"MMX sucks" part II

DJ Supreme · Jan 25, 2000

Regarding what total quoted Hannibal on:>In the grand scheme of things, Altivec adds complexity and instructions to the CPU, and thus represents a movement back towards a CISC-like design philosophy, where functionality is added to the die vs. being implemented in softwareThis it completely incorrect. This is an archaic and outdated definition of RISC being employed here. RISC today is defined not by the instruction count (ex. a G3 and P3 is roughly equal in the respect), but the instruction TYPE count, such as add, mul, permute etc.Furthermore, the AltiVec unit still uses the universal load/store unit in the G4, which still only has 4 addressing modes, which again conforms the modern RISC definition - few adressing modes. Furthermore, the PPC series' instruction set does not mix memory/register adressing with arithmetic, which the x86 does - another threshhold between RISC/CISC.Lastly, regarding the "MMX sucks" thread. OK, maybe it was an offensive and inappropriate title. But some people thought it was a bash at the PC.People: Let me assure you, it's not! IT'S NOT! I use Macs and PC's side by side every day (I have one of each on my desk at home), and I like my PC for what I do with it. My criticism with MMX is not to be takes as a criticism of the PC in general, but merely as an objection to the lowly technical limitations of MMX and the mass-marketing approach Intel took when announcing it.Hope this clears matters up.

DrPizza · Jan 25, 2000

<BLOCKQUOTE>quote:<HR>Originally posted by DJ Supreme: Regarding what total quoted Hannibal on:>In the grand scheme of things, Altivec adds complexity and instructions to the CPU, and thus represents a movement back towards a CISC-like design philosophy, where functionality is added to the die vs. being implemented in softwareThis it completely incorrect. This is an archaic and outdated definition of RISC being employed here. RISC today is defined not by the instruction count (ex. a G3 and P3 is roughly equal in the respect), but the instruction TYPE count, such as add, mul, permute etc.<HR></BLOCKQUOTE> That's missing the fundamental idea of RISC. Instruction count, instruction type count, are both irrelvent to RISC. The idea of RISC is a simple one. Make the compiler complicated and the hardware simple. Altivec is completely opposed to that. The compiler is (relatively) simple, the hardware is complicated. <BLOCKQUOTE>quote:<HR>Furthermore, the AltiVec unit still uses the universal load/store unit in the G4, which still only has 4 addressing modes, which again conforms the modern RISC definition - few adressing modes. Furthermore, the PPC series' instruction set does not mix memory/register adressing with arithmetic, which the x86 does - another threshhold between RISC/CISC.<HR></BLOCKQUOTE> Those are all just minor features of RISC. The fundamental concept -- dumb silicon, smart compiler -- is broken by Altivec. <BLOCKQUOTE>quote:<HR>Lastly, regarding the "MMX sucks" thread. OK, maybe it was an offensive and inappropriate title. But some people thought it was a bash at the PC.<HR></BLOCKQUOTE> Since PCs are the only computers in the world with MMX, that seems fair enough.... <BLOCKQUOTE>quote:<HR>People: Let me assure you, it's not! IT'S NOT! I use Macs and PC's side by side every day (I have one of each on my desk at home), and I like my PC for what I do with it. My criticism with MMX is not to be takes as a criticism of the PC in general, but merely as an objection to the lowly technical limitations of MMX<HR></BLOCKQUOTE> Lowly? Well, yeah, it's integer only. But it's also ~4 years old. Altivec is brand new -- wouldn't you expect Altivec to be a little better than MMX? <BLOCKQUOTE>quote:<HR>and the mass-marketing approach Intel took when announcing it.<HR></BLOCKQUOTE> Intel's approach wasn't too bad. By making MMX a buzzword, they encouraged its use -- so that people making products could claim 'Uses MMX'. This is a Good Thing. It's lost its buzzword status now, because its usage has become far more common. This is thanks to Intel's advertising. It was a pretty smart idea.

resteves · Jan 25, 2000

<BLOCKQUOTE>quote:<HR>Intel's approach wasn't too bad. By making MMX a buzzword, they encouraged its use -- so that people making products could claim 'Uses MMX'. This is a Good Thing. It's lost its buzzword status now, because its usage has become far more common. This is thanks to Intel's advertising. It was a pretty smart idea.<HR></BLOCKQUOTE> Hhmmmmm.... When Intel pushes a mediocre product through marketing muscle and a ton of advertising, it is a 'good' thing. I wonder what PeterB's reaction would be if Apple did the same thing?

DrPizza · Jan 25, 2000

Resteves: It wasn't mediocre. It might be today when compared with Altivec, but in 1997 it wasn't mediocre. So Intel weren't pushing a mediocre technology with huge amounts of advertising, they were pushing a good technology with huge amounts of advertising.

