Please create an account to participate in the Slashdot moderation system


Forgot your password?

More AMD K7 Details 77

arbustus writes "If you're wondering what's going on with the AMD K7, check out PCVelocitiy's preview. They go over the K7's decoders/IEU's, FPU, bus speed, cache, etc. "
This discussion has been archived. No new comments can be posted.

More AMD K7 Details

Comments Filter:
  • by Phroggy ( 441 )
    Yeah, that's what I was thinking. Mac OS 8.6 is supposed to do SMP, as well as (of course) Mac OS X (I don't remember if Mac OS X Server does or not). I bet LinuxPPC would be pretty happy on a 750MHz PowerPC G4...
  • I think they said it's physically identical to a Slot 1 but electrically incompatible, so that would mean very definitely not. Remember, the motherboard is licenced from DEC and runs at 200MHz...
  • I've got a Zenith 286-8 that doesn't have a motherboard, just a backplane. The normal motherboard stuff is on a pair of expansion cards... one with the 286, 287 (yup, got a math coprocessor... the thing'll eat a 486SX-25 for lunch doing CAD work), base 512k system RAM, and supporting circuitry. The other's got ttyS0 and lp0 and, I think, the stuff that's handled by the chipset on modern motherboards. It won't run without the second card, at any rate.

    They aren't ISA cards, though... they're some funky non-standard Zenith local-bus format. The 8-bit tab is normal ISA, but the second, 16-bit tab is fully as wide as the first tab. The backplane's got four special slots on it that can take either these cards (I've got a couple of RAM cards that use the same format) or standard ISA.

    I've never tried putting one in another machine, for the obvious reason that they just don't fit, but I did try booting it once with two CPU cards in it. No dice... the first POST light wouldn't even come on. I'd suspect that any similar attempt would run into trouble with the CPUs getting in fights over the bus and suchlike, unless the CPU cards were actually designed for SMP.

    More on topic... I've got of those CPU cards here in my hand as I type (yes, typing one-handed)... the CPU has an AMD logo on it, and underneath that, "(c) INTEL", which I find tremendously interesting...
  • Posted by Nr9:

    multiple cores not execution units
  • Posted by Nr9:

    nop, the 1 meg xeons are ppros with mmx
    pIIs are lower
  • Posted by Nr9:

    imo just the opposite cuz G4 dies are much smaller and they've got better fabs(ibm) and there will be multiple core G4s(G2K?)
  • Posted by Nr9:

    anything competes against the xeon? whtts so good about xeon?
  • SSE requires an entirely new processor mode, not just a mode switch. Think "Kernel Level Support Required".
  • I kind of agree with you but it's still an alternative and while AMD's products aren't intel killers they are very solid.

    Ultimately, in a free market, AMD will offer processors with almost identical performance at an almost identical price. AMD has enough credibility to do that and then it becomes a different matter, who includes chip id? who's chipset uses rambus memory? Who uses closed chipset hardware which increases the over all cost? Who will produce another chip I can put in my current motherboard? (Intel has burned me with the Socket8 and the PII slot now...) Who supports SMP?

    Intel is getting lazy anymore, the PII is really just a PPro in a differ box with MMX. The PPro wasn't that much better than the pentium. The P3 is humorous. If AMD can cleanly beat Intel with the k7 it will be good for the industry. I'm not holding my breath though because it will probably be a draw...

  • How exactly are the PII's lower? They've got the exact same core as the PPro.
  • Umm... it's called a G4.
  • Alpha does it by interleaving the memory. (kinda like raid0 for memory).

    AFAIK it interleaves across multiple memory banks. it might also interleave across the memory modules in each bank as well, i'm not sure.

