http://www.ddj.com/hpc-high-pe...ce-computing/216402188
An interesting read. Putting in here because larabee is more a CPU than anything else.
An interesting read. Putting in here because larabee is more a CPU than anything else.
A First Look at the Larrabee New Instructions (LRBni)
- Thread starter dmens
- Start date
This is making my nether regions tingle. Me likes.
Basically supports what Nemesis has been repeatedly stating in his more lucid posts. Larrabee is (/will be when released) what NV wished Cuda would have become (eventually someday)...a true general purpose GPU.
Nemesis 1
Lifer
- Dec 30, 2006
- 11,366
- 2
- 0
Thanks. But ya kinda Iwas saying that . Remember when we were discussing the Elbrus compiler. I brought up that link. About transmedias morphing . And how if Transmedia would have had great luck if the design would have included more cache and vectorizing unit with the use of Elbrus compiler.
I thought Larrabe was going to be EPIC backend with intel. Like ATIs but differant. Than Intel would simply morph x86 programms. Intel shocked me here. Truely they did. Intel could have gone either way. Why this way. Seems for now intel is doing what I thought they would force others to do. Strange it is I couldn't see that .It still doesn't make sense. But I think sandy is the key were we shall see just how good intels compilers are on the frontend of these Chips.
When you look at Sandy @ 32nm. Than Whatever at 22nm. Look at what each brings than it gets a little more clear. MAYbe maybe not.
Sandy has AVX and 2 additional operands on sandy only 3 are active . The extra operand for FMA on 22nm.
Nehalem has to operands sandy has 3 active 2 plus 1 for vectorizing. Ivy Bridge @ 22 has 4 operands. 2 + 1 vector+ 1 FMA= 4 operand
Old AMD slides showed Bulldozer adding 1 more operand 2 +1 for FMA=3 operand.
Now . If AMD develops FMA on X86 Than Intel can use AMDs. AS AMD won't be porting to FMA. LOL. Intel gets all the benefits or MOST all the benefits of FMA buy going AVX and since intel is porting x86 instructions to vector instructions AMD can't copy intel . Intel finally got flea off back.
SO Idontcare. Heres what I think I don't believe for one second that AMD or Intel Can DO FMA on there present backends. Both AMD and Intel should be looking at EPIC as new cpu backends . Intel already does FMA on a cpu epic (VLIW). Atis backend is epic(VLIW)
So I think Vectorizing code is intels chance to use epic backend. If you read the article you already know it does't matter what the backend is with these vector processors. The scatter gather section was best info I have read. I was wondering how intel was going to pull this off . That guy did good job of laying it out. SO I think sandy will be Epic backend based solely on the fact that IVY Bridge is suppose to have FMA something both Intel and AMD have is EPIC that make use of FMA. Intel has ll the ducks lined up in a roll. Looks good for them not so good for others. I not tossing AMD out with water here as all AMD will need is the right compiler and its look out intel. But Keep in mind Intels Compilers were already good . In 2004 they bought Merced and elbrus. The Elbrus Compiler has been in development since 1980 for VLIW(epic) AMD would give anything for these compilers. The Elbrus compiler can issue 9 or 10 instructions . If you Had a cpu that could do 8 issue and really tiny whinney. You would double the output of todays larrabee without increasing core count. If FMA doubles FP and Vectorizing code doubles FP. Than your looking at 4x the work of present larrabbe. Thing all has to recall here Intel it has seemed to make C+C look really intersting and CL is going to benefit both the cpu and gpu equally . So Intel AVX+FMA combined with Larrabe native is going to be great for you guys. That old queston which is better to buy Cpu or GPU to upgrade is going to finely fixed. Because Sandy is going to bring a hugh staggering extra gaming power to a Modern made for intel game. Older games somewhat .
I thought Larrabe was going to be EPIC backend with intel. Like ATIs but differant. Than Intel would simply morph x86 programms. Intel shocked me here. Truely they did. Intel could have gone either way. Why this way. Seems for now intel is doing what I thought they would force others to do. Strange it is I couldn't see that .It still doesn't make sense. But I think sandy is the key were we shall see just how good intels compilers are on the frontend of these Chips.
