Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
Wbmw:
Are you particularly dense or do you not see the safe dropping on your head? The measured TDP difference is what Dothan and its chipset uses more power than the Turion and its chipset do when being worked hard. They had indentical GPUs and memory, nearly identical disks and screens, so the differences could only be attributed to the CPU + chipset. Assuming Intel's chipset isn't a power hog (and I believe it isn't), then only difference is that AMD's Turion core, on die memory controller and NB is less power hungry than Intel's Dothan and the NB portion of its chipset.
Actually you would hope that the 6-7W difference is being eaten up by the NB, else you have only one place that could have used that much extra power, Dothan itself. Turion's I/O power is 2.9W TDPmax at halt. Active I/O power has a higher amount than that. Intel's NB chip must include FSB power, backbone bus power and the power going to routing logic, memory controllers, arbitration, power control and synchronization logic. The only things not included is the PCI express to the GPU and the SB link. Given that both are point to point low voltage differential serial links (20 IIRC), that doesn't take all that much power compared to the 128 bit DDR(2) interface and the 64 bit FSB. It wouldn't surprise me that 80-90% of all NB power comes from those things equivalent on a Turion die. Having those on 90nm copper SOI versus 130nm aluminum bulk doesn't help Dothan either.
So the difference, by your argument, means that Dothan itself is using 4-5W more than Turion's die and the NB is using 2 more W. So much for Dothan's supposed energy efficiency. To fix the paper comparisons, commonly done by you, we must add 4-5W to Dothan's TDPtyp and add in 2W more for the NB portion. So Dothan's TDP goes from 27W TDPtyp to 33-34W TDPmax. And Yonah's 31W estimated TDPtyp spec must be pushed to 43-45W TDPmax (of course assuming that Intel doesn't lower their TDPtyp standards any more).
The rest of youir estimates are similarly based on the wrong set of numbers. Their family TDP numbers have so much margin that individual members actually use far less than the power specified. Heck 2.4GHz Venice by test uses less than 31W and its Family TDPmax is 67W. Applying the same ratio to your post has AMD 1.8GHz DC puts out only 25.4W vs 45W for 2GHz Yonah. The 2.4GHz Opteron DC (Sun just announced the x80) puts out only 44.0W vs the same 45W for 2GHz Yonah. The 2.8GHz FX-57 puts out 50.9W to the 45W of 2GHz Yonah.
Here is a link showing even higher Dothan 2.13GHz (770) power use: http://www.cpuheat.wz.cz/html/PentiumM.txt
And: http://translate.google.com/translate?hl=en&sl=ja&u=http://pc7.2ch.net/test/read.cgi/jisaku/...
Pete
Dear Mas:
Of course I did. BTW, thats 400/2*333 to 533/2*533.
Pete
Dear Mmoy:
I thought you would see that as a reference to throughput and not latency. With an additional FPadd and FPmul units, you could start a SSE2 DP pair add and a SSE2 DP pair mul each cycle compared to every other cycle as it is now. In addition, the two new FP units would only need to support 64 bit instead of the 80 bit support needed in the current two (x87).
Pete
Wbmw:
Using the much older icc8.0 vs icc9.0 used as the basis for Dothan SPECint2000 speculation. I calculated that 9.0 over 8.0 was at least 5.6% boost. Doing the same to FX-57 and Opteron x54 and you get 1970 * 1.056 = 2080 minimum. Boost to 3GHz and we get to 2230. Add in faster memory (Current Rev E can use DDR533) and later will run with DDR2, likely to get past 2450. K8 will likely clock faster than Conroe. K10 may even be faster yet.
The icc and ifc compiler is optimized for P4/PM. Pathscale at least optimizes for K7/K8. FX-57 gets 2261 in SPECfp2000. I use peak scores since icc and ifc gets base=peak on Intel CPUs. Pushing this to 3GHz and we get to 2399. Add in faster memory to DDR533 and DDR2 too, and we are beyond 2500. K10 could get 2 additional FPU units to do single cycle execution of packed SSE2 instructions. That would take SPECfp2000 to heights where Conroe can't match.