DJ Supreme · Jan 25, 2000

<BLOCKQUOTE>quote:<HR>Originally posted by PeterB:That's missing the fundamental idea of RISC. Instruction count, instruction type count, are both irrelvent to RISC. The idea of RISC is a simple one. Make the compiler complicated and the hardware simple. Altivec is completely opposed to that. The compiler is (relatively) simple, the hardware is complicated.No it isn't. The build of the PPC chips follows simple and general rules. Even-sized registers, few addressing modes. Please tell me in which way you define "complicated" - boasting a lot of transistors?

Furthermore, the AltiVec unit still uses the universal load/store unit in the G4, which still only has 4 addressing modes, which again conforms the modern RISC definition - few adressing modes. Furthermore, the PPC series' instruction set does not mix memory/register adressing with arithmetic, which the x86 does - another threshhold between RISC/CISC.<HR></BLOCKQUOTE> Those are all just minor features of RISC. The fundamental concept -- dumb silicon, smart compiler -- is broken by Altivec.How? The altiVec unit is still following the simple rules which the rest of the PPC CPU does - evensized registers, a homogenous register file, even-sized instructions. et al. <BLOCKQUOTE>quote:<HR>Lastly, regarding the "MMX sucks" thread. OK, maybe it was an offensive and inappropriate title. But some people thought it was a bash at the PC.<HR></BLOCKQUOTE> Since PCs are the only computers in the world with MMX, that seems fair enough....Don't assume. It wasn't. <BLOCKQUOTE>quote:<HR>People: Let me assure you, it's not! IT'S NOT! I use Macs and PC's side by side every day (I have one of each on my desk at home), and I like my PC for what I do with it. My criticism with MMX is not to be takes as a criticism of the PC in general, but merely as an objection to the lowly technical limitations of MMX<HR></BLOCKQUOTE> Lowly? Well, yeah, it's integer only. But it's also ~4 years old. Altivec is brand new -- wouldn't you expect Altivec to be a little better than MMX?The age is irrelevant - Intel could have done it a lot better than they actually did. Do you remember when the P55C (Pentium MMX) was released? intel preached that it accelerated all applications, implicitly due to MMX. Well, the general acceleration of execution of code was due to the unmentioned fact that the L1 cache size was doubled... <BLOCKQUOTE>quote:<HR>and the mass-marketing approach Intel took when announcing it.<HR></BLOCKQUOTE> Intel's approach wasn't too bad. By making MMX a buzzword, they encouraged its use -- so that people making products could claim 'Uses MMX'. This is a Good Thing. It's lost its buzzword status now, because its usage has become far more common. This is thanks to Intel's advertising. It was a pretty smart idea.In other words, when Intel uses buzzwords, it's a Good Thing, but when Apple does it, they're a fucking bunch of freaks appealing only to morons? (well, that's what the word has been here in the battlefront for the last year or so)

Dan · Jan 25, 2000

The guy says, oh, sorry for the offensive title and then uses it again. And why so hung up on this stuff anyway? Who really cares about MMX? It was great, like, 4 years ago.

DJ Supreme · Jan 25, 2000

<BLOCKQUOTE>quote:<HR>Originally posted by PeterB: Resteves: It wasn't mediocre. It might be today when compared with Altivec, but in 1997 it wasn't mediocre. So Intel weren't pushing a mediocre technology with huge amounts of advertising, they were pushing a good technology with huge amounts of advertising.<HR></BLOCKQUOTE>PeterB, please realise this: MMX was SIMD integer-only, along with the fact that when running, it made use of the FP register file, resulting in long context switches and huge resulting performance penalties. This is NOT the hallmark of a great product (IOW: MMX was 90% marketing and 10% technical innovation).Time is NOT a factor here, and if Intel really wanted, they COULD have made a great SIMD solution, but instead they chose to push a half-baked piece of garbage. And they didn't even allow you to compile for the thing.You can't defend this rationally, Peter.

DJ Supreme · Jan 25, 2000

<BLOCKQUOTE>quote:<HR>Originally posted by Dan: The guy says, oh, sorry for the offensive title and then uses it again.Well, sorry if I hurt you, Dan, but it was the only way to get attention around the previous thread. BTW, didn't you notcie the quotation marks?>And why so hung up on this stuff anyway? Who really cares about MMX? It was great, like, 4 years ago.<HR></BLOCKQUOTE>It was great only in hype and marketing. It's reality to anyone using a Pentium MMX or newer today. If you want to do SIMD on a Wintel box, you have to dig into MMX.[This message has been edited by DJ Supreme (edited January 25, 2000).]