    21164 has a 128bit bus width, so needs 2 dimm's or 4 simms in a memory bank. The UX board has 6 dimm sockets -> 3 banks, and there is a noticable increase in performance as you fill up each bank.

    the 21264 has 4 dimm's in a bank (256 bit bus), so the potential for bandwidth gains via interleaving are even greater.

    so effectively memory interleaving pipelines memory requests from the high-speed host bus and spreads them across multiple banks/modules of memory on a slower bus.

    latency is poor, but you do get a huge increase in bandwidth for large memory access.

    if this is what they're going to do with K7, it will mean you'll need to install your memory in pairs (or even quads for high-end boards), like you used to have to do with simm's on pentium machines.
  • Actually, latency would be better with 256bit wide bus vs 64bit wide bus.

    uhmm.. width of a bus has nothing to do with latency..

    Since the time to get the first piece of data is the identical, but with a 256 bit bus you get 4 times as much data in the same amount of time.

    if the time is identical then the latency is identical. and relatively speaking a 2GB/s bus with a latency of x ms has a poorer latency than a 1GB/s bus with latency of x ms. (ie the latency has not scaled with the bandwidth).

    what the interleaving does do is to reduce wait states during the data transfer, thereby giving a higher bandwidth.

    so you can't use the 66MHz memory clock to calculate the bandwidth, because then you're ignoring the effect of interleaving. and that effect is that memory access is multiplexed from the 4 64bit 66MHz dimm's onto a 200MHz 256-bit EV6 bus..

    256bit @ 200MHz = 6.4GB/s theoretical peak rate.

    which is something else...

    although i'm not sure whether the 21264 really has a 200MHz host bus.. could be 100MHz which would equate to 3.2GB/s peak memory bandwidth.
  • Everything I have seen says the instruction sets are functionally equivalent. Basically anything that can be done with one set of instructions can be done with the other. The only real functional difference is that KNI can also operate on 2 double precision floats not just 4 single precision floats.
  • Yes, but unlike Intel anyone will be able to make a processor with slot A. And, the licensing fee, if any, won't be stifling. It would be interesting to see Rise or IDT make EV-6 based chips.
  • The only difference with the K7 is that the hype is not coming from AMD. So, far the only things AMD has officially said about the K& is the basic architecture at Microprocessor forum last October, and the demos at Comdex and Cebit. I think they have learned a lesson about overpromising and underdelivering. So, right now they are basically lettign the device and architecture speak for itself. I do think a lot of people outside of AMD are hyping up the K7 a lot more than tey shoud lat this point. Basically giving free advertising.
  • No it is a doubling of bus speed. The EV-6 bus is a 64bit wide bus, that in the K7 incarnation will run at 200MHZ. But, since the Ev-6 is point-to-point between the chipset and CPU. The FSB speed only tels you that the maximum memory bandwidth to the CPU is 1.6GB/s. But, aside from that, it has no implications about the actual memory bus width or speed. In fact the memory could easily be 64bit 100MHZ SDRAM. So the memory bandwidth would be 800MB/s, but the FSB bandwidth would still be 1.6GB/s. That is an unlikely scenario except in the possibility of a cheap K7 motherboard and chipset. The more likely options will probably be a 128bit wide 100MHZ SDRAM bus, 64bit wide DDRSDRAM, or 800MHZ RAMBUS.

    But, the point is that 200MHZ FSB tells us nothing about the memory speed when it comes to an EV-6 bus.
  • I think you are talking about the Win95 bug with K6-2 over 350MHZ. This is a result of a timing loop in Windows95 that goes too fast on a K6 at or over 350MHZ. There is a patch for Win95 on this. I don't think the problem exists in Win98. Note this is a MS Windows bug not a hardware problem with the K6.
  • the 21264 has 4 dimm's in a bank (256 bit bus), so the potential for bandwidth gains via interleaving are even greater.

    I believe that they also use 66MHZ ECC SDRAM for this. Giving a bandwidth number around 2GB/s.

    latency is poor, but you do get a huge increase in bandwidth for large memory access.

    Actually, latency would be better with 256bit wide bus vs 64bit wide bus. Since the time to get the first piece of data is the identical, but with a 256 bit bus you get 4 times as much data in the same amount of time.