Lastly I go back to history. Digit-life speculations are like theirs for Prescott, badly overblown.
Pete
Wbmw:
Yet with that 1 watt of power difference Yonah will clock 566MHz faster (at first), and the roadmaps eventually have 2.33GHz showing up sometime in the second half of next year (presumably early in H2, since Merom will be launching later in the quarter). That's 733MHz faster than what AMD is promising at 90nm (assuming they can actually deliver).
Its 6-7W difference (and it could be higher given the differences in conditions) when you include the memory controller in the NB. Adding back that power in higher clock, it likely will get to 2.2GHz, faster than Yonah with 64 bit and much faster FP. In 90nm, no less.
Measurements will confirm how close the power dissipation of Yonah is relative to DC K8. My guess is that once reality sets in, you will realize that Intel will have a huge power advantage in the mobile segment next year.
I think you will see Yonah as not being as good as hoped wrt K8.
It's also worth noting that the Intel system lasted 25% longer in the Office application battery life test, which is important to many business users. Wattage wise, the Intel system dissipated 22W system power, while the AMD system dissipated 27W. Now, the Intel system demonstrates 5W better power dissipation for business tasks.
I think what that shows is that the chipset is still an advantage for Intel. Likely the Turion's LCD needs more power to get the same brightness (this was shown in a previous page). So the performance difference is even higher. As to the graphics speed, you better look at the setup on previous pages. They deliberately overclocked the Turion's GPU to the same clocks as Dothan's GPU.
So you have one set of data on the X2, but the same measurement methodology was not used for Dothan, so you do not have an apples to apples comparison, just more wild speculation.
When 2Ghz Dothan is measured using the same setup, it gets higher than the listed 21.5W TDP at 25.4W. The 2.13GHz Dothan was not tested. But if it gets the same ratio as the 2GHz, it would get 31.9W. The only wild speculation is how Yonah would do performance and power wise.
Pete
Wbmw:
Try looking at what VM people like VMware talks about it. They have VMware running on Pacifica (per them) right now. They state that they don't think VT was good enough. Virtual memory acceleration is key for virtualization. To do it in software is slow and quite complex.
This comment though says it all: "Brrr... Digitimes... I wouldn't trust anything out of that." http://www.realworldtech.com/forums/index.cfm?action=detail&PostNum=3577&Thread=11&roomI...
Pete
Dear J3pflynn:
THe article has some glaring errors in the decode section. The 4-1-1 P-III/PM decoder does not issue 6 MOPs per cycle, but only 3. Only one decode path can do complex decodes, but it still issues one MOP per cycle. K7/K8 can decode 3 complex decodes per cycle generating 6 MOPs per cycle (3 executing and 3 load/stores). In K8, a executing MOP paired with no load/store MOP can be combined with a no executing MOP paired with a load/store MOP into a single MOP pair.
So using their terminology K7/K8 has a 4-4-4 decoder generating 2-2-2 MOPs per clock. P6 is supposed to have a 4-4-4-4 decoder, but generating only 1-1-1-1 MOPs per clock.
As to the performance estimates, 2.8GHz K8 already beats the SPECfp2000 score. A 3GHz K8 would likely still outrun Conroe in SPECfp2000. A 3GHz K8 would also likely beat Conroe's SPECint2000 score using the same compiler. Of course K10 may be out at that time with an additonal FPadd and FPmul unit to do SSE2 packed instructions at 1 FPadd_pair and 1 FPmul_pair per cycle or 4 DP flops per cycle. This will likely push K10 far out of Conroe's reach in FP and even exceed Power and Itanium SPECfp2000 scores.
Pete
Wbmw:
VT isn't good enough by those who know what they are doing look it over. Almost all say Pacifia is better and VMWare is already running on Pacifica.
You can give a system with too small memory, run programs that need far larger amount of memory with virtual memory (using hard disk to extend apparent memory). It works, but is dead slow. So a 32 bit VT may work, but performance will leave a lot to be desired (think of those software AMD64 emulators performance for a good example).