DrPizza · Jan 25, 2000

<BLOCKQUOTE>quote:<HR>How? The altiVec unit is still following the simple rules which the rest of the PPC CPU does - evensized registers, a homogenous register file, even-sized instructions. et al.<HR></BLOCKQUOTE> So? Those features are not what make the PPC ISA a RISC ISA. The thing that makes the PPC RISC is that the nature of the instruction is simple. They do simple things. The difficult things get done by combining the instructions, which is controlled by the compiler. The thing that Altivec does is to have single instructions that do complicated things. The complexity is moved from the compiler to the processor. That's directly contrary to RISC's intentions. <BLOCKQUOTE>quote:<HR>Don't assume. It wasn't.<HR></BLOCKQUOTE> Why? MMX is a feature that all PCs sold in the last 3-odd years have. It is a standard PC feature, specific to PCs. It's a constituent part of every new PC. Every last one. That makes an attack on MMX an attack on PCs. <BLOCKQUOTE>quote:<HR>The age is irrelevant<HR></BLOCKQUOTE> You're kidding, right? So we might as well criticize Motorola/Apple because they're 1990 vintage machines are slow and crap compared to modern computers. <BLOCKQUOTE>quote:<HR>- Intel could have done it a lot better than they actually did.<HR></BLOCKQUOTE> Could they? Could they really? How do you know? By comparison with what, say, AIM had at the time? Oh, I forgot. They didn't have an SIMD instruction set at that time, did they. If Intel could have made something as good as (say) Altivec in 1996/7, why didn't they? Why didn't Moto? Why didn't IBM?I find that assertion pretty ludicrous, actually. <BLOCKQUOTE>quote:<HR>Do you remember when the P55C (Pentium MMX) was released? intel preached that it accelerated all applications, implicitly due to MMX. Well, the general acceleration of execution of code was due to the unmentioned fact that the L1 cache size was doubled...<HR></BLOCKQUOTE> I don't believe that Intel did do that, actually. Yeah, they said that the MMX processors were faster, and it was obviously because of the bigger cache, but I never got the impression that they claimed it was because of MMX. They touted the performance of MMX-enhanced applications (and for them it can make a big difference). <BLOCKQUOTE>quote:<HR>In other words, when Intel uses buzzwords, it's a Good Thing, but when Apple does it, they're a fucking bunch of freaks appealing only to morons? (well, that's what the word has been here in the battlefront for the last year or so)<HR></BLOCKQUOTE> I don't believe that Intel and Apple are particularly comparable, actually. Intel don't sell direct to consumers (there might be a small number of sales, but not very many). Apple do. Intel make the processors. Apple do not. Intel have competition. Apple do not. MMX was new and different when it was released. Altivec was not. Intel have to actually strive to get market acceptance -- because of the competition within the PC processor market. Apple do not; Apple write the OS, so Apple can ensure that all the important parts of it use Altivec. The positions of the two companies are very different.MMX wasn't at all bad, but in the PC market it's often difficult to encourage the uptake of new technology. Intel have to market it to consumers so that consumers demand of their software vendors support for MMX. Apple don't really have to. Apple have no great interest in whether consumers demand Altivec or not. It makes no difference to Apple -- they're not the ones making the processors. It doesn't matter to Apple if there's no uptake of Altivec -- it doesn't cost them much. It does matter, however, to Intel if there's no uptake of MMX, as it will cost them a lot.

DJ Supreme · Jan 25, 2000

<BLOCKQUOTE>quote:<HR>Originally posted by PeterB:

