Re: IA64

["Thomas W." <kirk@colinet.de>]

Hi,

> As a non-Alpha user (but an aspirant one), I'm curious to know how the
> 'community' sees IA64? Is it perceived as a threat? Would alpha users
> be persuaded to switch over easily?

The IA64 is very agreesive and new design wrt to microarchitecture. 
As ooposed to that, Alpha (21264) is a state-of-art "traditional"
architecture, which is already 64bit and performs quite well.
Based on my experience, Alpha delivers, what you pay for: superior
performance at a superior price-tag ;-)

The linux port to Alpha ( which is, what I use all of the time ) if
proven, reliable and stable. Although GCC on Alpha is crap, the
freely available Compaq-Compilers for Linux/Alpha do *very* well.

Important to me is workstation-class firmware: seriel consoles, netboot,
decent development support and console-callbacks, a fair set of diagnostics.
However, this counts for me, YMMV.

Based on my experience of Linux/IA64 on a dual IA64 prototype, performance
is - at the moment - fairly poor, compared to a dual 21264. Of course, this
is due to a lousy and unoptimized pre-production run of the sillicon.
At the moment, no one can *really* tell how fast an IA64 will be. Too much 
of the performance will also come from the surrounding chipset, like memory
and I/O buses.

GCC on IA64 sucks, period. GCC has no means of optimizing towards a performant
sequence of instruction-packets to be fed into the IA64 pipeline. The SGI
compilers, which are free as well, do a much better job. I don't see
any advantage on either side of the fence here - although the Compaq
compilers appear very mature, from what I have seen so far. I can tell
far less on behalf the SGI Compilers.

As for the Firmware stuff, I am *very* pleased how Compaq/Alpha handle
things. They emulate x86-Bioses of standart I/O cards upon POST, which
basically allows users to stuff an Alpha with a decent set of PCI Boards
cheaply available on the market. As for IA64, I cannot tell how Intel
will handle this - it might well be crucial for IA64's success.
The same goes for the Firmware functionality in general. Intel has the
unique opportunity of getting rid of the PC-crap, which is nothing more
than a glorified DOS-loader, even by today. I assume, IA64 will not start
in some real-mode and will not support Gate-A20.

As for the future, I doubt the extensions like ISSE, 3DNOW or even MVI will
persist. Instead, I could imagine, that having fast network-links on
chip for large-scale parallel machines will be a key point for any 
architectures' success. This is planned on future Alphas, and I hope
the Alpha team does this well. Putting a boat-load of ALU's in a VLIW
architecture is also an interesting approach, but the validity of this
concept will be proven in future IA64 versions, definitely not in the
first run.

The biggest threat for Alpha is - IMHO - Intels manufacturing capabilities.
If they can run a 256Kbyte on-chip L2-cache, accessed 256bit wide, at 1 GHz,
they sure have some superior fabs at hand. If they can turn that advantage
over to IA64, Alpha is in *serious* trouble.
Given Intel's ability roll-out huge amounts on chips, along with their
financial backing, I am certain they will try to buy-in the server market
on price, if the acceptance of the IA64 does not develop fast enough in
their opinion.

Whatever comes in the near or far future, Alpha is here already. It's proven,
it's fast and I can actually *buy it*. I doubt, Intel will make it in
the first attempt, but I assume future generations of the IA64 will be
*very* interesting - but yet, this remains to be proven.

Since I have no religious or whaterver bias towards Intel, I don't really
care. I would like to see Alpha succeed, since I think it's a Good Thing(TM).

Just my $.02

Sincerely,
Thomas Weyergraf

-- 
Thomas Weyergraf                                                kirk@colinet.de
My Favorite IA64 Opcode-guess ( see arch/ia64/lib/memset.S )
"br.ret.spnt.few" - got back from getting beer, did not spend a lot.