In Q4 of this year, 1.6GHz DC Opteron EE (30W TDP) will be released. That already is less power than Yonah. And that includes 3 cHT ports and 2 DDR channels which are not included in Yonah's TDP. In fact this is 30W TDPmax where Yonah's is more like 31W TDPtyp.
In fact in a head to head comparison between 25W TDPmax MT-34 and 27W TDPtyp Dothan 1.86GHz, Turion based system actually used 6W less under high load (gaming) conditions at Tom's Hardware. A64 X2 4600+ was tested at 55W (at VRM input). That is 2.4GHz 512KB DC. 3 speed grades above your x65 HE example. Because the test measures power input to the VRM, it should be reduced by the VRM efficiency (Intel pegs this at 90% when drawing 50-60W). 55W * 0.9 = 50W at 2.4GHz dual core 64 bit running 2 copies of Prime95, a so called "thermal virus" by Intel that will cause their CPUs to exceed their specified TDPs.
We will see what Yonah actually uses using the same test setup. If it acts like all the other Intel CPUs under this test, it will have a far higher power usage than the specified TDP.
Pete
Wbmw:
But Virtualization, which comes out next year, will force 64 bit checkoffs much sooner. And corporate purchasers will worry about it far sooner than you think. Some are worrying about it now. The market can be unforgiving of missing checkoffs. "Are these 64 bit capable laptops?" they'll ask. And the marketeer says, "They will be in a year (maybe)." "See ya next year then, bye!"
Pete
Wbmw:
Engineers use notebooks. There has been one who already asked for a notebook that does workstation applications. Many pointed to Turion and X2 based notebooks (the pricey VooDooPC notebook was mentioned several times). Alas, Intel has nothing in this space as Dothan can't run 64 bit apps. These are being looked at in mobile form so you are wrong!
Workstation grade apps will be used in the mobile environment. They are moving these to x86 servers, workstations, desktops and now, mobiles. Proves the AMD64 everywhere argument. You just don't like it because its a glaring Intel deficiency. Now its a good to have option. Soon it will be required, even if it is normally unused.
Soon virtualization moves from servers down to laptops. Companies can run their corporate software used in their servers on their laptops using the same OS. Saves a lot of retraining. Sorry, 32 bit only mobiles need not apply.
Pete
PS you do know that DC Turion code named Taylor will be out early 2006 (although as AMD has been pulling in schedules, it may show up this year). Now you get 64 bit and dual core in a T&L package. Ahead of Intel, yet again.
Wbmw:
I can think of a few that won't work in a 32 bit environment. But since you can't see them, you don't think they exist. Hint, they need 4-8GB of memory for a single task and that won't run in 32 bit Windows. Yes they're workstation grade apps, but since Windows XP64 hasn't been out long, there isn't many of these. They are available for 64 bit Linux however.
And it only takes one. It will probably be some near reality based video game with ultrarich textures, huge map and lots of details. Yes there may be a 32 bit version with lousy textures, small map and little detail. And those users will get killed easier. And because they will need the greater detail to enter the upper echelons. The latter will be a higher pull for the 64 bit versions.
Lastly there is a rapidly expanding trend which needs 64 bit greatly, virtualization. That will like multiple cores, but first, 64 bit will be a requirement. So a 32 bit dual core will be SOL. Of course you forgot about that. Or conveniently forgot to mention it. It bolsters dual core, but only the 64 bit ones.
Pete
PS, You also forget the huge number of single threaded apps that run faster on the faster SC, and even get a boost from 64 bit, than the slower DC. 30+% performance increase is a lot to make up over the large installed software base. Likely most of the time, Yonah just gets run over by the faster A64 and Turion SCs.
Dear Alan81:
You forget the Turion analysis at Tom's hardware (of all places)! DVD watching required only 19% of the Turion at the lowest clock. The Turion has a much more powerful FPU than a Dothan. It could crunch that spreadsheet much faster at that lowest clock (if the power management scheme sees that spreadsheet at a garbage level priority which mean it can chew idle cycles only at the level needed for the primary app, DVD watching). So you get that sreadsheet crunched with no more power used at all. Besides that recalculate probably will take under a second anyway, far shorter than the period of a DVD movie.