When you look at Sandy @ 32nm. Than Whatever at 22nm. Look at what each brings than it gets a little more clear. MAYbe maybe not.
Sandy has AVX and 2 additional operands on sandy only 3 are active . The extra operand for FMA on 22nm.
Nehalem has to operands sandy has 3 active 2 plus 1 for vectorizing. Ivy Bridge @ 22 has 4 operands. 2 + 1 vector+ 1 FMA= 4 operand
Old AMD slides showed Bulldozer adding 1 more operand 2 +1 for FMA=3 operand.
Now . If AMD develops FMA on X86 Than Intel can use AMDs. AS AMD won't be porting to FMA. LOL. Intel gets all the benefits or MOST all the benefits of FMA buy going AVX and since intel is porting x86 instructions to vector instructions AMD can't copy intel . Intel finally got flea off back.
SO Idontcare. Heres what I think I don't believe for one second that AMD or Intel Can DO FMA on there present backends. Both AMD and Intel should be looking at EPIC as new cpu backends . Intel already does FMA on a cpu epic (VLIW). Atis backend is epic(VLIW)
So I think Vectorizing code is intels chance to use epic backend. If you read the article you already know it does't matter what the backend is with these vector processors. The scatter gather section was best info I have read. I was wondering how intel was going to pull this off . That guy did good job of laying it out. SO I think sandy will be Epic backend based solely on the fact that IVY Bridge is suppose to have FMA something both Intel and AMD have is EPIC that make use of FMA. Intel has ll the ducks lined up in a roll. Looks good for them not so good for others. I not tossing AMD out with water here as all AMD will need is the right compiler and its look out intel. But Keep in mind Intels Compilers were already good . In 2004 they bought Merced and elbrus. The Elbrus Compiler has been in development since 1980 for VLIW(epic) AMD would give anything for these compilers. The Elbrus compiler can issue 9 or 10 instructions . If you Had a cpu that could do 8 issue and really tiny whinney. You would double the output of todays larrabee without increasing core count. If FMA doubles FP and Vectorizing code doubles FP. Than your looking at 4x the work of present larrabbe. Thing all has to recall here Intel it has seemed to make C+C look really intersting and CL is going to benefit both the cpu and gpu equally . So Intel AVX+FMA combined with Larrabe native is going to be great for you guys. That old queston which is better to buy Cpu or GPU to upgrade is going to finely fixed. Because Sandy is going to bring a hugh staggering extra gaming power to a Modern made for intel game. Older games somewhat .
Originally posted by: dmens
of course the *really* interesting details are not in the article.
Do you mean that there are technical features of interest that the article has omitted? Like it's incomplete?
Or do you mean that the article doesn't delve into the interesting consequences?
IntelUser2000
Elite Member
- Oct 14, 2003
- 8,686
- 3,787
- 136
I corrected you on other thread but Ivy Bridge is a mere shrink of Sandy Bridge. Haswell is the Tock after that and features FMA.
Can you expand on your premise for why no FMA in Sandy Bridge precludes Ivy Bridge from having FMA?
Penryn was a mere die shrink of Conroe and yet they shoved SSE4.1 instructions into it.
I haven't seen anything to date that says FMA on Ivy Bridge can't happen or won't happen, outside of your posts that is, so if I am going to discard my vision of the future based on extrapolation of historical events then I'd prefer to do so with something more substantial than faith alone.
Convince me.
I'm with you, not disputing your position. But I see Inteluser disputing your position and personally I'd like to know more about Inteluser's logic behind their position on the topic.
If the logic behind their opinion is superior to my own then I will discard mine and adopt theirs.
At this point in time sufficient evidence/logic does not exist to compel me to discard my opinion, but I am fully open to this changing. Inteluser has yet to steer me wrong in many threads to date, but I want to know what's behind the steering as I am not much of a faith-based kind of person in these types of situations.
FWIW SUN architecture should never be invoked as being exemplary of how something could not be implemented very well...having worked with them and their products they are in every sense of the concept an american company operated by american management the likes of which attended the same business management schools that the CEO's of the auto industry were indoctrinated.
