Slashdot Mirror
Alpha 21264 And Athlon 850 Review
Arg! writes: "Hi. There's a comparison review of an Alpha DS10 with a 466 MHz chip running AlphaLinux up at this link on Ace's Hardware. The system is compared against an 850 MHz Athlon box (also running Linux) and there are a lot of different benchmarks also showing some memory comparisons with some other PC chips, like the Celerons, as well as some Sun Ultra benchmarks. The config tested was $5500, but a base DS10 is supposed to be around $3500 ... maybe not too bad for a nice 64-bit box. ;)"
An excellent point. Sort of like 286/386 days with Harris & AMD outdoing Intel on their own masks.
The 21264 won't ever see 1.6GHz, or even 1GHz. The now shipping 21264A (EV6.7) and upcoming 21264B (EV6.8) will see those numbers, though. The EV6.8 will be 0.18u with copper interconnects, and should arrive before the end of the year. Of couse, the EV6 was supposed to have scaled to 1GHz by last year, and we're just now at 733MHz. Now that Compaq has finally settled most of the confusion caused by the DEC buyout, things should get back on track very quickly.
And to those who think the Alpha is a dead architecture, you've got another thing coming. Sure, its more expensive than an Athlon system, but it's not meant to compete in that market. Plus, no PC can offer the sheer memory throughput or clustering capabilities that are offered by VMS and Tru64 Unix on Alpha (and no, beowulf doesn't even come close).
--
http://gammatron.weblogger.com
I recall from many years ago that the byte benchmark sucked when you failed to compile the benchmark using proper byte alignment. This was exposed when gcc suffered compared to ms compiler on windows I believe. The deal is, the cache works twice as hard to feed the cpu if words are not aligned on cache line boundaries, wich might be 64 or 128 bytes or something. The gcc compiler did much better when alignment was fixed. This may be more of a problem since the floating point numbers may actually be double precision by default (?) I wonder if this benchmark aligned properly for gcc (or for the compaq compiler for that matter).
The problem is that artifical benchmarks are dumb. The only thing that really matters is application performance. If the applications you use run faster on one CPU than another, voila, there you have your benchmark. Of course, good artifical benchmarks are possible if they emulate application code (like the 3D 2000 max or whatever benchmark that uses actual game engine pieces to test performance.) Cross platform testing is a bit hard. You'd have to test app that were equally optimized for both platforms. It shouldn't be too hard, however, with good optimizing compiles. Take GIMP and compile it using the Intel x86 compiler (even more optimizing than the Microsoft compiler) and test GIMP compiled with the Compaq Alpha compiler.
A deep unwavering belief is a sure sign you're missing something...
One thing to remember is that no matter how good GCC is, Compaq's compiler is *amazingly* good.
--GnrcMan--
Trying to categorize RNA or DNA or protein sequences using suffix arrays with dynamic programming, including sometimes large datasets.
A substantial portion of the gene sequencing being done now is on the Alpha. This is the sort of thing that the Alpha really excels at.
--GnrcMan--
I disagree with some people when they say that Alpha is a dying platform. In reality, it is far from dead. Aside from the hideously expensive POWER3 CPU, the Alpha is the fastest CPU out there. With the release of the EV6.7 and EV6.8, they may again capture the fastest proc known to man crown again. A 1 GHz Alpha is not that far off, Compaq demoed air cooled 1GHz Alphas a few months ago. In fact, Alpha may be the only competition left for Merced (I refuse to use that Power Ranger-esque name Itanium!) in the mid to upper range market. There was an article a while ago (on Tom's Hardware I think) cogigating on whether or not Alpha would still be faster than Merced. They cited various points about how plain RISC may beat out VLIW if implemented properly (like in the Alpha) and how the DEC guys have quite a leg up in clock speed (at least at that time.) Judging by when Merced will be out the 1GHz + Alphas should be out, and should provide quite a strong competition since
A)Alphas could just be faster, and
B) Alphas have much more market penetration and a larger support network built up. With Intel's power, that will change, but Alpha will have quite a head start on Intel.
Then there is the fact that Alpha is aimed at a much broader spectrum, and Intel might just not be able to take the market with Itanium.
A deep unwavering belief is a sure sign you're missing something...
The compiler situation presented by the article hilights a key advantage that Compaq has over Itanium. The compiler needed to take advantage of Alpha is already freely available, while (if intel keeps with it's current practice of charging for it's compiler) one would have to pony up quite a bit for Intel's compiler. The compiler is critical because Itanium (based on VLIW, which places all code ordering work on the compiler) is heavily dependant on a good compiler for any semblance of performance. Sure GCC compiles Itanium code, but if it can barely optimize x86, do you really think it has any hope of producing good Itanium code? This might not be important on Windows, but as OSs like Linux and BeOS, which depend on the freely available gcc to compile many of the applications, the availibilty of a good free compiler is very important.
A deep unwavering belief is a sure sign you're missing something...
