Slashdot Mirror
Alpha's Going Going Gone
WildCode writes "Get your Alphas now cause HP is releasing the last of the Alphas (the final one expected to be released in 2004), and there will be no more." I was already under the mistaken impression that Alpha was dead, so this story is rather bittersweet for me. Still, as far as architectures go, Alpha will probably be among my favorites. It was once vastly ahead of its time, if not severely cost-prohibitive.
I'm sorry, I just can't put my support behind a server that doesn't have stupid rhyming commericials for it.
Are Alphas "utterly buyable, give 'em a tryable"?
I think not.
No, it wasn't a mistaken impression at all... Alpha died about 1997, when Compaq bought DEC, and squandered the assets of a great company. Sure, they were still turning out machines, but the Alpha was as good as dead from that point on. I figured maybe HP would know what Compaq didn't, and resurect the Alpha, but they are beholden to Intel, so that didn't happen.
Believe it or not, even though it's been dead for the past 6 years, it could still be resurected...
For one, Intel has bought the rights the the Alpha, so they could use some of the same ideas in their Itanium and Pentium chips. The miserable failure of the Itanium is quite encouraging, because that could mean the only way they can get a leg-up on AMD64, would be to start making Alphas... It's wishful thinking on my part, but Intel would have much to gain.
Long-live the Alpha.
Slashdot gets worse every day... Pipedot: News for nerds, without the corporate slant
I have liked alphas since I first laid hands on one back in 1993 when it was running OSF.
I have a few of them kicking around my house, a couple of EV6s and an EV4 that still does sterling service.
Alphas were always fast, both integer and floating point, however it was the FP performance that really made them popular for scientific applications. Once the EV6s arrived the integer performance increased considerably over the previous EV5s with the addition of four integer pipelines (EV5s had two each for integer and FP) and also register renaming and instruction reordering. Alphas have 32 64 bit integer registers and 32 floating point registers, but an additional 32 showed up in the EV6 which allows the processor to do a lot more work per clock cycle along with the extra pipelines and out of order execution. The other great thing that EV6 introduced (well, actually the 21164PC chip introduced it but it wasn't as useful as the 264 EV6) was MVI. This is the equivalent of the Intel SSE/MMX instructions but the MVI instructions had direct access to memory as they were 64 bits wide just like any other instruction on the chip. This meant it was trivial to load up a 64 bit word and then do parallel work on the data such as sum or max.
Instructions also executed very quickly, typical of a RISC chip, and the processors had very large caches for their day (2MB being typical but server chips had much more than that). Even access to main memory was very quick, the EV6 bus was also used on the Athlon for this reason.
So, you essentially had in Alpha a processor that has able to crunch integer and FP data very quickly, had a fat bus to memory with a cache large enough to be useful and lots of general purpose registers, extended parallel instructions that worked easily with the existing instruction set, one which looks like an high level language it is that easy to use. They also had very high clock speeds for their day and used those cycles very efficiently, Alphas were running over 500Mhz when Intel was putting out sub 200Mhz 32 bit chips that struggled to do more than a no-op in the time that an Alpha could do a fourier transform!
Unlike Itanic, the Alpha was designed to make compilers easy to write, Compaq released the DEC compiler to work under Linux and it was amazing to see the boost in speed that came about when that was used. The fact that the EV6 was so smart meant that the compiler didn't need to be all that clever to make code that Alpha could run very quickly. It was pretty simple to avoid cache misses and other performance sapping problems.
Compared with other processors of their day such as SPARC the Alpha was at least twice as quick if not more. It is only as Compaq took over and took their foot of the pedal and speed ramps dried up that other architectures caught up but it took some time. If they had continued to keep pace with die shrinks and clock speed increases Alpha would have been embarrassing its competitors even today, in fact it still is if you witness the fact that HP won't release benchmarks for EV7 until Itanium can beat it.
So, yes, Alpha was great and I haven't even touched upon EV7 as I never got my hands on one and I'm not likely to now. Damn HP. Damn them to hell!
From what I have heard it is quite likely that Alpha EV8 technology will live on as the next gen Itanium, effectively something like the current Pentium where the Itanium instruction set will be wrapped around and Alpha style core with translation to make it seem like an Itanium. Yuk.
As I said, Damn HP.
"I have the attention span of a strobe lit goldfish, please get to the point quickly!"