The finest engineers on the planet could be working at GM and at SUN, but you wouldn't know it from the products their management directs them to create and market.
Nemesis 1
Lifer
- Dec 30, 2006
- 11,366
- 2
- 0
Not alot of info available but this is as good as any others. If you shop to read you will find much confusion On FMA when where why. Best source is Intel blogs but be carefull these guys are sharp . Alot of people angry over delay of FMA on sandy.
http://www.dailytech.com/Detai...f+IDF/article12685.htm
http://www.dailytech.com/Detai...f+IDF/article12685.htm
Nemesis 1
Lifer
- Dec 30, 2006
- 11,366
- 2
- 0
Actually Haswell would be 3rd generation Sandy so it would be 2 ticks in aroll. As the operands are already in sandy only three active. So its when intel desides to use . It could happen on Ivy bridge. Haswell is not a TOCK. Sandy goes like this Tock tick than another Tick with FMA active. Ivy Bridge =22nm . Haswell= 22nm. FMA and power saving added. I would imagine Intel will do FMA when it suites them . Keeping eye on AMD/ NV/ IBM at same time.
lopri
Elite Member
- Jul 27, 2002
- 13,310
- 687
- 126
I just wanted to say hi to you, Vee. You've been missed!Originally posted by: Vee
Originally posted by: dmens
of course the *really* interesting details are not in the article.
Do you mean that there are technical features of interest that the article has omitted? Like it's incomplete?
Or do you mean that the article doesn't delve into the interesting consequences?
As for the article, I am not knowledgeable enough to understand it without some aid so I'll leave it up to you and others. Larrabee looks like a kind of 'brute force' approach and with the rumored die size of 600mm^2, I can't imagine what the power consumption or how it will fit the propaganda "Performance per Watt" that Intel's been pushing.
PlasmaBomb
Lifer
- Nov 19, 2004
- 11,636
- 2
- 81
Considering that a Larrabee-enabled product (GPU, whatever) is appearing to be more like 10% hardware and 90% software my guess is there are teams of programmers working at breakneck speeds trying to get compilers and drivers fleshed out and ready for optimization still.
how fast a machine using the described LRBni ISA run, and how fast can it do graphics, how energy efficient is this software pipeline versus the established models? those are the details which are most interesting imho.
on a sidenote, anyone know of good technical books on computer graphics?
No textbooks, but...
http://graphics.stanford.edu/c.../cs248-07/schedule.php
http://courses.ece.illinois.edu/ece498/al/
http://nehe.gamedev.net/ The tutorials are good.
http://graphics.stanford.edu/c.../cs248-07/schedule.php
http://courses.ece.illinois.edu/ece498/al/
http://nehe.gamedev.net/ The tutorials are good.
IntelUser2000
Elite Member
- Oct 14, 2003
- 8,686
- 3,787
- 136
Sure I will. Number 1: http://www.canardpc.com/dossie...Nehalem_a_Haswell.html
Where its basically the only site Ivy Bridge and Haswell is EVER mentioned outside of Intel, and all other sites(like wiki) pulled their data from.
2. Fused Multiply-Add isn't a simple instruction you just slap on like the SSE. It's a big change in the order of adding single cycle SSE. Actually, think of it as adding single cycle SSE on top of a CPU that has absolutely no SSE. Probably more than that which is why only the handful of CPUs support it(Itanium/Power).
Nemesis 1
Lifer
- Dec 30, 2006
- 11,366
- 2
- 0
Just Googled still a few sites with info . But ya most have pulled the old info . Which said same as what this said. So why is this stuff being pulled? But Like the link a gave yesterday / There is still info out there . Like the 4 operands being present in Sandy. Just 3 active. Now what all that means I not sure . But If intel has to wait till Haswell the Cost most be to high right now performance gain Vs. power efficiency. Yep I want to see INtel pull this off. FMA on there present backends . Same as AMD . I want to see this LOL. How many times do I have to say that . SHOW ME. I believe in GOD. This Not so much LOL!