    But, as I said previously memory configuration is chipset dependent, and therefore can be just about anything.
  • The CPU bus is still 200MHZ*64bits if this is true. Something I did not think of though was what if you have 2 CPUs. Each of which has a 200MHZ*64bit bus. Aggregate bandwidtht hat those tow CPUs could use would be 3.2GB/s, and 200MHZ*128bits would be a reasonable design in order to keep 2 CPUs fed.
  • This kind of backward-compatible technology might solve some of the problems mentioned in this
    thread. If the K7 takes off, it'll be only a matter of time before somebody does something similar to the Spectra for it, so any investment one makes in computers now isn't necessarily lost when one wants to use the next generation of CPU.

    Actually, this probably will not happen witht he K7 since it is a total break with previous technology except for the instruction set. The buses are too different for there to be an S7 upgrade for the K7. With the various K6s the issue wasn't pinouts or bus protocols, but primaily voltages and multipliers, so making upgrades isn't too difficult.
  • Well, I remember Microway used to do a transputer card for the jolly old PC that would give you a MegaFLOP, and you could have multiple ones, I think. They did compilers.

    Those were the days. I was still on a Spectrum 128 with a Microdrive in those days...

  • I've just ordered a Spectra 333 from Evergreen as a quick way of upgrading a legacy P100. It's basically a 333MHz K6-2 on a module that fits into the Pentium socket 7 on non-K6-compatible motherboards. Very slick, and a good alternative to swapping in an entirely new motherboard which sometimes just isn't possible, eg. on industrial PICMG or embedded cards.

    This kind of backward-compatible technology might solve some of the problems mentioned in this thread. If the K7 takes off, it'll be only a matter of time before somebody does something similar to the Spectra for it, so any investment one makes in computers now isn't necessarily lost when one wants to use the next generation of CPU.

    If the Spectra turns out bad, I'll post a problem report followup here. [It seemed worth taking the chance as Evergreen stake their business on compatibility ... I hope. :-)]
  • Hmmm, it could be on topic if the alleged problem is not acknowledged by AMD and makes it into the K7.

    Please give URLs + Usenet references where those problems have been discussed, to avoid starting a rumour without foundation.
  • Thanks Dastardly, that had me worried for a few minutes. False alarm then.

  • Only one fab plant? I can see the headline now.

    "AP: AMD only chip making plant mysteriously exploded only two weeks before the release of the K-7. The only clue the FBI has found so far is a scrap of what is believed to be the leather briefcase the bomb was in, bradishing a shiny metalic sticker on it. FBI director Guy Smilely said at the scene, 'The burnt sticker is printed with an insignia, some kind of swirl and two uninteligable words. We ask that anyone information about this symbol come forward.' Baffled authorities are searching for any clues leading to the the swirl gang."

    No I don't think the people at Intel are really terrorists. I'm just dissapointed that their number one competetor is so fragile. In fairness I have a dual PII and love it, but you won't catch me dead with a PIII.

  • the most important thing about the k7 is that it's the first chip that will compete against the xeon. That's where all the price markup is and the K7 will bring it down- and intel will be hurt by it.
  • It exists!

    The K6 *has* multiple (can't remember if it's 4 or 6) execution units. Yes, the plain old K6.

    ...I think.

    /* Steinar */
  • Here's an idea:

    Why not do SMP on a single chip? Have a single instruction decoder, with a cache on both sides so that instructions (hopefully) need to be decoded only once. Then have several execution units, which would be bunched together into superscalar 'CPUs' of two or three execution units. There would be four of these 'CPUs' on a single chip. Write a Linux kernel extension to take advantage of this SMP, sell the new chip for about the same as Intel's most expensive Pentium (but with probably twice the performance), and watch the money roll in!

    (OK, of course it's not that simple. What is wrong with my idea?)
  • The K6 may have multiple execution units, but to the programmer it appears as a single CPU, programmed sequentially. I am suggesting running groups of execution units as separate 'virtual CPUs', which could run different processes or threads in the same way that different CPUs run different threads in a normal SMP box. I think this would get better performance, because there must be diminishing returns from the number of execution units you can put in a normal superscalar CPU.
  • It's refreshing to see someone with a sense of humour.
  • AMD has always been playing catch-up. Here, they have a chance to take the lead and get a jump on Intel.