How? The altiVec unit is still following the simple rules which the rest of the PPC CPU does - evensized registers, a homogenous register file, even-sized instructions. et al.<HR></BLOCKQUOTE> >So? Those features are not what make the PPC ISA a RISC ISA.Yes, these things are closely connected. Fx. that the first 4 registers in an x86 chip are 8-bit segmentable bears witness to the differentiated and complex nature of the x86 instructions.>The thing that makes the PPC RISC is that the nature of the instruction is simple. They do simple things.Rather that inside their groups, their function is very homogenous - same artithmetic types, same I/O commands...>The difficult things get done by combining the instructions, which is controlled by the compiler. The thing that Altivec does is to have single instructions that do complicated things.Such as permuting? It's not complicated, it's just a lot of simple things done a lot of times. That doesn't make it complex. The rules followed are still VERY simple.>The complexity is moved from the compiler to the processor. That's directly contrary to RISC's intentions.I don't agree. <BLOCKQUOTE>quote:<HR>Don't assume. It wasn't.<HR></BLOCKQUOTE> >Why? MMX is a feature that all PCs sold in the last 3-odd years have. It is a standard PC feature, specific to PCs. It's a constituent part of every new PC. Every last one. That makes an attack on MMX an attack on PCs.ARRRRRGH! GOD DAMMIT, Peter, will you PLEASE listen to what I'm saying?!? I'm refuting the claim that my post was an attack at the PC! I say it wasn't and that IT WAS NOT intended that way! <BLOCKQUOTE>quote:<HR>The age is irrelevant<HR></BLOCKQUOTE> You're kidding, right? So we might as well criticize Motorola/Apple because they're 1990 vintage machines are slow and crap compared to modern computers.In the context of this discussion, that is irrelevant, too. Read what I write, Peter. <BLOCKQUOTE>quote:<HR>- Intel could have done it a lot better than they actually did.<HR></BLOCKQUOTE> Could they? Could they really? How do you know?Oh, more bickering... Motorola did SIMD for both Int and FP types, and gave the unit it's own register file. That's not too hard, or shouldn't be for the world's largest semiconductor chip company.Besides, the fact that Intel finally made SSE available a year ago is a proof of this, too. Don't you see this, Peter?>By comparison with what, say, AIM had at the time? Oh, I forgot. They didn't have an SIMD instruction set at that time, did they.No. They took this time and made a thoroughly crafted product, contrary to Intel half-baked counterpart.>If Intel could have made something as good as (say) Altivec in 1996/7, why didn't they? Why didn't Moto? Why didn't IBM?Intel started out earlier. It's as simple as that. The AIM consortium started out later, but took their time. <BLOCKQUOTE>quote:<HR>Do you remember when the P55C (Pentium MMX) was released? intel preached that it accelerated all applications, implicitly due to MMX. Well, the general acceleration of execution of code was due to the unmentioned fact that the L1 cache size was doubled...<HR></BLOCKQUOTE>>I don't believe that Intel did do that, actually.Believe it. It wasn't until PC World struck it up bigtime that Intel really admitted it. Even Intel's own engineers thought that it was a big hoax...>Yeah, they said that the MMX processors were faster, and it was obviously because of the bigger cache, but I never got the impression that they claimed it was because of MMX.Hey, you just agreed that they hyped up MMX. I'd say this pretty much confirms what I said - you know, every company is so wild to make it look like a new technology of their does wonders. In Intel's case, it was MMX.>They touted the performance of MMX-enhanced applications (and for them it can make a big difference).Yeah, a single Photoshop filter that performed 4x faster. He he he... View image: /infopop/emoticons/icon_smile.gif <BLOCKQUOTE>quote:<HR>In other words, when Intel uses buzzwords, it's a Good Thing, but when Apple does it, they're a fucking bunch of freaks appealing only to morons? (well, that's what the word has been here in the battlefront for the last year or so)<HR></BLOCKQUOTE>>I don't believe that Intel and Apple are particularly comparable, actually. Intel don't sell direct to consumers (there might be a small number of sales, but not very many). Apple do.This isn't relevant. Sales methods isn't relevant - this is about the fact that both companies hypes up their products. Why should Apple's direct-sales methods make their hype different to Intel's???>Intel make the processors. Apple do not. Intel have competition. Apple do not.Even though you're taking the discussion down an irrelevant track, Apple DOES have competition. A Mac is a PC, right? Well, Apple has competition from the entire PC world. It's obvious. If I were one day to think that my Mac sucked so much, that I couldn't really live with it, I'd naturally go buy a PC.Ergo, Apple has competition from the rest of the PC marketplace.>MMX was new and different when it was released. Altivec was not. Intel have to actually strive to get market acceptance -- because of the competition within the PC processor market. Apple do not; Apple write the OS, so Apple can ensure that all the important parts of it use Altivec. The positions of the two companies are very different.I don't see the relevance in this at all.>MMX wasn't at all bad, but in the PC market it's often difficult to encourage the uptake of new technology.Agreed. USB is an excellent example of that. It was first availble in 1996 (ah! there's that mistake I made eralier). A friend of mine got a Pentium 200MMX with USB back then. But it's only today thay USB is starting to gain acceptabce (or, rather, a year ago in the Mac workd due to the appearence of the iMac w/ USB).>Intel have to market it to consumers so that consumers demand of their software vendors support for MMX. Apple don't really have to. Apple have no great interest in whether consumers demand Altivec or not.Not true. This implies that Apple is completely blind to the competition. Remember the competition in the PC market, Peter. >It makes no difference to Apple -- they're not the ones making the processors. It doesn't matter to Apple if there's no uptake of Altivec -- it doesn't cost them much. It does matter, however, to Intel if there's no uptake of MMX, as it will cost them a lot.

No, since MMX is integral to their processors, the importance of it's uptake is just as big or as small as you claim it is to Apple. In Intel's case, the cost is switched over to the chip purchasers. In Motorolas case (the G4 manufacturer) case, the cost is just switched over to Apple.BTW, you argue that the uptake of MMX is critical to Intel. Then how come Intel didn't release a development kit that would allow you to compile C or C++ for MMX? IMO, this is critical if you want to get acceptance for a new product.I must conclude that MMX was merely a marketing stunt for Intel. MMX was only supported through ASM, which is about as bad as it can get if you're a serious coder, and it was thoroughly hyped, as discussed earlier.QED.

DrPizza · Jan 25, 2000

SSE is quite different from MMX. It's for FP, not integer.The age of MMX does make a difference, because when it was released it didn't suck. <BLOCKQUOTE>quote:<HR>ARRRRRGH! GOD DAMMIT, Peter, will you PLEASE listen to what I'm saying?!? I'm refuting the claim that my post was an attack at the PC! I say it wasn't and that IT WAS NOT intended that way<HR></BLOCKQUOTE> The two things cannot be separated.Intel started out earlier, and so had to suffer lower transistor densities, and so on. There were restrictions imposed on MMX that were very much relaxed for both SSE and Altivec. The age does matter. Intel couldn't have released an MMX that was identical to what Altivec now is, for instance. They wouldn't have been able to make the processors. There are reasons for why MMX works the way it does.

