Is the Sparc T4 Too Little Too Late? 128
packetrat writes "Ars Technica reports on Monday's launch of the Sparc T4, and how it finally (nearly 20 years after everyone else) brings out-of-order execution to Sun Sparc ... er, Oracle Sparc. But the benchmarks that Oracle has thrown up (surprise) are a smokescreen for the fact that the processor is still woefully behind state of the art, and it serves mostly as a placeholder to keep the remaining Sparc user base from defecting to Intel — even as Oracle is selling systems based on Intel and Oracle Linux. With the right benchmarks, my minivan outperforms a Maserati. The T4 is a minivan."
Old news (Score:5, Funny)
Or are we going to see this story once per core?
Re: (Score:1)
Or even worse once per thread.
Re: (Score:2)
Are they out of order?
Re: (Score:2)
Gah. Beaten to it. By an hour.
Balls.
Re: (Score:3)
Yesterday reported the announcement of the T4, this story reports on an analysis of the announcement.
Really, this is just an example of the underlying truth: slashdot always reposts ArsTechnica stories
Re: (Score:3, Interesting)
An incomplete, and incompetent analysis by Ars Technica. Let's keep the facts straight here.
The T4 cpu isn't designed to work as someone's gaming rig cpu. It's not designed to run the Linux environment (thank god! - there's enough bad Linux servers out there, don't need any more - no it's not the hardware's fault, it's the OS itself, coupled with bad drivers from vendors)
It's designed to make workloads run on the Solaris OS.
With things like thread parking, lightweight processes, many threads per core, wir
Re: (Score:2)
Re:Old news (Score:5, Funny)
Re: (Score:1)
No mod points, but the P post had milk coming out of my nose. Okay, TMI...
Re: (Score:2)
Yesterday's was the final T4 post, but it's out of order.
That was because they used the FTL neutrinos ( http://science.slashdot.org/story/11/09/26/1330253/faster-than-light-particle-results-to-be-re-tested [slashdot.org] ) to send the post. Which validates CERNS results.
Re: (Score:2)
no .. yesterday's wasn't biased enough and had too much data
car analogy (Score:5, Insightful)
With the right benchmarks, my minivan outperforms a Maserati. The T4 is a minivan.
When you're moving lots and lots of boxes, then yes, a minivan does outperform a Maserati. It's a pretty good analogy IMHO.
As Seymour Cray noted: Anyone can build a fast CPU. The trick is to build a fast system.
Personally I've always found SPARC boxes to be good with I/O.
Re: (Score:2)
that depends on the dealerships in question.. i bet Maserati dealerships sell next to no minivans.. but then who would go to a Maserati dealer looking for a minivan?
Re: (Score:2)
I bought Mustang(ford) at a Volkswagen dealership at almost $2000 under blue book.
Mycroft
Re: (Score:2)
to be fair at most dealership of the price range of Maserati - if they do have a used car department they will normally limit it to cars of the same class, and the rest they blindly send to auction.
Re:car analogy (Score:5, Interesting)
Re: (Score:2)
Concur. As a storage guy, I would absolutely love to have one of these. I'm sure they're going to be pushing ZFS at this thing, hard. It's (almost) purpose-built. Nothing else can even come close to touching it.
Re: (Score:2)
Re: (Score:2)
fsck on a ZFS datastore? Maybe I misunderstood; as far as I know, that's (still) not possible.
When did you have these 100TB stores? That's still "a lot" of storage by modern standards, to some degree.
Re: (Score:2)
Re: (Score:1)
Re: (Score:3)
Re: (Score:1)
With the right benchmarks, my minivan outperforms a Maserati. The T4 is a minivan.
When you're moving lots and lots of boxes, then yes, a minivan does outperform a Maserati. It's a pretty good analogy IMHO.
As Seymour Cray noted: Anyone can build a fast CPU. The trick is to build a fast system.
Personally I've always found SPARC boxes to be good with I/O.
This was pretty much my reaction.