The problem with your arguments is that DVD watching and spreadsheet recalcs are not very demanding jobs anymore. In fact using nVidia GPUs, much of the decode power is provided by the GPU with power saving dedicated hardware. Besides most times the bottleneck is not the CPU, but disks, I/O and/or memory size.
Pete
Dear Mmoy:
My father just retired. He was the leading professor teaching programming to students at MSOE (Milwaukee School Of Engineering). He and I use Linux for the most part. Grub is far easier to use than Windows multiboot. It is quick to setup and allows one to boot an OS even on a damaged filesystem. I use it to recover from many failures on client's systems. It easily fits on a floppy and allows you to boot into any partition on any disk. I also have grub on a bootable CD (just use the grub floppy as the boot image) as many client's systems no longer have floppies.
On the computer I'm currently posted on, I have 16 lines of images to boot. It boots into windows just fine. One is Win95, one is Win98SE, and one is WinXP Pro. The rest are two flavors of Suse, 6 are 3 different Gentoo kernels with 2 different kernel options each and the last one is a RedHat 3.0 AS. I use grub's command mode to boot into CD/DVDs or floppies. I do use a boot partition to have known good copies of the various kernels and images separate from any filesystem just in case. It has saved me a few times.
Pete
Wbmw:
More unfounded assumptions.
You can't find one and thus, there can't be one. There are quite a few under Linux. Many in Solaris too. I can think of a few in Windows. Less than the other two, of course. They were 64 bits far longer. And it shows. Multithreaded applications have been out for years (longer than 64 bits and I was doing mutithreading on a 4P MP/M Z80 box in the late 70's) and even after all of this time, few are available today. Very few programmers do well in going multithreaded. Less than 10% of todays programmers can make efficient multithreaded code and that may be wildly optimistic.
The bulk of todays code is single threaded. The easy multithreaded code is available. OEATLTTMT (obviously easy applications that lend themselves to multi threading) like 3D flow simulations and graphic processing are here and likely run on dedicated hardware, like the latter with GPUs. The low hanging fruit has been plucked.
Even hot games programmers like Carmack will tell you that games speed ups will not likely get beyond 30% or so and that is after years of work. They rather have faster SCs than slower DCs. And 64 bit versions of games are coming out. Far cry is the first beyond the standard ones in Linux.
All it takes is one application that only comes out in a 64 bit version and the gates open up. Everyone without a 64 bit laptop becomes one to be pitied and subsequently ignored. And they will kick themselves for buying a 32 bit only laptop. The reverse will not be true. There won't be a single case where the SC won't run any MT app. They won't be in anyy different position than most are in now. They run slower than the latest and greatest, but it does run. They can be consoled that most of the time their laptop runs faster for the things they normaly do. And it still outruns that DC in games.
Pete
Wbmw:
Another highly unfounded assumption. AMD will have a mobile 64 bit DC. And there are already some 64 bit application uses that can't be done with 32 bit versions. Photographic panorama stiching is one that comes to mind. As to a single core not outperforming a dual core any single threaded app will do that quite hansomely. There are few multithreaded games for exapmle. A faster SC beats a slower DC hands down. There exists plenty of evidence for this.
Also we know of one use where 64 bit CPUs run circles around 32 bit ones, encryption. 5 times speed improvements with some algorithms. There a single core at the same frequency (using half the power) would run circles around that 32 bit DC.
Lastly, don't assume that SC and DC will operate at the same frequency at the same TDP. Evidence shows that AMD loses two speed bins with their DC and Intel loses about three with theirs. So a 2 GHz DC should be compared with a 2.6GHz SC. For most users, the SC will appear to be faster in normal day to day use.
Also a 2.6GHz 64 bit SC will perform, at worst, 2/3 as fast as a 2GHz 32 bit DC in 32 bit apps, However the 2GHz 32 bit DC will perform 0 times as fast as a 2.6GHz 64 bit SC in 64 bit software.
So I would take the 64 bit SC over the 32 bit DC overall. Of course a 64 bit DC would satisfy both our desires. AMD has one, Intel won't. VoodooPC puts one in their DTR gaming laptops. Soon AMD will release a T&L flavor.
