Hacker News Reader

Comment by titzer | Hacker News Reader

I worked for Sun. The company mindset never shifted from building big-iron megaservers. Which is a shame because the workstation market lulled for a bit, but is actually back with a vengeance nowadays with heavy-duty GPUs, beefy CPUs, etc, for media and video editing--think $20k workstations. Apple has a slice of that market and a few PC manufacturers that focus on that, and it's very profitable. Not million dollar server profitable, but profitable. Sun neglected its own very profitable workstation line in favor of the server market, which was completely rocked by the dot com bust.

Sun, with Java, which hides SPARC vs x86 could have made the entire cloud happen a decade earlier. But SPARC was a performance loser compared to Intel and Sun just stubbornly refused to see the horizontal scaling forest for the giant E10k trees.

View on HN

>Sun, with Java, which hides SPARC vs x86 could have made the entire cloud happen a decade earlier.

All they had to do was live their trademark tagline, "the network is the computer."

Which I learned today was abandoned by Sun or Oracle and was adopted and re-registered as a trademark by Cloudflare in 2019.

They did, from SUN RPC, they moved into Distributed Objects Everywhere with Objective-C/OpenSTEP, which then gave birth to RMI and J2EE.

Containers are basically Java application servers, rebooted as language agnostic OS services.

I don't know that SPARC was such a performance loser to x86 so much as Intel simply having a commanding lead in process technology at the time. Even Apple capitulated. That lead gave them a market domination that allowed them to reinvest (think Amazon) in their own fabs. x86 was a dog that Intel spent $1B/yr on improving so that your code ran fast. They took that strategy to the bank for as long as they could which was at least a decade longer than anyone thought they could.

As I think you're pointing out, Java wasn't a pivot. It was rather an excuse not to pivot. Sun really needed a Gates style pivot with Internet Explorer or a Bezos style pivot with his services manifesto. Instead you got a ponytail.

Also, before missing the cloud, Sun missed routers. Still, they were a pretty successful failure. We should all fail so well.

Oh, it was totally. Sun lost at least two generations of microarchitectures to mismanagement. They resisted doing out-of-order execution far too long and they just could never keep up with single-core performance. Even process technology would not have been able to rescue the comparatively sophomoric micro-architecture. Intel and AMD and IBM just totally outclassed them.

Sun didn't miss routers... The part of Stanford University Network that went to do routers used an other name: Cisco.

SPARC rapidly became a performance loser to x86. And, at the time, Intel had no process advantage. None at all.

Intel Pentium II was almost exactly 100% as fast as UltraSPARC IIi, on the same process (350 nm), at about 1/20th the cost. That was the last competitive Sun SPARC product, after that they were all brutally slow. The less said about US III and US IV, the better.

By the time that Sun put itself on the auction block, their internal CPUs weren't even good enough for legacy customers. They were obliged to start re-selling Fujitsu SPARC products just to keep SPARC loyalists inside the tent.

You're right. When I compare SPARC with x86, they seem to have the same process node at roughly the same time. I probably meant manufacturing rather than process. Intel at the time was colossal. It was possible for an innovator like Transmeta to come along and develop low power. But then Intel could motor past them on a marketing promise, engineering and finally out manufacturing them. This manufacturing prowess was a lot of the reason for Apple's PowerPC to x86 switch in 2005. Moreover, advantage begat advantage until x86 ran into various walls.

Things have changed since and now TSMC, Transmeta's fab then, is the new colossus. And Apple has even dumped x86 for their own ArmV8 designs (manufactured by TSMC).

Didn't the UltraSPARC lines have greater throughput as compared to x86 offerings, even when they couldn't keep up on single threaded performance?

Sun experimented with an UltraSPARC-branded ultra-wide CMT design in their last days -- Niagara. It was very unsuccessful, because, realistically, what can you do with 32 low-performance cores that you couldn't do with 8 high-performance cores?

The answer turned out to be "nothing"

------------

The "conventional" UltraSPARC designs -- US IIi, US IIIi, US IV, and the canceled "Rock" project -- were just really very mediocre chips sold at tens of times the cost of a similar x86.

It really is amazing that Sun and HP had the hubris to design their own RISC processors in SPARC and PA-RISC; everything that was wrong with the Unix workstation market in the 90s can be summed up there, in a way. Proprietary CPUs and custom repackagings of Unix, all slightly different, but in reality they weren't different enough to justify. And this just left room for Microsoft and Linux and x86 to just come up the middle and make them all obselete.

Lots of vendors throwing around "open" this and "open" that but none of them meant it. Just try writing a GUI app for a Unix machine in the mid 90s. Motif was closed source, with licensing fees. Things like GTK and QT were still primitive. Xt was garbage. It was just a hodge podge of things with the assumption that anybody entering the market had hugely deep pockets.