DJ Supreme · Jan 25, 2000

<BLOCKQUOTE>quote:<HR>Originally posted by PeterB: SSE is quite different from MMX. It's for FP, not integer.I know that. Still, SSE is a superset of MMX.>The age of MMX does make a difference, because when it was released it didn't suck.Yes it did. Didn't you read what I wrote? Is it not appearent to you what technical shortcuts Intel took?

ARRRRRGH! GOD DAMMIT, Peter, will you PLEASE listen to what I'm saying?!? I'm refuting the claim that my post was an attack at the PC! I say it wasn't and that IT WAS NOT intended that way<HR></BLOCKQUOTE> The two things cannot be separated.Intel started out earlier, and so had to suffer lower transistor densities, and so on. There were restrictions imposed on MMX that were very much relaxed for both SSE and Altivec. The age does matter. Intel couldn't have released an MMX that was identical to what Altivec now is, for instance. They wouldn't have been able to make the processors. There are reasons for why MMX works the way it does.

What you're writing is NO EXCUSE WHATSOEVER for not making MMX better than it was. Intel should have given MMX a seperate register file (for one thing).

Detnap · Jan 25, 2000

<BLOCKQUOTE>quote:<HR>This it completely incorrect. This is an archaic and outdated definition of RISC being employed here. RISC today is defined not by the instruction count (ex. a G3 and P3 is roughly equal in the respect), but the instruction TYPE count, such as add, mul, permute etc.<HR></BLOCKQUOTE>everybody today has their own cute definition of risc, can you tell me why your definition is better than the other? (ie, do you have a source saying that one definition is outdated?)also, about mmx, i think the needs back then (1997) were different than today. back then, very few people cared about fp (not to mention that the pentium fp wasn't so hot to begin with). we needed to run stuff like windows, powerpoint, multimedia(video) faster and mmx didn't seem like such a bad idea.also, in your original "MMX SUCKS", you said that mmx didn't have a lot of apps written with it in mind. but i'm sure that's not correct. there's a lot of stuff, such as direct x and windows 98 and.. well, i don't know, i don't have a list, but it's no longer a "buzzword" so developers don't necessarily put it on the cover anymore.also, a couple more things. can you please fix the quoting in your articles?and lastly, did you know that you can reply to a thread in progress and you don't need to create a new thread? just hit the reply button at the bottom.Ted

DrPizza · Jan 25, 2000

Yeah, they should have given it separate registers, to make a processor that they got lower yields on that was hotter and bigger, that was too big for the packaging they wanted to use.Whatever

JB · Jan 25, 2000

Some random thoughts: Apple will always have the advantage in this area. They can fully integrate a feature like this with it's OS. The fact that MS hasn't done this with MMX/ SSE speaks to the size of Intel in the industry (hard to assimilate a fellow borg). The fact they haven't done this with 3D Now! speaks to the size of AMD in the industry ( T-Rex's don't munch on small prey) and after all there is always the most magnificent DirectX (tongue planted firmly in cheek). AV is a must for Apple. I'm sure there is a strong desire to shore up the base: the graphics industry. NT has made significant in-roads due more to non-SIMD type features and apps that are NT only (3D Max, Maya, etc...). But lets face it MMX/SSE/3D Now/AV are not necessarily features that make an impact on the average comp user, so how much it will help Apple expand it's market-share is debatable. My guess is that it will be less than MMX's impact, but will help Apple (along with OSX) maintain their base. It's interesting that with AV Apple has done two things that the WinTel world has done. One is to create a CPU centered feature that enhances application (MMX) performance, the whole concept was fully derided by the world of Mac users. To find them now crowing about AV is, well, deliciously ironic. MS finally caught up with the idea of the web in 97 and tied together the OS and the browser in enhance web/system performance (albeit with a bumpy start, but they slowly have come to near realization of this model three years later). And what happens in the Apple world with the G4? Integration of the OS and a feature that enhances application and system performance. Maybe at the next major get together the leadership of MS and Intel can stand before their adoring masses and proclaim that "it's okay, it's okay" for Apple to copy their ideas?

Paul Hill · Jan 25, 2000

<BLOCKQUOTE>quote:<HR>Originally posted by JB: Apple will always have the advantage in this area. They can fully integrate a feature like this with it's OS. The fact that MS hasn't done this with MMX/ SSE speaks to the size of Intel in the industry (hard to assimilate a fellow borg). The fact they haven't done this with 3D Now! speaks to the size of AMD in the industry ( T-Rex's don't munch on small prey) and after all there is always the most magnificent DirectX (tongue planted firmly in cheek). <HR></BLOCKQUOTE>Eh? DirectX has been MMX, SIMD and 3DNow! compatible for over a year now, and DirectX is part of Windows last time I looked.I like MMX. It's getting on a bit, and I like 3DNow! better, but MMX isn't doing any harm (except for taking up all that die space), and I'm pretty sure there wouldn't be SIMD, 3dNow! or, yes, AltiVec, if Intel hadn't made a big deal about MMX back at launch time. C'est la vie, eh?