Pete
Wbmw:
That is not noted on that graph. Others stated that it refers to the CPUs in their lowest non idle state. There is just so much missing from that graph that states what is being measured and the conditions used therein. You are making assumptions that will likely prove to be untrue.
You can't assume the same conditions most assume from a performance graph. There one assume the best the processor can do. You can't assume that if the object is the most performance per watt of core TDP. That doesn't prevent AMD from taking a Turion MT-40 at 2.2GHz and programming it to naturally run at 2 times the 200MHz HT clock, at the lowest voltage it can be at and running it with PC3200. Its SPECint2000_rate score is 30 to 40% of the standard 2GHz MT-37. However its TDP power is just a watt or two, one fifteenth of the ML37 which is itself much lower TDP than the Mobile Athlon 64 Clawhammer at 2GHz.
You have no proof Intel isn't doing the same. Only your assumptions say different. Show a link to where Intel states what the conditions getting those numbers were. The raw data would satify many of my concerns, but not all. Many of those future cores are likely estimates and we all know how good the initial Prescott TDP and performance estimates were.
It is silly to make so many unfounded assumptions. Even for you.
Pete
Wbmw:
Sure I can. If I make the performance of Yonah at half the clock rate each core will score a little more than half the full rate score. And each half rate core will be at 10-20% less voltage so the power drops to between 35% to 40% per core. Add the two together and you get more performance for less power with the ratio climbing 50% but not much more performance. Surprise, AMD can do this too with its current dual core.
So unless the tests are spelled out exactly with real numbers, they are meaningless. Just look at the Turion. Reduce its clock rate to 1GHz from 2GHz. The TDP drops from 25W to 5W. Take out the on die controller and I/O power to get core power, and the number goes from 21W to 1W. SPECint2000_rate drops to about 55-60% of the score at 2GHz. Whoops, AMD is far ahead because they can boost performance per watt by 23 to 25 times that of its higher speed cousin.
And thats without 65nm, dual core or other tweaks. I guess Intel has to go back and get another 5x improvement just to get to where AMD is now. Thats another two years into the future. You would cry FOUL at this expansion. And you should too at Intel's presentation without the backup data. That graph is meaningless without the raw data used to make it. As it was debunked on other forums.
Pete
Dear Combjelly and Keith:
AMD has stated that they will release a 860 or 865 EE Opteron DC. Thats a 30W TDP. Lose 2 HT ports, registration and a memory channel and you can have a 25W TDP DC mobile at 1.6-1.8GHz on 90nm. AMD also stated that there will be a 15W TDP mobile Turion and a 7W LV Turion for subnotebooks coming out in 2006. I do not recall if these were to be 65nm or mature 90nm CPUs. There is also the Dragonfly (A64 + extras) for embedded apps and handhelds. I do not recall the TDP, speed or process node.
Pete
Wbmw:
It is Intel that uses dirty tricks. It knows that its future roadmap is filled with hope, hand waving and histronics. It, like many other crooks, doesn't like to be caught "red handed". It also doesn't like its tactics to be made public. It wanted to lock in its customers forever. The problem is people hate to be locked in even if it is a gilded cage. It gets worse when people have to overpay to stay there. Soon when they breakout, they leave and won't come back until all cages are destroyed. And they cage their would be jailers.
Pete
Chipguy:
ARM is not raising frequency as fast as when x86 was at those frequencies.
BTW: When 600MHz Alpha 21164s were sold in 1998, Intel was selling 450MHz Xeons. Your recall is faulty. The Alpha did provide much higher performance, but at higher power and transistor counts. 200MHz P6s were sold in 1995. (this is according to www.sandpile.org )
Pete
Chipguy:
No shipping ARM is at 1GHz. 1GHz and up x86s sell 9 digit figures per year. No ARMs sell at 1GHz. Thats 0 digit figure per year. As to embedded cores, 16 bit MCUs and DSPs sell even more than all ARMs. The total processing power though of x86 based systems sold in a year is larger than ARM based ones in the same year. The amount of software for x86 far exceeds that of ARM. And that is worth more than all of both types of CPUs.