I think it's a great analogy as well. Sometimes, a minivan is what you need.
Re: (Score:1)
Well, funny enough the T4 also really IS a minivan. ;-P [wikipedia.org]
Re: (Score:2)
This is an Oracle product now.
It is worth noting that Oracle specifically dumped Sparc as it's reference platform for it's flagship product.
The ship has already sailed here. People already defected to x86 en masse. The real question now is whether or not this product is good enough to encourage people to move back to Sparc. Oracle's own treatment of Sparc as a "3rd party vendor" will matter a great deal as will their pricing model.
Oracle simply may be too greedy to make the whole package a great deal.
Re: (Score:1)
Re: (Score:2)
Exactly what rock have you been hiding under?
Attend an Open World or talk to the user community sometime.
Re: (Score:2)
The situation is though the boxes are now smaller relative to the car.
While the T4 is for higher workloads the fact is the standard x86 processor can handle most of workloads quite well. And the T4 usage is getting more specialized.
Back in the late 1990's the Ultra Sparc Chip was widely used for Web Servers and DB Servers. Because most companies needed a server that can 50 units a second. Today the standard x86 server can take 500 units a second and now say the T5 can take 5000 units a second. But today'
Re: (Score:2)
Meh, from the cars available in the ancient Test Drive III game, the Maserati was a real dog. The Lamborghini or even the Chevy prototype sports car was cheaper and faster... and the money didn't even matter in the game :-P
Don't blame Oracle (Score:1)
Well, not for this. The T range of processors has been GPLed for some time. You know the rule with Open Source - if you don't like it, you can fork it. True, the fab plant might be a problem, but if the design was good and seriously competed with the official version, you'd find someone to sponsor it.
I don't find the new T4 compelling. There's a lot to be said for the MIPS64 and the current POWER chips, and even the Itanium 2 has some novel features that are interesting. I also very much prefer the Transput
Re: (Score:1)
As an ex-Sun/now-Oracle CPU designer, I find your comment offensive. Although the focus of what we design has changed (from "innovative yet commercially unsuccessful" processors, to "steadily, incrementally better" processors), there are still a lot of smart people working here, both from "the old days" (the traditional UltraSPARC lines) and more recently brought in (from the Afara/Niagara lines).
Just like the Oracle software, the hardware (chips + systems) is expected to make money, not necessarily having
Re: (Score:2)
Linux isn't restricted to the x86. It will run fine on Alpha, PA-RISC, MIPS, Itanium, PowerPC, POWER, and most likely the T4 (it runs on pretty much every other SPARC out there). There are various flavors of BSD that will run on all of those platforms as well, including many others.
The question becomes why you'd want to run Linux on any of these. Or perhaps, why a hardware manufacturer would want to encourage you to run Linux on one of these. It might make sense to buy a high-end POWER based system to r
Better than nothing if you're locked in (Score:1)
It's easy for a company to get locked into an architecture when using home grown or proprietary software. I would bet that there are a bunch out there that really need an upgrade, and this will allow them to postpone an expensive and business threatening change for a few more years.
Oracle is extending the life of it's investment in Sun, but I don't see evidence that it is really developing it.
Sparc?! (Score:1)
SPARC T4 is only the beginning (Score:2)
SPARC T4 maybe late but not too little, it may only succeed in delaying the migration to Intel, AMD and IBM, but it will be more than capable to match them and by the time the T5 is out -if Oracle commits to it- Oracle will start to eat into the other players.
Re: (Score:2)
Re: (Score:1)
Re: (Score:2)
Except he has a point. Some organizations still have an investment in a Solaris infrastructure. For some small to mid sized companies the M series are a too expensive for what you get, while the T3 is too underpowered for a database. Hopefully the T4 will try to fill that gap, especially in the price category. Will the T5 eat into other platforms? I have my doubts there as I think this is all in the 'too little too late' category.
It's too bad really, as Solaris is a nice OS. Time to work on yet anothe
Re: (Score:2)
Now hold on here, don't mix up Solaris and SPARC.