    The previous lines of AMD chips also had crappy FP compared to Intel. 3DNow was cool, when you could find software that was written to take advantage of the instructions (good luck). The K7, on the other hand, looks like it has damn good normal FP, and even better MMX/3DNow FP.

    I thought that this review was pretty good. They never said that it was going to kill Intel or anything, just that the K7 might be a temporary leapfrog over Intel until Intel gets its seventh-gen technology out the door. And it'll be cheaper, probably.

    The other thing is that AMD is pushing this new motherboard type. Could be a good thing, could be bad. One hand: faster FSB speeds. Other hand: maybe less companies making Slot A boards in quantity, so they're more expensive.

    Dunno. 'Tis all speculation until we get them in our hot little hands, but I'm optimistic about AMD getting this one right. (Now, let's hope their one-and-only fab doesn't blow up or something)

  • K6 - it's the best thing since sex! it'll rock all over intel.

    K6-2 - it's the best thing since sex! it'll rock all over intel.

    K7 - it's the best thing since sex! it'll rock all over intel.

    I've lost faith in AMD. They promise the world then don't deliver. I'm sure it'll be a decent processor, but nothing special, judging from what AMD has put out in the past.
  • Well, that's a definate maybe. The K7 will be a SLOT and not a socket type chip, at least at first. I heard from a reliable source (Maximum PC Magazine) that AMD will eventually make a Socket7 version of its K7, but things can only go so far.

    Looks like you will have to bite your lip, and wait and see.
  • if AMD gets over the intitial production problems, the K7 will be a SERIOUS blow to Intel. Ouch.

    I'm rooting for AMD all the way.

  • Can it handle PIII special instructions?
    does it have it's own?

    As I understand it, the K7 supports the PII instruction set, but doesn't support the SSE/KNI instructions. Programs that depend on these won't work on the K7 (or on a PII, for that matter).

    The K7 certainly supports 3D-Now, and I remember hearing rumours of an extended version of 3D-Now at one point. I don't remember hearing anything about it since, though.

  • "All processors wait at the same speed"


  • You will only need to get a BIOS upgrade.

  • Actually K7 will only work on MBs *with* 100 MHz FSB. See previous post.

  • Well, whether the G4 or the K7 will be faster should, naturally, depend on what you're doing. A 200 MHz bus on the K7 to the system RAM would certainly speed up certain apps, but the optimized altivec instructions would make other things faster for the G4. (or am I stupid, by not knowing the bus speeds for the G4?)

    Other important considerations will be AMD's track record for poor yields (resulting in lower clock speeds), and the tendency for Motorola PowerPC chips to be available to the average person only in a Mac (limited choice and competition for motherboards). Although, at this time, we don't know how many motherboards will be out for AMD's new Slot A line, so motherboard selection may not be an advantage for the K7.

    All and all, it's too soon to tell. AMD boosters (like me) are hoping for Sharptooth to tear up the competiton, but I've also been having an occasional fleeting thought about running linux on one of those G4s.
  • by bwz ( 13374 )
    KNI != 3DNow!. They are very similar but the opcodes are different so they're not compatible.


    Has it ever occurred to you that God might be a committee?
  • by bwz ( 13374 )
    Sorry if I was ambigous, I was trying to point out that a KNI binary won't function on a 3DNow! chip. IMO that makes KNI != 3DNow!.


    Has it ever occurred to you that God might be a committee?
  • Because US DoD demanded more than one supplier of "strategic materials" or somesuch (that's the way I heard it anyway). Then they (US DoD) figured that the x86 line wasn't so important anyway and slackend the constraints.. And thats why Intel did the 486SX - About as cheap as Cyrixs and AMDs 386es and about as fast but "superior" (rumor has it that it was the same chip with a pin removed and that the 487 also was the same chip but with another pin removed) :-/ :-/.... And after that day AMD and Cyrix have been playing catch-up, I'm quite happy to se that they're closer today than they were a few years ago..