Pete
Chipguy:
Perhaps I should be more clear, I said a shipping ARM product at 1GHz. There have been shipping x86 CPUs over 1GHz for years. The first was an Athlon 1GHz CPU.
We had a future 10GHz core (which won't happen for the foreseeable future) and certain cyrocooled AMD64 CPUs in the >4GHz range. And we already have L1 caches running over 3+Ghz on shipping x86 CPUs. What may be available in the future on some as yet unproven process is quite easy to suggest. Performance on paper is easy, real performance on real hardware is hard. Doubly so when sold in the 6 figure range in units per year to the public at large at a price they can afford.
Pete
Chipguy:
Name one ARM core that runs 2GHz. Name one ARM core that does even 1GHz. Compare cores at same clocks instead of some ARM running sub GHz compared to a 2GHz Turion or even PM core. At even the same performance, these latter two can be significantly under volted so that their TDPs become far less than their ratings.
I don't know, do the ARM cores include the DRAM controller? If not, then the claimed TDPs of ARMs must be increased greatly to make even a lousy apples to apples comparison. Now either overclock the ARM to 2GHz (or whatever to perform similarly in SPEC or some other performance measuring benchmark, if it can even get close (I think it will require the Turion or Dothan to slow quite a bit to allow for a match)) and then make some power measurements at roughly the same performance.
Most of those transistors in the x86 designs are to enable them to run at high performance. Even the RISC and VLIW architectures seem to need a lot of transistors to run at the high performance per core required by today's software.
If two or three speed bins down allow for operation at half the power (the dual core numbers seem to bear this out), then going down by a speed of 1/2 on the same exact core will knock down the power requirements by a order of magnitude or so. Reworking the core to perform at no more than half the performance of the current top bin, will take much less than that. Going to a top performance of a quarter of the top performing bin, will knock the required TDP to under 0.5W. That still may be faster than the ARM core can go.
Lets see. The last x86 core that was under 1M transistors was the 486. That 486 reduced to 90nm would likely run at 1GHz at about 1W. Where that would compare to the top speed ARM core? I don't know. But Intel claims that their next Xscale (ARM) core will reach 1GHz, but that the performance will only be 25% faster than their current top bin 624MHz Xscale core. It looks like Intel will give up a lot of IPC for higher clock rates on their Xscale cores (<sarcasm on>So what else is new?<sarcasm off>).
Pete
Dear Mas:
I see you forgot to notice another compiler change, icc9 and ifc9. Comparing scores from two different compiler versions has caused differences to occur on the exact same hardware before with the later version scoring much higher. Comparisons between different hardware practically require using the same compiler. IIRC the 2GHz PM got only 1541 using icc8. Even using linear scaling, which no modern system can achieve, that translates to 1741 at 2.26GHz, a difference of at least 5.6%. I don't think it would score that high on icc8.0 and ifc8.0.
However these AMD CPUs beat that score anyway: A64 FX 55, 57, Opteron 152, 154, 252, 254, 852, 854 all using icc8 or older.
Using the lower 1741, these additional CPUs beat it: P4 3.8Ghz with 2MBL2 in a Dell Workstation 380 using icc8.1 and without HT, P4 3.73GHz with HT. P4 3.46GHz with HT, P4 3.6GHz 2MB L2 without HT.
All of the above using SPECint2000 peak per common usage.
Of course, all PMs score 0 using 64 bit versions.
Pete
Dear Ixse:
Since OPteron x75 already gets a SPECintRate of 67.2, AMD has already passed it with the equivalent of a A64 X2 4400+ (same core, same L2 and same clock, different socket 939 vs 940). Given that, you can backtrack to see what slower speeds can do and a A64 X2 running at 1.8GHz with a 512K L2 each, which is Manchester, would do about 57 on SPECintRate. Given that the slowest A64 X2 3800+ runs at 2GHz, that means that every A64 X2 currently sold will outrun the Conroe. If AMD would sell a 1.8GHz Manchester, likely be called a A64 X2 3400+, then and only then would the Conroe, that is supposed to be released 15 months from now, catch up to AMD's slowest A64 X2. The Opteron 265 HE (1.8GHz, 1MB L2, 55W TDP) would be slightly faster at about 60 and uses less power (55W vs 65W) than the slower top end Conroe.