The Alpha was formerly Compaq and even before that a Digital invention. HP has their PA-RISC architecture, of which Itanium was planned from the start to replace (one of the design requirements for Itanium was that a software translator for binary PA-RISC code was made possible).
Furthermore, as far as I know, the Alpha is still produced by Intel, not by HP/Compaq/Digital, as Compaq sold their alpha plants, personell and all associated IP to Intel (and thus avoided a lawsuit, as well as ensuring the Alphas future for a few years). There were also plans for a Itanium version of Tru64 (formerly Digital Unix), but I am unsure as to whether it was ever commercialized.
All in all. It seems like a pretty sound business decision to me. This is what "they" have planned all along, for many years, whether "they" are Digital, Compaq, HP, or Intel.
Having worked on Alphas (and VAXs for that matter) in my previous job for six years, I have to say good-bye to an old friend. It was, for me, an incredibly powerful platform that did so many data oriented tasks so easily. The multitasking performance was amazing. We, for the longest time, ran on a VAX that was the equivalent to about a 486-120 MHz machine that could handle thirty developers. At the same time, it could handle thirty clients running reports. In general, we didn't notice each other. The Alphas put that system to shame. I often had to remember that I was working with multi-gigbyte files and processing them in seconds, not hours like on a PC. But, I suspose we'll have to use the "future" of PC hardware until it eventually catches up to the past.. =)
Bel, the mostly sane.. "Of course I can't see anything! I'm standing on the shoulders of idiots." -- Me
That's what we called them when I worked at DIGITAL (up until Compaq canned a lot of people right after the purchase and almost killed the Alpha then).
They are still ahead of their time. The fastest Alpha's (EV8, over 1GHz) were still far faster than an equivalent speed x86 processor.
I've heard repeatedly that Samsung will still be producing the processors. I have not looked into this recently though.
It's a shame to lose such a great architecture. Yet another example of the best ideas not always being the most popular or surviving. At least part of the architecture will live on in AMD chips (for now at least).
PGA
The fact that the EV6 was so smart meant that the compiler didn't need to be all that clever to make code that Alpha could run very quickly.
I guess "smart" is a relative term I guess how you apply it. If you say the designers were smart in making the chip very simple (dumb) which made compilers easy to write because there weren't that many ways to do a particular thing well, then I'd agree with this statement.
The thing about the Alphas were they really were RISC. The 21064 has *horrible* performance with strings and byte oriented operations because byte level addressing was not present in the instruction set. String manipulation and the like were synthesized in the libraries by using shifts, masks, and such. Not a very *smart* CPU... quite dumb actually (not that this is a bad thing... it made it simple). The story was that the designers were a little too purist with the first Alpha and "underestimated" the amount of byte level operation/addressing that was used or something or just wanted to really make the Alpha a number cruncher. At this level of the architecture, there was no out of order execution or the like. A pipeline stall caused *everything* to stop until the memory request was handled.
Very shortly after the 21064, byte level addressing and manipulation was put into the Alpha line. Of course, this was really nice and improved a lot of those type operations. Also, out of order execution was put in at a later point. That was the really nice Alpha.
As far as living on... there are many CPUs today that have at least a part of their lineage with the Alpha. Either their FSB (Athlon), their philosphy (high clock speeds), or direct decendents (StrongARM).
Another bit of history: at the time the Alpha was released, HP had another CPU that they liked a lot. It was the PA8000 family. The PA followed the "wide" philosophy of processor speed - it had lots of execution units, out of order execution, and all the other stuff that we see a lot of today. The problem was that all of that extra circuitry made it *extremely* difficult to ramp up the clock speed. People referred to this camp of design as the "Brainiacs" because they did a lot per clock cycle and the clock speed was fairly low because of the complexity of the chip. The Alpha camp was called the "Speed freaks" in that they believed high clock speed was first, then later design in the complexity. Anyway, the PA8000 and the Alpha started out about the same speed but HP just couldn't keep up with the clock speed ramp up of the Alpha.
So... I guess back to the original topic... the Alpha was easy to write compilers for early on because of the simplicity of the instruction set (the chip is "dumb"). Later, when the Alpha crew added the byte operations/addressing it simplified some library writing. Later, out of order execution and such were added which didn't really impact the complexity of the instruction set as much as just made the compiler better. The OOE is where the "smarts" came in.