SPARC has been dwindling for years - I know, I still manage several hundred SPARC servers. Solaris, if it's dying at all, has only started that path in the last year. Solaris on x86 is damned robust, and in a politics-free world, would be my preferred platform for most computing.
In order is sometimes worthwhile (Score:2)
In fairness, for some of the dense, massively multithreaded stuff that SPARC has been targeting in the last few years, in order execution gives you some power and transistor budget savings. Compare the Intel Atom, which ditched the out-of-order capability of their other mainstream processors.
But I was surprised to learn that Sun hadn't previously done out-of-order SPARC, although apparently Fujitsu have.
Re: (Score:1)
Compare the Intel Atom, which ditched the out-of-order capability of their other mainstream processors.
Yes and the performance of atom processors is quite shit because of that.
Re: (Score:2)
Yes and the performance of atom processors is quite shit because of that.
Only if you don't run enough threads. It's also slow because of low clock speed and limited cache, but hyperthreading recovers much of the performance lost due to in-order execution if you run two threads.
Re: (Score:2)
Well, that's the price some designs pay for power efficiency - various ARM implementations make this tradeoff too, which is admittedly a nicer architecture than Intel's stuff and offers good performance per watt. (http://stackoverflow.com/questions/4072183/which-arm-architectures-have-out-of-order-execution). Depends on the characteristics the design needs; for a workstation, in-order would generally be sub-par.
Ignorant article (Score:5, Interesting)
Sun had out of order SPARCs for years, contrary to the article's claims. Sun had a two pronged strategy, one aimed at single thread performance (the UltraSPARC series), the other at multithreaded performance (the T series). The UltraSPARCs were never really that good, so were eventually dropped in favour of the Fujitsu SPARC64 series, and the replacement (code named "rock") was dropped by Oracle because progress seemed stalled forever, but they did indeed have out of order execution, register renaming, and "Rock" had a promising "pre-execution" thread that was supposed to alert cache controllers ahead of time to pre-fetch data that can't be statically predicted, dropping cache misses to near zero.
The purpose of the multithreaded processor was to support mainly I/O bound tasks, and lots of them - web servers are like this, though more in the past where web content was more static. In those systems, a T series SPARC system noticeably outperformed similarly priced competition (with similar reliability - you could get a lot cheaper if you didn't care about component quality).
The single threading improvements in the T series are being added because even I/O bound systems often have compute-bound tasks. In particular, the T4 lets you assign one high priority thread which gets to hog CPU resources, in addition to out of order execution and other techniques that all threads benefit from, so I/O bound threads don't get hung up waiting for a single CPU-bound task to finish.
Re:Ignorant article (Score:5, Informative)
Sun's first out-of-order execution CPU was code named Millennium and was due out in y2k but it was late. They canceled the project in 2005. It was going to be called UltraSPARC V. Rumour has it that it was going to have a mode without register windows as well to further increase performance.
None of the Sun UltraSPARC CPUs (I, II, IIi, III, IIIi, IV, IV+) had ooo.
Sun just couldn't get ooo to work. Fujitsu had no problem, on the other hand. Their SPARC64 CPUs were miles better at the same clock frequency and went to higher frequencies too. Sun always made sure that you always got the last generation version of Solaris when you bought a Fujitsu box to make the Sun boxes more compelling.
Sun then gave up on "conventional" CPU design and went for the highly multi-threaded designs that they bought in from Afara (formed by ex-Sun staffers) and the Rock which turned out to be a dud. There was a good article about that after Oracle bought out Sun explaining why it wasn't a good design but I can't find the link.
So Sun did a deal with Fujitsu to re-badge their SPARC64 boxes with the latest Solaris on them...
I've no doubt that the T4 will be very good for certain loads. I know that my current employer bought a couple of T2 boxes as ClearCase view servers a few years back and the performance was abysmal since ClearCase doesn't scale well on multi-threaded systems. They had to be reassigned and replaced by M-series boxes with the (more conventional) Fujitsu SPARC64 CPUs.