Ken Fisher · Jan 25, 2000

DJS,Stop creating threads just because you have a reply.This is silly, and it should be in the main thread.

JB · Jan 25, 2000

All are hacks. They attempt to fill in the gaps in the CPU architecture. AV is a better implemented hack. One path as opposed to two (directX/MMX) How revalent will MMX/SSE be when Intel's EPIC CPU arrives? My guess; not much.

Easy Rhino · Jan 25, 2000

DJS, instead of whining around here, why don't you go fix the myriad of technical errors on your web site?

Dan · Jan 25, 2000

Heh! Word.

dent · Jan 25, 2000

<BLOCKQUOTE>quote:<HR>One is to create a CPU centered feature that enhances application (MMX) performance, the whole concept was fully derided by the world of Mac users.<HR></BLOCKQUOTE>This is all from memory, but as a Mac user at the time, I was under the impression that Intel rushed MMX out the door (and started the marketing blitz) to help further differentiate their products from x86 chip cloners. Further, at the time, the only thing that Intel could show as an MMX-derived improvement was one or two Photoshop filters sped up my 4x or something. (And yes, this is all that AltiVec could show when it debuted as well.) Before MMX, the Pentium debut was a relative disappointment performance-wise compared to high-end 486 systems of the time (anyone remember the P75?) so it was hard to get excited about "it'll be REALLY fast in the future, trust us!". It turns out that it DID get faster as time went on, but the initial marketing hype was shown to be unfounded. Somewhere in the same timeframe was the whole Pentium FP bug, but I don't recall the exact timeline. I don't think it was unreasonable to be skeptical (and as further details came out, skepticism of MMX was in fact warranted.)As for deriding MMX for adding CPU functionality to benefit specific app performance, at the time my understanding was that software-based modems was one of the touted examples of MMX benefit. The Mac world had just recently experienced that particular evil and it certainly seemed silly to me that WinTel would want to not learn from Apple's mistakes in that area. On the other hand, even if it was just a speedup of a few Photoshop filters, that was still a potentially significant benefit in my book (since I happen to work in graphic arts).As I said, this is all from memory, plus I don't consider myself a particularly rabid Mac user either. Perhaps you remember some of these same details; maybe I'm just mixed up on the timing (it was a while ago after all).-----On the topic of AltiVec moving the smarts back into the hardware: I may be mistaken, but I thought it was the job of the compiler (i.e. software) to reorder instructions so they could be run through the vector unit rather than trying to do this reordering on the fly in hardware? Isn't that similar in concept to what Transmeta is doing? (And wasn't that the big deal over Motorola's C compiler being ready at the launch of AltiVec, unlike compiler tools to support MMX at MMX's launch, since it helped minimize the delay before end-users would start to see apps compiled to take advantage of the new chip features?) I would have thought that still would fall into the "smarter software" camp.

DrPizza · Jan 25, 2000

AFAIk, Moto's compiler has limited support for automatically using Altivec (it'll use its large registers for some things, that kind of thing) but as for actually using the Altivec unit proper, you have to use special instructions. The difference is, these instructions are in C/C++, rather than assembler. It's moot, now, IMO, because Intel have a similar compiler for C/C++ and Fortran.Actually, the P66 offered huge performance gains over the then king, the 486DX266. It was amazing that this processor -- running at the same clockspeed -- could beat the pants of the 486.And as for Intel trying to differentiate themselves from other x86 vendors, that seems a bit suspect since they licensed MMX to other vendors pretty damn quickly, IIRC.

DrPizza · Jan 25, 2000

Oh, and softmodems really aren't that bad. For probably 90% of people (if not more) they're ideal. Cheap and effective.Things like softmodems were all part of Intel's NSP initiative, which _they_ claim MS tried to kill (tho' I can't quite see how MS could do any such thing. Never mind).[This message has been edited by PeterB (edited January 25, 2000).]

JB · Jan 25, 2000

From the September review of the G4 by Firingsquad: "On a side note: The irony of Apple adding SIMD instructions should be noted by PC and Mac fans alike. Several years ago, when Intel added MMX to its Pentiums, Mac zealots worldwide started screaming about the complexity MMX represented and how it outlined the differing design philosophies of CISC and RISC. Now, several years later, Apple is touting the much more complex "Velocity Engine" as the biggest thing since the transistor, and Mac fans on the net are gobbling it up with little sign of ironic acknowledgement. Also, Intel is now making chips that internally run RISC instructions, and the Merced is a huge jump in that direction. It goes to show you: some things never change, and sometimes compromise is the best engineering decision." Here's the page: http://www.firingsquad.com/hardware/g4processor/page5.asp Note the use of the word: "screaming". I can vouch for the legitimacy of it's use. Like many other Window users, I heard a lot of that screaming.

Evil_Merlin · Jan 25, 2000

JB,Um since Windows95 OSR1 MMX has been part of Windows, and it has been included in the Win32 API since that time.SMID, 3dNow and MMX are part of Windows98 SE and Windows 2000, and have been since their inception.