If they'd all just united around POWER architecture at least maybe they'd have had a chance?

In the 1990s it actually made some sense. If you go back to, say, 1993, every RISC vendor was selling chips that were much, much faster than x86. It was working like gangbusters.

The problem is that development costs rose and rose and rose without bound. Even if you could outsource fabrication, just designing your chip would bankrupt you.

HP, to their credit, saw the writing on the wall. The Itanium started as a joint Intel/HP project. It was HP's attempt to get Intel to pay for their new generation of chips, and sell it to the broader industry, in the hope they could have a new generation that would be (somewhat) backwards compatible with PA-RISC, but get off the treadmill of escalating development costs.

It was a clever plan, but as we know now, it didn't work worth a damn. Practically nobody but HP actually bought the IA64 chips. Intel quickly realized it wasn't worth any further investment and let HP's badly-needed chip series wither on the vine.

Sun didn't even have that much of a plan. They kept making chips they could not afford, and when "Rock" threatened to bankrupt the company... they just gave up. Even though giving up also threatened to bankrupt the company.

All of this is true except that there was no need for every vendor to have their own RISC ISA.

Which is why I'm saying with 20-20 hindsight it feels like they should have all just piled in on POWER. Of all of them born in that era, it's the only one still around. (Apart from ARM of course whose target was something else. MIPS is kinda just on life support.)

I don't remember the politics around the consortium there. I guess it could be that Sun and HP and the others just saw IBM as too much of a competition to want to be in bed with them. But given most of these vendors had shipped 68k arch workstations previously, I would have thought they'd be down with Motorola's offerings?

The problem in the end is these people all saw the other workstation vendors as competition rather than the actual competition, which was Windows/x86 (and later, Linux).

POWER was not originally open for use by others, under any license terms. It also was on the trailing edge of performance for many years. The original POWER 1 and POWER 2 boardsets were enormous power-sucking bastards with price tags to match. (Yes, boardsets. They were not even microprocessors.)

SPARC and MIPS were relatively open from day one but there were fewer takers than you would expect

SPARC had a number of licensees, but most of them chose to either re-use Sun software (Fujitsu, Solborne) or sell their chips directly through Sun (Ross, Weitek)

MIPS had many licensees but it ended up being bought by one of its most successful customers, SGI. Turns out, controlling the product direction is pretty important!

The point is, there were a lot of market forces piling up to drive people to play their own game, even it looks very silly 30 years on.

Ironically Solaris with SPARC ADI is one of the few UNIXes where C is properly tamed with hardware memory tagging, while Intel borked yet again their attempt to provide such capability on their CPUs (MPX).

I have been using SPARC for a decade, in 1998-2007.

In all this time all the SPARC servers and workstations that we had were ridiculously slow in comparison with Intel/AMD CPUs.

Nevertheless they were used because there were a lot of EDA/CAD tools that were available only for Solaris.

As soon as Cadence, Mentor etc. have ported their programs to Linux, SPARC was dead.

Towards the end of that decade, the difference became extreme, my laptop with a 64-bit AMD Turion CPU could run a SPICE simulation much faster than a very large and very expensive multiprocessor SPARC server.

Despite their slowness, the Sun or Fujitsu computers had many nice features before they became also available on Intel/AMD computers.

For example, in 1998 I was impressed that the Sun workstation we had could be powered on and off from its keyboard, because this happened before this became normal also for Intel/AMD computers. At that time we still had PCs with PC/AT power supplies, not with ATX power supplies, so power on and off had to be done from traditional switches.

I remember impressing my boss in about '97 - I got the first Mac in the company as my work computer, and I could switch it on (and off again) with a dedicated key on the keyboard. That was unheard of in his PC-centric world. ATX power supplies did become popular soon after that, though.

Yep, complete waste of talents and specialties -- for a company which had its genesis in amazing workstations to have lost its edge there is sad.

Also, who knows how it would have turned out if they had actually doubled down on OpenSTEP and DOE, instead of ramping them down in Java's favour.

Or if they had invested into AOT and JIT caching options for Java instead of leaving them to third parties.

Both options could have made Solaris more interesting, maybe.

It was really hard to resist the tidal wave of commodity x86 hardware at that point in time. Nobody did, not even Apple. And the server market got eaten by Linux, Solaris had no serious advantage anymore despite being a quality product.

It really is a different story now, though. There's room for more diversity, though they'd probably have to have made the switch to Linux.

Open sourcing Solaris earlier on before Linux ate the whole market -- or embracing Linux, or taking some hybrid approach, could maybe have kept them relevant.

Many would state that even today there are quite a few features that Solaris does much better than Linux, but so is "progress".

Reply on news.ycombinator.com