Ephraim Schwartz - are you Jonathan's bro?

I'm always curious about these assumptions by the press that SPARC costs more and delivers less than cheaper x86 counterparts. It's like the old school folks saying "Java runs slow." (The older JRE versions did. They're definitely fast now! - and about to get faster!) In my experience, SPARC does run on par with x86 at times; other times, SPARC is the faster CPU. How about citing some benchmarks when making the comparison? I still think SPARC has long legs.

Conversely, I'm rather curious what makes people think that SPARC runs so much faster than x86 that they can justify the large cost difference.

First, let's make no mistake: SPARC hardware costs more money. You can weasel your way around the statement if you would like, but if I have $500 and only $500 to spend per machine, it isn't going to happen. SPARC never came anywhere close to that.

Second, let's make no mistake: (recent) x86 is not designed to scale past > 8 processors. Currently, Intel stops at 4 dual core chips and AMD stops at 8 dual core chips. SPARC has been known to scale to 72 chips and more.

For raw number crunching, I just don't see SPARC as the way to go. It can hold its own, but without a floating point SIMD instruction set such as SSE, it will fall behind pretty quickly. Luckily, most uses of SPARC are not HPC.

Now, granted, a 450MHz UltraSPARC II, cycle for cycle is far faster than a 450MHz Pentium II, but with Intel's new Core2 architecture, the gap on a cycle by cycle basis is closing, if it has not already converged.

With the UltraSPARC T1 however, SPARC does have a performance lead over x86, but not in the traditional "My ALU is faster than yours" sort of way. By having 32 threads and an OS to match, highly threaded applications can have an overall level of performance that is higher since more will be done at once, but each individual task takes slightly more time due to the rather low clock speed (1.0GHz or slightly higher). For highly parallelized tasks with basically no FPU code, this works well and the price/performance is amazing. For single threaded tasks, well I have heard horror stories about tapping fingers, waiting for the job to complete.

Now, the SPARC workstations still use the UltraSPARC III 1.6GHz. Those are pretty fast, yes, but I doubt that they would really compete with a similarly price x86 system. Moreover, I don't think Sun has updated their UltraSPARC line since IV+, which was a number of years ago. I think they are throwing in the towel for everything but insane throughput computing (i.e. UltraSPARC T1). Fijitsu pretty much showed that its SPARC64 VI line beats UltraSPARC IV on single threaded performance. It is no wonder the new enterprise servers from Sun are using Fujitsu's processors. Just checking out the cheapest config. of a high-performance UltraSPARC workstation Sun Ultra 45 for $3,695, I get 1GB RAM (333Mhz), a 1.6GHz UltraSPARC III, a modest graphics accelerator, and 250GB of disk space.

$3,269 at Dell gets 2.66GHz Quad core processor, 1GB of RAM (667Mhz), 2x146GB SAS disks at 15K RPM, a more than decent graphics accelerator (supported by Solaris/NVIDIA no less). And for you people who break your computers, 3 years of "Gold" level support is included in that price. Save about $430 too.

Let's be honest now, do you really think that a single 1.6GHz US-3 will even come close to matching that? Yeah, I highly doubt it. =\
If you said "Yes", you might just be ignoring the facts. Basically everything is faster in that system. Faster processor frequency (using an extremely [as far as x86 goes] efficient processor), 4x the cores, twice the RAM speed, 15K RPM disks with higher throughput, PCIe x16 NVIDIA graphics card. It just hurts how much faster that is.

Now, if I wanted to match the performance of the Ultra 45, I'm sure I could dumb it down, but then the price would be glaringly lower and make the Ultra 45 look like a worse and worse deal.

-> If you do enough R&D into a borked processor architecture such as the x86 and never update a superior one, yes, the stupid one will pull through by the sheer amount of mind power and dollar power thrown at the problem.

I love SPARC hardware as personal user and a novelty, but when the day is done, I would not be replacing my HPC and workstations with it -- it is just too expensive. If it costs $3,695 to get a workstation that is essentially shadowed by its retarded x86 cousin on crack, then there just isn't a point in getting one. Even if you absolutely must get a Sun system (the name usually means high quality and great support), Sun's Ultra 40, based upon the dual-core Opteron in its weakest configuration could blow that system out of the water for about $1,600 less.

If your goal is to scale to large numbers of processors in a single system that just cannot be done with a cluster of smaller, cheaper x86 boxes, then by all means, accept higher costs, get a super SPARC server -- x86 just won't take you that high. But this "low-end" SPARC market is just stupid. If the Ultra 45 was < $2000, then I might see it being justified (even though a sub $2000 x86 could *still* overpower it). RISC/UNIX vendors just don't care to cater to cheaper prices since they sell fewer machines. And with higher prices, less will want them. It is a rather deadly spiral, isn't it?
I really want a Sun Ultra 45 despite having said all of this, but being the decision maker for my company, I have to put aside the childish nature of want and make choices that benefit my company.

Java? What the heck?
Java does run more slowly in comparison to native C code. Have you done enough programming in both languages yet or did you just repeat Sun's "Java makes things fast!" speech too? It isn't that black and white either. It is fast enough for many uses (there has been some performance gap bridged, yes), but it still uses more memory than native C code (and pretty much always will considering the JVM's job), but also has all of the advantages of Java's language and allows people to develop applications faster using the rich class library.
If you have ever whiled away a week porting, writing and testing makefiles, or working out kinks in a program to make it crossplatform, then you can appreciate the portable by design. Java usually means lower development cost compared to C, but slight hit to performance -- usually unnoticable with our fast processors and the advanced JIT of the JVM. It has its uses and the number of uses are increasing, but there is a reason that 3D games and HPC software are not written in java. Let's not get religious about it.

@ Ephraim,

You may change your mind on Darkstar when you see the announcment coming out within the next 10 days...

@ Patrick

"but there is a reason that 3D games and HPC software are not written in java."
This is starting to change. 20,000 downloads of the Darkstar SDK within 3 months is a strong indication of a savvier game developer audience that understands how the latest speed benchmarks show that Java can now go head to head with C.