    Has it ever occurred to you that God might be a committee?
  • not really, but the this is a catchy subject.
    Besides the "it rules" post,

    Can it handle PIII special instructions?
    does it have it's own?

    Does anyone has pricing and schedualing info?
    maybe a chart of L2 cache - MHz speed?

  • Does anyone know anything about the K6-IIs hardware incompatibility in the 350Mhz range?

  • The floating-point numbers _seem_ particularly impressive, but isn't that peak speed?

    Guess I'll have to wait until someone runs SPECfp or whetstone on on it... my 180MHz PPro is seeming quite sluggish these days :-/

    The article also says that this will be aimed at servers... argh, jack up the price $500... not good.

    I bet it will STILL cost way less than the Xeon :-P

  • You seem to forget that:

    a/ Until the release of the PII, the K6 was the best PC processor out there. There was really no competition between it and the original Pentium-- half the price and faster.

    b/ The K6-2 IS superior to the Pentium II-- within its niche. I admit that software optimized for 3DNow isn't as widespread as it should be, but my $100 K6-2 300 gets better frame rates using 3DNow than a PII 450 ($$$)
  • Hmm....well my 8088/8086 was on a card. ISA as a matter of fact...hmm, I wonder what would happen if I plugged it in to one of the free slots on 686 system...A computer running inside a computer??

  • nope accually it was an AMDEK. Mine was ISA sized though, I guess I could plug it into my 686 mobo if I wanted to, but considering you people think it might blow up or something I guess I better not. I'm not sure who made the processor in this bad-boy, since I'm not sure which chip on here is even the processor in the first place. I found the BIOS because I had to switch it out to be able to use CGA , but the rest of the chips all look alike.
  • Not the processors that crack RC5! 8^)

    / K-DaK
  • K6-2 - it's the best thing since sex! it'll rock all over intel.

    But I think everyone knew the K63D was going to be a dud -- it was pretty well known that nothing had really changed in the core except for the addition of some new instructions and official support for 100MHz FSB.

    The K7 is actually a new product, not a warmed-over K6. Unfortunately, there's those yield problems everyone is so sick of hearing about (ever notice that AMD hasn't had the lead in raw CPU clock -- a clumsy measure of fab yield & quality -- since the 386/40 days?).

    Intel is throwing production technology at the problem of increasing CPU performance: crank up the clock speed and hope nobody notices it's really not all that much faster.

    AMD has been steadily improving the design of their processor (with the exception of the K6-2, which I don't consider terribly impressive), and is applying what they learned from that to the K7.

    If AMD can ship a K7 at a competitive clock speed [face it -- those numbers are what sell CPUs; just ask Intel] it will potentially stomp all over whatever it is Intel has lined up for the year.

    [For that matter, does Intel have anything lined up for the year that isn't reheated P6 cores with uncountably various cache architectures and in strange and incompatible packages?]

    But like everyone else, I'm tired of seeing problems, problems, problems from AMD. If I were a shareholder I would be very unhappy right now. Eventually Intel will actually learn something about x86 engineering from AMD, incorporate that knowledge into a near-generation product, apply their "Fab The Hell Out Of It" strategy, and moosh AMD back into their cage.

    And on that day, I will feel sad.
  • I really doubt it would get to the POST check - if it didn't make some nice sparkles, that is... The circuitry is simply non compatible.
  • Well, my computer has a K6-II and the Pentium II I also use around here are MUCH slower... I became a convinced AMD fan.
  • The Pentium III has the new Katmai instructions which were already in the AMD 3dnow chip (K6). The new K7 should have more features then the PentiumIII
  • I am looking for an honest answer, will this thing blow the crap out of the G4, or will it be the oppostie. Or will they be about the same?

    Also, will this thing actaully come out in June or is this more crap?
  • First of all, the K7 is going to be a Socket 370 chip (basically a Socket 7 with an extra row of pint), and therefore not compatible with any older motherboards.

    Second, it will have a 200MHz bus, and again, not compatible with older boards.

    keep researching

The relative importance of files depends on their cost in terms of the human effort needed to regenerate them. -- T.A. Dolotta