15 months to almost catch up to AMD's slowest dual cores (Taylor may be out by that time and it could go slower, I guess, at much lower power), is a big statement on how far behind Intel has fallen.
Pete
Chipguy:
I notice you won't say how many sold. Dell stopped listing them. Ditto for HP. IBM doesn't. They all sell lots of x86 and AMD64 workstations. So do a lot of white box vendors. The ratio here is probably far worse than the overall one. IA64 enabled CPUs probably sell 3 (or more) orders of magnitude less than AMD64 enabled ones in the workstation market.
Pete
Wbmw:
I don't buy into those arguments. All K8s and AMD64 enabled Celerons, Pentiums and Xeons can be placed into servers. We are comparing CPUs, not what they are used for. We are comparing ISAs as well, if you would rather look at it that way. IA64 failed in workstations. Failed in high end PCs. Failed in entry level servers. Is failing in mid level servers. You want to eliminate all of those markets because IA64 couldn't compete. And there is no logical reason for doing so. There is a rose colored glasses reason though. It makes IA64 look somewhat better. But it is still losing. Getting eaten by AMD64 on the low side and Power on the high side.
Pete
Chipguy:
Do all ARMs have memory management? No! Every x86 CPU sold by AMD does. Unless Intel still sells some classic 8086s, all of Intel's do too. How much memory is attached to a typical ARM? 64MB is the max according to Intel (Xscale 297x). That is not a lot, but enough to run X windows on Linux. Yes they could be given more, but how many even have that much? But what is its performance level? How does it compare to the slowest Celeron being made today? The slowest Sempron? How about the slowest Geode NX (1GHz Athlon x86)? The fastest ARM (Xscale 297x) is 520MHz. Not a lot of processing power. But enough to be a file server. About 5 years behind in the GP market though.
I do not know how many applications are available for ARM. But, I speculate that they don't even come close to those for x86. But you are right, I stand corrected that at least some of the ARMs can be classified into the GP market. I think that at a minimum of 250MHz, an MMU and at least 32MB of memory (it doesn't need a FPU) (mid level of 8 years ago), they could be added to the GP market. Its amazing how grey those lines have gotten over the years. But those that meet those minimums likely sold far less than 1 billion last year. Do you even have a revenue figure for 2004?
It still doesn't invalidate the argument that AMD64 enabled CPUs out sold IA64 enabled ones by at least 2 orders of magnitude in 2004.
Pete
Dear Chipguy:
Prove it! Of course storage servers don't need much horsepower anyway. At WTMJ (TV CH 4) their tape libraries run on a Z80 which serves up programming going out on the air. Its not that big a task. A PC tells it what to do. The Z80 gets the appropriate tape, scans the bar code and inserts into the directed tape deck. It then tells the deck to play from x to y starting at time z. When done it puts it back. (sarcasm on)Hard Job!(sarcasm off)
Application serving however is likely beyond it. Performance is required.
Pete
Wbmw:
I understand the difference better than you evidently do. AMD64 grandfathered all of the applications. IA64 practically none. Yet AMD64 fixed many onerous things. Added things that Intel said was impossible. Its so obvious in retrospect that you just do not see the big deal. To paraphrase Yoda, "that is why you failed!"
Pete
Wbmw:
It is not succeeding in its original intended market. Intel and you keep shrinking its intended market as pieces keep getting knocked off as it loses out in them. Workstations was definitely an intended market. There are very little in there. That market slice have almost completely been taken by x86 and AMD64 enabled CPUs. It was to get down into the high end PC market. It never got there. And it keeps getting eaten alive in the low end server market slice. Now its losing out in the mid range server market. Soon you will only have it in the ultra high end market and boldly claim that that was its intended market all along. Sorry our memories are better than you desire.