Many cheap Multias come with a soldered 166 MHz CPU, no floppy drive, and no hard drive. The more expensive ones (~$25 extra) have a socketed 166 or 233 MHz CPU, a floppy drive, and sometimes even a small hard drive (400 or 500MB).
These Multias run Alpha Linux just fine. They're around the speed of a Pentium 100 at integer computations, and a Pentium 200 at floating point.
You also might want to look at the 21164A, which is quite cheap these days. You can get a 533 MHz CPU, motherboard, and perhaps even a DIMM for around $500. This will be about the same speed (or a little faster) than a 450 MHz Pentium III at integer ops.
The nice thing about the mismatch is it really highlights the power of the alpha when fed properly optimised executables.
As the article shows, using the C|O|M|P|A|Q compiler, the Alpha still defeats the Athlon at floating point work. Pretty amazing stuff.
As a long term fan of the alpha, I am so pleased to see the architecture entering the consciousness of the enthusiasts. Perhaps someone can lend these guys an XP1000 (XP stands for "Extreme Performance") or even an ES40, just for fun.
Course it would be more fun if they had started pumping out 0.18 micron Alphas so they could return to the days of eye-popping clock speeds - rumour has it the 21264 could hit 1.6GHz!!!
Just don't mention the cost - always the weak spot of the Alpha.
I have a 21164 500Mhz box with 256MB of memory, total cost was around $1k. This was when memory was expensive, too...
Not a bad system at all. Nowhere near as fast as the 21264's, but much, much, cheaper. It would have been pretty usable with less memory, too...
My blog: http://www.seebs.net/log/ --- My iPhone/iPad app: http://www.seebs.net/seebsfrac/
The MHz's must be smaller on an Athlon than on an Alpha. This is similar to the US gallon vs the UK gallon problem. It's time we switch to metric units when reporting computer speeds.
That the DEC compilers for AlphaLinux were that good. Wow. If only they would merge their optimizations into GCC... Incidentally, I just noticed that DEC has available a new version of em86 that should work on the 2.2 kernels, and that they also have been kind enough to make a nice RPM of netscape for DUNIX with the needed DUNIX libraries that Linux needs to run it.
For every problem, there is at least one solution that is simple, neat, and wrong.
The article said that you need a special graphics card for an alpha. Can you use all your other pci cards on an alpha?
Yup... I'm using an Alpha right now with a Matrox Millenium II PCI.
--
Every AMD 750 and Via KX133 motherboard supports the Athlon's 100 MHz DDR front side bus. DDR stands for "double data rate", which is chip-speak for "it transfers data effectively twice as fast". If you take an electrical engineering class (microprocessors, for example), you'll learn that data is transmitted on the bus only at certain times (defined by the clock).
For example, consider that the bus is on a 24 hour clock. The Pentium II with its 100 MHz bus transfers data at 12:00 noon. The Athlon, with it's 100 MHz DDR bus, transfers data at both 12:00 noon and 12:00 midnight.
I hope that explains it.
Don't get me wrong, I very much like Alpha's. I have three Multias, one each running Linux RH5.2, OpenBSD and Compaq Tru64 (hobbyist).
But take a look at the 21264 datasheet: 107W [sic] maximum power/heat at 633 MHz (2.35V & 46Amps). This is HOT, but it's also 0.35 um process. If they shrink it to the current 0.18 um, DEC could bring it down to ~20W @ 1.4V. Very competitive.
But will Compaq make this investment?
Cool! Where do I find the -$3500 deal? ;-)
;-)
If you go here and do a search in the bottom frame, for "alpha", you'll see some of the $3000 models listed as being available!
I still want to actually build my own Alpha... as the Penguin-Ferrari article points out, we need diversity, and anyway I like messing with different stuff. Anybody know of more recent articles, or have personal experiences to offer?
I bought a refurbished Alpha Digital Alpha Workstation 433au (21164 version, either EV5 or EV56) from egghead.com for US$1,199 a couple of weeks ago.   It just came in last week and it's up and running Red Hat 6.2 for alpha, kernel 2.2.14-6.0.   The box (without monitor and keyboard) came with a 433Mhz alpha CPU, 128MB RAM, 2MB cache, an Adaptec 2940 SCSI controller, 2.1GB Seagate SCSI drive, 12X CDROM, Intergraph video card (and something with Cirrus Logic video too), built-in EtherWorks ethernet, and built-in SB-compatible sound.   It also has 2 64-bit slots and 2, shared ISA/PCI slots.   It's one of the older boxes (circa '96-'97), but it's processing seti@home work units (non-optimized) at around the same speed as a PIII 450.   One of the main reasons why I bought it was because it is upgradeable to a 600Mhz 21164 processor (although I am finding that the CPU alone is exorbitant and I would do better buying a mobo w/CPU and memory!).