And I'm very angry with what Oracle has done with Solaris 11 licensing. I've got a pile of old Sun workstations for playing with that have now become landfill. Oh well, my Solais skills can rot. It's Linux all the way now.
Re: (Score:2)
If you are really into zombies, you might want to disinter BSD. OpenBSD is really well supported on USparc, and you don't have to deal with Oracle. My Solaris skills were lost long ago, along with my last Oracle 7 install disk set.
Re: (Score:2)
Hmm.. OpenBSD. Sounds dangerous! Good idea, though. I want to keep some heterogeneity about for the purposes of writing good code.
Re: (Score:2)
The canceled UltraSPARC V and canceled Rock don't count because they never shipped. It's pretty sad that Sun, the inventors of SPARC, were getting out-designed by Fujitsu.
Re: (Score:2)
> The UltraSPARCs were never really that good
That's unfortunately true. I had an UltraSPARC I workstation, and a Pentium 133 could beat it for "desktop" workloads. Sure, disk, network, probably even memory was faster, but the CPU just didn't deliver. And it was crazy expensive by comparison. Sun just lost it around that time. They opened up to a point, and that delayed the decline, but Sun has been a shadow of its former self for a long time.
Re: (Score:2)
Was that Pentium 133 running a proper operating system or a glorified program-loader with lipstick on (e.g. DOS/Win 3.11 or Win95)?
That UltraSPARC I could whip the 133MHz Pentium at integer and floating point. Solaris was doing far more than DOS or Windows were at the time.
Now the DEC Alpha and MIPS processors of that time could make the UltraSPARC look slow.
Re: (Score:2)
Was that Pentium 133 running a proper operating system or a glorified program-loader with lipstick on (e.g. DOS/Win 3.11 or Win95)?
As long as the work gets accomplished effectively, does it matter what the OS is?
Re: (Score:2)
If the UltraSPARC is running a multi-user operating system with a multi-threaded kernel, sophisticated networking stack, an over-the-network graphics system and the other is running a 1970s-vintage CP/M rip-off (which is basically doing no work while anything else is running) and a 16-bit GUI or a Frankenstein 16/32-bit cesspool of DOS and goodness knows what else, then the comparison is hardly fair.
Re: (Score:2)
Except that -- even though I love and use Linux -- you don't actually need a a multi-user operating system with a multi-threaded kernel, sophisticated networking stack, an over-the-network graphics system to do "desktop" workloads.
OS/2 Warp and Win2K were more than adequate, as was MS-DOS with DESQview.
Re: (Score:2)
> Back then, was it really the Ultra SPARC CPUs that weren't suitable for desktop use?
Maybe, although I think I checked, and it was beefier than what the SuperSPARC (Sparc 20) had to cope with. Those were a generation before, and it showed, but at least at the time they were reasonably competitive (and a lot smaller). I was running an older SuperSPARC as a server for ages, until 10BaseT just didn't cut it any more.
Why are you asking me? (Score:2)
You're the reporter, why don't you do the research and report your findings? If you want a poll, then do that. But your audience isn't going to have the answer to that kind of philosophical question; that's your job.
BTW, my policy is to never read articles where the title is a question. But I'm such a nerd, I have to click on everything related to the T4.
Re: (Score:1)
Re: (Score:2)
I saw a 40MHz computer in 1990 and was in awe. I think it was a Mac IIfx. It had 640x480 graphics with 16.7 million colours, a 68040 processor and I think it might have had a DSP that could do CD-quality sound.
In those days anything above 33MHz was very special indeed, and I think we were just about getting to 1 million transistors on a chip.
My dad used to have a 286-12 laptop (Compaq SLT). I learned MS-DOS on that and a bit of C coding with Lattice C.
SPARC is dead (Score:2)
Sorry, for all intents and purposes SPARC is a dead architecture as is Itanium. Moving forward you'll have X86, X86-64 (AMD-64) and ARM...
