are there GPU chips that are > > significant above 1 SP TFLOPs? According to Wikipedia there are not. > > AFAIK the ATI 5870 can achieve up to 3.04 SP TFLOPS at 950MHz. That's a > single chip. And the data is on Wikipedia - where did you look? Look here: > > http://en.wikipedia.org/wiki/FLOPS > > -- > Bernd Paysan > "If you want it done right, you have to do it yourself"http://www.jwdt.com/~paysan/ I looked at the same page but didn't pay attention to a fine print at the bottom :( Anyway, Radeon HD 5870 is in the field for 2 month or something like that? Somehow I don't think Terje's buddies had it for 8 rounds of optimization. Also it seems up until very recently very few organizations tried non- NVidea GPGPU.
From: Torbjorn Lindgren on 10 Dec 2009 13:14 Torben �gidius Mogensen <torbenm(a)diku.dk> wrote: >"Andy \"Krazy\" Glew" <ag-news(a)patten-glew.net> writes: >> isn't Intel using PowerVR in some Atom chips? > >I know ARM uses PowerVR, but I hadn't heard Intel doing so. The chipset for the MID/UMPC Atom's is Poulsbo (aka "Intel System Controller Hub US15W") and contains "GMA500". GMA500 is entirely unrelated to all other GMA models and consists of PowerVR SGX 535 (graphics) and PowerVR VXD (H.264/MPEG-4 AVC playback)... IIRC only the special MID/UMDPC Atom's can be coupled with Poulsbo (ie Z5xx/Silverthorne)? The other Intel chipsets are uses a lot more power than the Atom itself (extremely bad when you have a low-power CPU) and are pretty anemic to boot. In the other end is Nvidia Ion which also uses a bit more power than one would hope but at least have an usefull GPU/HD playback accelerator (like Poulsbo but faster). http://en.wikipedia.org/wiki/Poulsbo_(chipset) http://en.wikipedia.org/wiki/Intel_GMA http://en.wikipedia.org/wiki/Intel_Atom#Power_requirements
From: jgd on 10 Dec 2009 16:15 In article <hfk1mu$i7j$1(a)smaug.linux.pwf.cam.ac.uk>, nmm1(a)cam.ac.uk () wrote: > Yes. But the word "planned" implies a degree of deliberate action > that I believe was absent. They assuredly blithered on about it, > and very probably had meetings about it .... Indeed. Intel don't seem to have become serious about multi-core until they discovered that they could not clock the NetBurst above 4GHz, but that their fab people could readily fit two of them on a single die. I did some work with the early "Pentium D", which was two NetBursts on the same die, but with two sets of legs, and no communication between the cores that didn't go through the legs and the motherboard FSB. Locking performance was unimpressive, to say the least, and early Opterons beat it utterly. I'm going to take a lot of convincing that this was a long-planned product; the design just isn't good enough for that to be convincing. -- John Dallman, jgd(a)cix.co.uk, HTML mail is treated as probable spam.
From: Robert Myers on 10 Dec 2009 16:47
On Dec 10, 4:15 pm, j...(a)cix.compulink.co.uk wrote: > I did some work with the early "Pentium D", which was two NetBursts on > the same die, but with two sets of legs, and no communication between > the cores that didn't go through the legs and the motherboard FSB. > Locking performance was unimpressive, to say the least, and early > Opterons beat it utterly. I'm going to take a lot of convincing that > this was a long-planned product; the design just isn't good enough for > that to be convincing. The charts I remember, and I'm sure they were from the last millennium, observed the rate at which power per unit area was increasing, had a space shuttle thermal tile number on the same slide for comparison, and concluded that the trend was not sustainable. There were, in concept, at least two ways you could beat the trend: go to multiple cores not running so fast (the proposal in that presentation) or bet on a miracle. Apparently, the NetBurst team was betting on a miracle. From the outside, Intel looks arrogant enough to believe that they could do multiple cores when they were forced to and no sooner. In actuality, they weren't far wrong. Most people don't remember Pentium D. Robert. |