JB · Jan 25, 2000

A hack is a hack no matter how well you slice it. Maybe you can point me to all those Windows features that take advantage of MMX that don't go through DirectX?

Zer · Jan 25, 2000

Things that I can think of that use MMX... Custom programming things that you do, and, hmm... MoveMemory, CopyMemory, FillMemory, ZeroMemory. I'm sure there's more but basic memory features can...

Taifun · Jan 25, 2000

<BLOCKQUOTE>quote:<HR>Such as permuting? It's not complicated, it's just a lot of simple things done a lot of times<HR></BLOCKQUOTE>Gee. By this definition every processor is simple. <BLOCKQUOTE>quote:<HR>Intel should have given MMX a seperate register file (for one thing).<HR></BLOCKQUOTE>Between speculative execution and direct load forwarding a lot of the effect of having few registers is negated. Having fewer registers saves die area. [Edit - misread your statement to be the same stupid few registers argument you brought up last time. Using the FPU and MMX unit at the same time is relatively rare, and you save die area and complexity by doing it this way]Why would you compare AV to MMX when SSE and 3dnow are around? Might as well compare a K7 to a G3 and point out how much the G3 sucks.[This message has been edited by Taifun (edited January 25, 2000).]

Venture · Jan 26, 2000

I have heard all the arguments that Altivec is superior to MMX and SSE.If this is the case, how come the independent benchmarks appear to show so little difference?Could it be - gasp - that for the sort of things that MMX and SSE do, Altivec doesn't do them any better?

DJ Supreme · Jan 26, 2000

<BLOCKQUOTE>quote:<HR>Originally posted by Caesar: DJS,Stop creating threads just because you have a reply.This is silly, and it should be in the main thread.<HR></BLOCKQUOTE>As on USENET, most threads have a bad tendency to go off topic once they pass X number of replies. On USENET, I usually count 5 replies. Here, it's usually 1 or 2 pages. After that, the main points drown in different discussions and personal flames.Besides, I've created 2 new thread based on replies. That is all. It that what you're getting so worked up about, or..?

DJ Supreme · Jan 26, 2000

<BLOCKQUOTE>quote:<HR>Originally posted by Easy Rhino: DJS, instead of whining around here, why don't you go fix the myriad of technical errors on your web site?<HR></BLOCKQUOTE>Tell me what "these myriads" of erros are. So far you've only quoted a simple capitalization error in the news section.As before - mail me any corrections, don't post veiled flames about it here.

hegor · Jan 26, 2000

MMX is mostly hype with little innovation. I wonder if it was cooked up in Intel's marketing dept? MMX does suck. They probably spent more money on those dancing clean suit people than its development.Hegor

DJ Supreme · Jan 26, 2000

Taifun:quote: -------------------------------------------------------------------------------- Such as permuting? It's not complicated, it's just a lot of simple things done a lot of times -------------------------------------------------------------------------------->Gee. By this definition every processor is simple.No, because the individual elements or atomic operations of the AltiVec instructions are uniform and merely performed many times. With permutations being an example.As I mentioned earlier, x86 (or CISC CPU's in general)'s mixing of adressing and arithmetic is more complicated and require the attention of different units inside the CPU (such as: "get the contents of r2, add 3 and writeback to r3). quote: -------------------------------------------------------------------------------- Intel should have given MMX a seperate register file (for one thing). -------------------------------------------------------------------------------->Between speculative execution and direct load forwarding a lot of the effect of having few registers is negated. Having fewer registers saves die area. [Edit - misread your statement to be the same stupid few registers argument you brought up last time. Using the FPU and MMX unit at the same time is relatively rare, and you save die area and complexity by doing it this way]In today's world of preemtive multitasking, you can't be sure of a case of MMX or FPU exclusive use. Still, it's not a stupid argument. It's very relevant - those flush stalls are real.>Why would you compare AV to MMX when SSE and 3dnow are around? Might as well compare a K7 to a G3 and point out how much the G3 sucks.That's a low blow. You can only claim that the G3 "sucks" if you are completely Mhz-centric, which most PC users unfortunately tend to be.

DJ Supreme · Jan 26, 2000

<BLOCKQUOTE>quote:<HR>Originally posted by Venture: I have heard all the arguments that Altivec is superior to MMX and SSE.If this is the case, how come the independent benchmarks appear to show so little difference?Could it be - gasp - that for the sort of things that MMX and SSE do, Altivec doesn't do them any better?<HR></BLOCKQUOTE>Sorry, Venture. In RC-5, a 450 Mhz G4 beats a 1 GHz-Athlon. Resulta available at the relevant RC-5 sites.When I tested Photoshop 5.5 (w/ the AV plugin) on a G4 Mac, it became 2 - 4 times faster, depending on the operation. Considering that the MMX and SSE/KNI/MMX-2 (whatever!) instructions only delivered relatively modest increases in performance, I will call AV prettly superior to both.