As for Montecito. Intel has exagerated performance claims in the past and even they do not get this 2.5x multiplier for integer based loads. Besides Montecito has to compete against dual core Opterons which have been out for a while. And Opterons speed likely will be bumped up before Montecito comes out. Lets look at this when it really comes out.
Pete
Dear Combjelly:
ARM CPUs are not typically used for mainstream computing. They could be used there given enough memory. But so could many 32 bit DSPs and Microcontrollers. And many 32 bit DSPs do floating point just fine. Its just that their overall performance is quite a bit below even entry level x86s. Another big problem is the small size of their RAM memories preclude it. Also the typical lack of memory management makes it fairly impossible.
Pete
Wbmw:
Your reading comprehension is way down. You should have it checked. I said go ahead and include them. And we could go all the way and include every other Embedded CPU, DSP and Microcontroller which together sold more than even ARM. Still in that market, AMD64 enabled CPUs out sold IA64 enabled ones by 100:1. And x86 ones sold more than 1000:1 over IA64 enabled ones in 2004 by your own figures! ARMs didn't sell over 100:1 over AMD64 enabled ones or even 10:1 over x86 CPUs in 2004. And ARM CPUs certainly didn't get more than AMD64 enabled CPUs in revenue. You can easily see that AMD64 enabled CPUs had far more revenue than IA64 enabled ones.
Pete
Wbmw:
You still haven't named one that can do it without a recompile. That was your original objection. Your logic is so faulty that you can't stand it. You try change the topic to try to wriggle out of it. Tough!
Pete
Wbmw:
And we have corrected you that AMD64 CPUs can all be used in servers. There isn't much of a difference anymore. Thus when we compare AMD64 CPUs against IA64 enabled CPUs, the whole market of GP CPUs is the market. Not just in those areas where IA64 CPUs may have some small success. Just because they lost in all those other areas doesn't mean we shouldn't include those areas.
Even if you want to include all ARM CPUs in that market, fine! Then we could add the many MIPS, DSPs and Microcontrollers of various sizes which are far larger in volume than even ARMs. Still IA64 enabled CPUs sell less than 1% of AMD64 enabled ones. By your own figures, IA64 enabled CPUs sold less than 0.1% of x86 CPUs in 2004.
You just don't want to admit that IA64 enabled CPUs are marginalized. AMD64 enabled CPUs are shredding IA64 enabled CPUs all up.
Pete
Chipguy:
ARMs are not GP CPUs. Both IPF and AMD64 enabled CPUs are. Every AMD64 CPU can be used in a server. Thus you should include all of them. In fact all x86 CPUs can and are being used in servers, even some embedded ones.
Pete
Dear Mmoy:
Hear hear!
I have done this since the early 80's. Mazda RX-4 that got an average of 17mpg using leaded regular. Then I went down to a F-100 Pickup with a 302 V8 that got 13mpg. Next, a F150 4x4 with the 300 I6 that got 17mpg. After that to a Aerostar Minivan that averaged 20mpg. I now drive a Ford Focus that gets 27mpg in the city and 38mpg on the highway. I average over 30mpg so far. I'm glad I got it.
If everyone else would do the same, the speculators would be pushing down on the price instead of up.
Pete
Chipguy:
We are comparing AMD64 enabled CPUs vs IA64 enabled CPUs.
AMD64 enabled CPUs sold over 20 million in the last 12 months. That includes Xeons, P4s, Opterons, Turions, A64s, X2s and now even Semprons and Celerons. Heck, AMD is likely to sell over 30 million AMD64 CPUs alone in the next 12 months. I'm sure Intel will at least match that.
Opterons are grew 89% from Q1 to Q2 alone. They are up over 300% over 2004.
Itanium boosters always try to limit the scope as it is far better than being blown away in the big picture.
Pete
Wbmw:
So you can't name a single one. Your objections are totally without merit. AMD64 is an innovation. Live with it.
Pete
Chipguy:
It is being outsold 100:1 to 1000:1 by units (CPUs). And 10:1 to 100:1 by CPU revenue. Servers have a lot of other things in them like memory, HDs, controllers, cabinets, software and service contracts.
Pete