During my difficult search for any alpha info, including parts (I am realizing how little is written about alphas as compared to x86s), I stumbled upon this site that sells DEC Alpha hardware (and others like Sun), and they have a link here called "Build Your Own Alpha System", where they list the various parts you can purchase separately to put a system together.   There are a few other places out there too, but I don't have the links handy.
I would say go for it - I've seen prices for 500-600MHz alpha mobos, w/64MB RAM for like $1500.
All in all, this is my first non-intel (or non-intel compatible) machine and it's been a really decent learning experience dealing with the Penguin on the 'ole Ferrari!
-- Win2k: "It's not so much that it's only 65,000 bugs, it's just that they stopped at 65,535 to prevent an overflow."
Cool! Where do I find the -$3500 deal? ;-)
Let me point out an article I liked from Linux Gazette on how to build your own inexpensive Alpha. It's almost 2 years old, so some of the details aren't as relevant, but I think some of the stuff explaining the gotchas of the Alpha platform for someone used to x86 (like me) are still informative.
I still want to actually build my own Alpha... as the Penguin-Ferrari article points out, we need diversity, and anyway I like messing with different stuff. Anybody know of more recent articles, or have personal experiences to offer?
"The best we can hope for concerning the people at large is that they be properly armed." - Alexander Hamilton
They're also comparing the 21264, which they're saying is around $3500 (though their system had extra stuff and was more like $5500), to the Athelon system which they seem to be saying is around $2000.
You're right. It's fair... NOT! :-)
"The best we can hope for concerning the people at large is that they be properly armed." - Alexander Hamilton
... in the comment that HP is betting solely on EPIC and MIPS has stopped processor development altogether. Given the slippage in Intel's original roadmap for Merced, HP were smart to keep a hand in the PA-RISC and while MIPS may not be revving their processors like the others, they earn a very tidy income from licensing their IP and creating variants of their processors for the embedded market while continuing to source high-end designs from SGI (who still have a 5 year roadmap for their R14K, R16K).
... more like unpaid advertising sometimes. I suspect that past 1 GHz, the CPU performance is of only marginal benefit (outside niche areas) as the speed limit is the bandwidth limitations anyway.
From a personal perspective, it is rather disturbing from an architecture point of view that so much attention is focused on the branding and MHz rating. If we use the analogy of cars, the peak revolutions per minute has got nothing to do with the actual real-world engine performance. A lot of factors depend on the I/O subsystem (gas tank + injection system), cache design (suspension), and more recently stlying (bodywork). Just because it runs hot, doesn't mean that it runs well, in fact from a thermodynamic perspective excess heat is an indication of inefficiency. Just as in real-life, there is a distinction between buses (good ol' shared memory Suns), industrial trucks (IBM workhorses), SUVs (SGI drool-designs) or motorcycle packs (Beowulfs). Pretending a souped up scooter with over-granished rusty frame can do the job of everything is a serious indication of cluelessness or delusion.
Just as in real life, the limitation is the overall transport system (network) which is still an information back-lane despite the heavy hype. Sure a speed-demon Porsche (Alpha) can outrun anything in a speed race but most people settle for a Ford (Intel) or Chevvy (AMD) to commute to work. Some may prefer a flashier Saab (Apple) or stick with the boring but solid Volvo (IBM) or even go upmarket with a BMW (SGI) but they all serve a basic purpose (mobility) and dominate specific niches. You get paid for maintaining a professional non-bias and correctly matching your company's needs to the available choices.
The quasi-technical mainstream press really has to
get their act together if reading the IT section in any general newspaper is any indication
LL
IMHO the current drive towards having one power-hungry-but-fast CPU in most commercial systems is not the best approach for delivering high-power and low-cost systems. A large gain in computing power could be made by changing the machine design to one based around several low power CPUs. A CPU design such as the ARM architecture, if modified to support multi-processing (the StrongARM does not support it as-is, I do not know how the situation is with other ARM chips) would be good for this as ARM CPUs are sufficiently fast and are power-friendly. (see http://www.arm.com for more details).
It would then be possible, once the chipset and motherboard have been designed and manufactured (not a small task by any means) to have a system with a lot of processing power but not costing a great deal more than a typical high-end system today. Also the power requirements for such a system would be on a par with a current system as well, if not less (depending on the number of CPUs employed and the system with which it is being compared).
Got the motherboard (with CPU) from a guy on the NetBSD lists. It woulda been $100 cheaper, but word got out and people started tracking down his source and the source raised the price... Memory from the chip merchant. Case from Antec - they botched it pretty bad, actually, but I soldered the relevant contacts to a switch. ;-)
After the recent spam problems and the amazing amount of lying they've done, I'm not buying from eBay, ever, under any circumstances.
My blog: http://www.seebs.net/log/ --- My iPhone/iPad app: http://www.seebs.net/seebsfrac/
Call Compaq about some of the discounts they have available. Through a reseller and with software developer discounts you can get the price REAL close to $2500.