Re: (Score:3)
That being said, SPARC is pretty badly behind, even compared to Itanium, and in the dust compared to last year's Power7.
Re: (Score:1)
x64 systems are here:
http://www.cray.com/Home.aspx
Re: (Score:2)
Re: (Score:2)
Nonsense.
Clusters do "perform on commercial workloads".
Infact, they perform better than big NUMA boxes. I've seen the migration away from Sparc NUMA machines to smaller clustered sparc machines myself. Some apps (most notably Oracle) actually do much better that way.
Re: (Score:2)
Clusters don't perform on commercial workloads.
What a blanket statement, I'm sure that all the people who do clusters or high performance computing (which is 99% commodity processors in clusters with custom high-speed interconnects) commercially would like to know they don't exist. In fact if you want performance for say a weather simulation or physics simulation or whatever where you can just fail nodes and keep going they're clearly the best. What they're not that good for is keeping one truth that is consistent and updated at all times, not distribut
Re: (Score:2)
Re: (Score:2)
No, more importantly why do you need it? You have to have 32 to 64 sockets because the T3's were so fricken slow... Slow Slow Slow. T4 is too little, too late and if you looked before the Oracle acquisition of Sun, it was the same way, a confused and faltering processor road map. Yes I can have hundreds of slow threads on my boxes or get faster Sandy Bridge Iron and buy a couple of hundred of those vs that one Sun box.
I went through it with the E class Iron a long time ago but sorry, SPARC is dead and
Re: (Score:2)
Re: (Score:2)
You need it for big databases that require massive amounts of I/O and memory bandwidth.
POWER.
Re: (Score:2)
> Too bad AIX is such a dog. It's the worst unix variant in existence.
Well said, although HP_UX was a close runner upper. Reliable - sure. Fast, elegant, user-friendly - no!
Re: (Score:3)
Where can I get an x64 or ARM system that scales to 32 or 64 sockets?
I don't know, but that'd have to be one big-*ss netbook.
Re: (Score:2)
Re: (Score:2)
What tasks greatly benefit from more CPUs in the same machine, as opposed to getting more machines with more CPUs and writing software that scales across them? You can put 80 Westmere-EX cores into one machine. To me, that seems like a fuckton. But I admit that I don't know much about scaling systems, so I'm curious why you would need more.
Re: (Score:2)
Because you're running Java servers, which typically use one thread per client?
This is SUN^WOracle, you know?
Re: (Score:2)
Where can I get an x64 or ARM system that scales to 32 or 64 sockets?
You can get them from SGI, although they're marketed for HPC, not for commercial apps. But keep in mind that you're talking about 256-640 cores; there's very little demand for such beasts. 80-core or smaller x64 servers are available from several vendors at reasonable prices and can satisfy 99% of the market.
The RISC/mainframe market is a totally different animal, one that those chips don't even play in.
And neither do the SPARC T series, which only scale up to 4 sockets.
Why bother with SPARC? (Score:2)
I can get a quad-socket T4 with eight cores per. That's 32 cores in a system.
Or I can get a quad-socket Opteron system with 12 cores per. That's 48 cores in a system.
Even better, the CPUs at 2.2 GHz and mobo for the latter arrangement can be had for under $6,000. Add the same memory and hot-swap hard drives as a 4-socket T4 and you're talking under $13,000.
The problem is that 3 GHz T4 costs over $90,000. I understand all the other stuff that goes into their server, but I doubt it's worth a $77,000 premium.
Re: (Score:2)
That's 256 threads maximum, that doesn't mean the hardware exists to process 256 threads simultaneously (not 256 cores).
In any case, 128 KB cache per core, AMD gets 512 MB per core. 4 MB L3 cache, AMD gets 12 MB (10 effective). Does it look like the T4 will be able to keep its pipelines relatively fed? Those threads need instructions and data.