Venture · Jan 26, 2000

DJ said:"Sorry, Venture. In RC-5, a 450 Mhz G4 beats a 1 GHz-Athlon. Resulta available at the relevant RC-5 sites.When I tested Photoshop 5.5 (w/ the AV plugin) on a G4 Mac, it became 2 - 4 times faster, depending on the operation. Considering that the MMX and SSE/KNI/MMX-2 (whatever!) instructions only delivered relatively modest increases in performance, I will call AV prettly superior to both."How does the G4 stack up to a G3 in RC-5? That should show what effect Altivec has on performance. As to Photoshop, the G4/450 results in Photoshop showed a nine percent improvement over a PIII/500. This was on Lighting Effects, which uses SSE rather than plain MMX. The PIII/500 showed a 2-3 times improvement over the Athlon 600, which has a better FPU but no equivalent to SSE.

Dan · Jan 26, 2000

<BLOCKQUOTE>quote:<HR>Originally posted by hegor: MMX is mostly hype with little innovation. I wonder if it was cooked up in Intel's marketing dept? MMX does suck. They probably spent more money on those dancing clean suit people than its development.Hegor<HR></BLOCKQUOTE> Sort of like Apple's marketing dept. eh? And you spelled sux wrong. View image: /infopop/emoticons\icon_wink.gif

Taifun · Jan 26, 2000

<BLOCKQUOTE>quote:<HR>In today's world of preemtive multitasking, you can't be sure of a case of MMX or FPU exclusive use. Still, it's not a stupid argument. It's very relevant - those flush stalls are real.<HR></BLOCKQUOTE>Context switches are expensive anyways. <BLOCKQUOTE>quote:<HR>That's a low blow. You can only claim that the G3 "sucks" if you are completely Mhz-centric, which most PC users unfortunately tend to be.<HR></BLOCKQUOTE>See the N way vs SMP thread for a long argument on this. The G4 is only competitive if you normalize for Mhz, which is stupid. The G3 isn't even competitive when you do that.

Venture · Jan 26, 2000

"You can only claim that the G3 "sucks" if you are completely Mhz-centric, which most PC users unfortunately tend to be."And I wonder why?Maybe it's because no matter how good certain chips may be, at the end of the day a sufficient disadvantage in MHz will render that moot, and from there on extra MHz is all advantage.As of now I can't buy a Mac faster than 450 MHz but I can buy a PC at almost twice that. Your inference is that, with all other things being equal, the Mac is at least almost twice as efficient as a PC.When the day dawns when we have 800MHz G4+s, we will also have 1 MHz Willamettes (and also Merceds). No doubt people will then be talking about how a 1 MHz G4+ - if it existed - would beat a 1 MHz Willamette.MHz is not the last word in determining how a chip performs, but claiming that PC users are "MHz-centric" implies that MHz can be ignored or at least placed low down on the scale of things to condsider.

Crash&Burn · Jan 27, 2000

Yo DJ!! Shut the fuck up and sit down already!!!!! I've had to watch this dumb-arse shit spewing from you on more than one forum. I would be the first person to admit PC's leave a good deal to desired but your fancy colored "super computers" aren't perfect either.... Mac's are good for graphics and I've known this for years before I got into computers but if I really wanted to do serious graphics I'd just buy an Alpha or SGI station and leave the PC and Mac world behind playing games and surfing the net where it should be. So like I had mentioned earlier stop bitching like the 12 year old you are and try to gain an IQ above 20!!!!!

Slagenthor · Jan 27, 2000

DJ:Good gawd quit whining about the register space.... View image: /infopop/emoticons\icon_wink.gifOne of the primary reasons MMX was designed to share the x87 register space was in order to insure compatability with non-MMX aware apps.Proper coding almost completely negates any performance penalties that the register-starved x86 suffers from, because direct load forwarding takes 0 clocks.Also, I see you mention that the "context switches" between MMX and x87 are devatating to performance, would you care to quantify that??? hmmmm?Like, how many clock penalties for context swicthes, and what sort of appliation performance hit do they cause???

"MMX sucks" part II

Smack-Fu Master, in training

Well-known member

Ars Tribunus Militum

Well-known member

Smack-Fu Master, in training

Ars Tribunus Angusticlavius

Smack-Fu Master, in training

Smack-Fu Master, in training

Well-known member

Smack-Fu Master, in training

Well-known member

Smack-Fu Master, in training

Ars Tribunus Militum

Well-known member

Ars Praefectus

Ars Legatus Legionis

Founder & Editor-in-Chief

Ars Praefectus

Ars Scholae Palatinae

Ars Tribunus Angusticlavius

Smack-Fu Master, in training

Well-known member

Well-known member

Ars Praefectus

Ars Legatus Legionis

Ars Praefectus

Ars Tribunus Militum

Ars Praetorian

Ars Legatus Legionis

Smack-Fu Master, in training

Smack-Fu Master, in training

Ars Centurion

Smack-Fu Master, in training

Smack-Fu Master, in training

Ars Legatus Legionis

Ars Tribunus Angusticlavius

Ars Praetorian

Ars Legatus Legionis

Ars Scholae Palatinae

Ars Centurion