People not use to me LOL at intel . But thats exactly what I been doing. THe only place were going to see FMA from intel is on epic backend. I will say FMA brings alot to table thats true. But I still LOL at BOTH AMD/INTEL thinking there going to make this work with present ARCH . I DON"T BELIEVE YOU GUYS. at AMD/INTEL. NOT this time . You show me FMA on desktop CPU and I can show you intels new Itanic. CAll it what they will.
Haswell is a tock tho . As its reported that Haswell actually gets undie vertex unit. Which explains the reported power efficiency improvements
I need to see how this relates to FMA. Cool more research .
People not use to me LOL at intel . But thats exactly what I been doing. THe only place were going to see FMA from intel is on epic backend. I will say FMA brings alot to table thats true. But I still LOL at BOTH AMD/INTEL thinking there going to make this work with present ARCH . I DON"T BELIEVE YOU GUYS. at AMD/INTEL. NOT this time . You show me FMA on desktop CPU and I can show you intels new Itanic. CAll it what they will.
Haswell is a tock tho . As its reported that Haswell actually gets undie vertex unit. Which explains the reported power efficiency improvements
I need to see how this relates to FMA. Cool more research .
- Sep 28, 2005
- 21,053
- 3,535
- 126
Originally posted by: PlasmaBomb
Your just all hormonal because you want a new toy to play with
well yeah true...
oh what if i told you, theres a W5590 thats expected to come out. Its a gainestown unlocked d0, kinda identical to my 975 but for gainestown side.
Also what if i told you guys gainestown is undergoing another stepping revision of E0?
Meh i want an i5 to upgrade kagami into a better HTPC.
Nemesis 1
Lifer
- Dec 30, 2006
- 11,366
- 2
- 0
Thanks much for that link . I read this befor . Usually I don't miss Important details but I did. I thought that Sandy AVX had an actual Vertex unit on die. Here In the same article you linked . It says Haswell gets the First on die vertex unit. Maybe I should find out exactly how this relates to FMA. Befor lol at intel and sticking foot in mouth. So Haswell is going to be More LArrabee like. Using intels CLONE won't even need GPU. or not much of one. To have great compute abilities for software render graphics.
2. Fused Multiply-Add isn't a simple instruction you just slap on like the SSE. It's a big change in the order of adding single cycle SSE. Actually, think of it as adding single cycle SSE on top of a CPU that has absolutely no SSE. Probably more than that which is why only the handful of CPUs support it(Itanium/Power).
Its interesting what larabbee is suppose to have accomplished. If it works as stated by intel . Intels IP on scatter/gather is worth a trillion dollars. If that mask to memory actually works. Its a crusher to everyone else. I just can't see how a vertex unit that has access to FMA would actual double a 256bit vertex unit. Like larrabbee has. That really seems like hugh power gains. If its true . Isn't intel infact selling a lemon to the public in the form of IVY Bridge. BAD BAD BAD INTEL . BOOOO ! OR is IVY a Shrink of Sandy . Than if thats all true how can Haswell be Third generation sandy. This stuff is coming way to fast. In 2010 you be crazy to buy Nehalem with Sandy so close. Buying IVy makes zero sense if you didn't buy Sandy . AS Haswell looks to be ass kicking.
Why do I get the Feeling intel can do all this at 32nm . But are holding back for future sales. That sucks.
ilkhan
Golden Member
- Jul 21, 2006
- 1,117
- 1
- 0
haswell is to ivy bridge what nehalem is to core. new arch. ivy bridge is the shrink of sandy bridge, like westmere is to nehalem. AFAIK, if haswell takes too long, they may do a third year of sandybridge (sandy/ivy/x), but that wont be haswell.
Listening to [Nemesis 1] is so strange, because I really don't know how much is impressive sounding bullshit and how much is genuine technical knowledge far beyond my level.
We're expecting larrabee around the 2009/2010 boundry, right? About the same time as 32nm comes out. But larrabee is 45nm, yes?
Listening to [Nemesis 1] is so strange, because I really don't know how much is impressive sounding bullshit and how much is genuine technical knowledge far beyond my level.
We're expecting larrabee around the 2009/2010 boundry, right? About the same time as 32nm comes out. But larrabee is 45nm, yes?
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads
- Started by Tigerick
- Replies: 21K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
-
Facebook
Twitter