Memory I/O? 25.6 GB/s for the AMD. What's the T4 have? Memory
Re: (Score:2)
Re: (Score:2)
Sequent was bought out by IBM in 1999. They had 32 processor machines. heck, before they got bought, they sold one that ran windows NT on a ton of Pentiums. IBM did pretty much nothing with them after buying them.
http://en.wikipedia.org/wiki/Sequent_Computer_Systems [wikipedia.org]
Re: (Score:2)
> Where can I get an x64 or ARM system that scales to 32 or 64 sockets?
NUMA, I assume? What is the point - once you have NUMA, the illusion of a consistent memory image is pretty useless. Clusters with fast interconnects exists (not cheap, but cheaper than SPARC), and they work perfectly well, so well that market is growing.
Re: (Score:2)
Good thing you posted as an AC since "intents and purposes" is grammatically correct. "intensive purposes" is incorrect. Now who's the "know-it-all?"
[John]
Re: (Score:2)
Re: (Score:2)
http://www.phrases.org.uk/meanings/all-intents-and-purposes.html [phrases.org.uk]
http://public.wsu.edu/~brians/errors/intensive.html [wsu.edu]
and of course, Wikipedia:
http://en.wiktionary.org/wiki/for_all_intents_and_purposes [wiktionary.org]
In the spirit of fair play, I did due diligence of searching for opposing opinions that would support your view, but came up empty handed.
Re: (Score:2)
Re: (Score:1)
Re: (Score:2)
We chose Solaris on SPARC T3 for media servers to drive a massive StorageTek SL8500 library because Linux on x86 can't keep up with the I/O. With real-world performance in excess of 1.5Gbit/sec, the latest T10kC drives with T2 tapes will bring many any backup media servers to their knees. And we can pump data to quite a few drives from a single T3.
Disclaimer: I work for Oracle because THEY pay ME to play with their giant toys :)
Re: (Score:2)
"many any" when I meant "many". I hate missing things on the preview...
Re: (Score:1)
Re: (Score:2)
You might want to look up "Larrabee" - there's a fairly significant thing preventing massive numbers of weak x86 cores being put on a single chip for handling thousands of small IO jobs: nobody cares.
Re: (Score:2)
Technology wise SPARC *was* ahead of Intel and IBM (Power Architecture) but not anymore.
Even Oracle acknowledges [crn.com] that their hardware business is suffering, one of the reasons Sun was a prime candidate for acquisition in the first place. The decline of SPARC is lamentable but no more than say the death of most RISC systems and manufacturers. Oracle is mentioned as something that has to be on big iron, so that's why they bought Sun? Look, Ellison making hardware makes about as much sense as them pushing th
Re: (Score:1)
Re: (Score:2)
Wait, wait, wait, comparing that SPARC based *system* to something you can buy from Oracle is like saying you can buy an F22 from General Motors. That is a highly modified system and has SPARC processors but it's nothing that Oracle provides as a SKU, unlike something you can get from IBM. But since you brought up the Top500, I notice that #2, is the Chinese Tianhe using what? NVIDIA GPUs?!?!?!
SPARC is dying, Sun couldn't sell enough of them because while they were "good" they were behind the curve in te
Yeah, but ... (Score:2)
Well, like all benchmarks, it depends on your needs.
If it's getting 6 kids to soccer practice in a vehicle with a high safety rating, then the mini-van is more suited to your purposes.
Of course, what CPU functionality in this car analogy corresponds to having the mini-van be preferable to a Maserati ... I don't think I can answer that. :-P
Re: (Score:2)
No (Score:1)
Minivan vs Maserati (Score:2)
And they sell thousands of times more minivans than Maseratis. Selling minivans isn't a bad business to be in at all.
Long live SPARC (Score:2)
It was great to have you around for the party. Too bad eventually Oracle got a hold of you.
At this point, even if it was the most awesome CPU ever designed, who wants to touch it now its attached to that evil corporate monster ( that has no choice due to legacy system investements )?