Sun plans to ship 8-core and 6-core versions of the Niagara processor in single socket servers before year end, according to sources. The decision to produce 6-core parts came as Sun tried to make best use of chip yields from its partner TI. In addition, Sun can now offer a couple of different prices with the Niagara-based systems, pitching the 6-core gear as lower-end. Based on 1.0GHz and eventually 1.2GHz chips, a Niagara server will consume on average 250 watts – versus close to four times that amount for similarly performing Xeon-based four-way servers.
From what I’ve seen of them, processor performance, cost and power consumption are all lightyears ahead of the competition.
This, coupled with the return of Andy, could turn Sun’s fortune’s around.
“From what I’ve seen of them, processor performance, cost and power consumption are all lightyears ahead of the competition.
This, coupled with the return of Andy, could turn Sun’s fortune’s around.”
mod -1: astroturfing
you work for Sun, and post this crap?
Have to ask, whats with the link? it just links to itself.
Fixed.
Whoever wrote the story, the HTML is href=”” ” rel=”nofollow”>http://www.theregister.com/2005/09/09/niagara_many_cores=””, it should have only one “quote mark” per a side, and no space between the first quote mark and http. That is the correct web adress for those trying to get to it.
Adrian
Also, the trailing = has to be removed http://www.theregister.com/2005/09/09/niagara_many_cores/ is the real correct address
From what I’ve seen of them, […] lightyears ahead of the
competition.
Since have seen absolutely nothing of them, this statement
seems perhaps undersupported.
I’ll withold judgement until they ship. Some things seem really
exciting: using hyperthreading (or whatever sun calls it) of multiple
instruction streams instead of speculative execution in order to keep
the execution pipes full seems like a good trade to me.
The clocks (1 GHz?) are very disappointing, and indicate that the
coordination with the TI fab isn’t working so well. Ditto for the
decision to ship 6 core parts to “make the best use of chip yields” —
which might be better translated as “OK, yields suck, but we can turn
off the broken cores and ship it anyway”.
And remember that the zillion-core tradeoff is only a win for some
workloads. Even for servers, having a single high-performance thread
is important for many tasks (locked database updates, for example, or
the final link of large C programs). The Niagra (which is an in-order
execution design) is going to perform something like a 733MHz P3 on
these tasks.
Anyway, I’d plug your ears to the spin until these things ship. It’s
a really neat idea, but it’s a huge risk for Sun.
It’s a chip designed with a server workload. It’s a safe assumption that you’re going to have many many threads running at the same time, at least one per connection. It’s the same story as always though, multiple threads working on common data are going to need locks and such which hurt performance. Web performance should be nothing short of astounding as far as I’m concerned. Remember serving up a website is basically listening for a connection, making some reads from a database and constructing packets. No locking, just reads from memory/disk and that’s it.
Seems you still don’t get it.
It’s not about how many Ghz a processor’s clock can have, it’s what you can do with all those Ghz..or Mhz!
Google for EV6 and learn something today, before start whinning on things you just DON’T KNOW..
My 500mhz ultrasparc does very well for my day-to-day stuff… not entering on the server-side of the talk, of course…
It’s not our fault if Intel is wasting power (Watts) and cicles (Ghz) with all those stupid processors..I mean, wait a moment… Yeah, stupid processors for stupid people. go buy one of it @compusa, I’m sure Dvorak have already told everyone they are hype! 😉
Intel rules p4 Rockz and linux is UNIX(tm)
using hyperthreading (or whatever sun calls it) of multiple instruction streams instead of speculative execution in order to keep the execution pipes full seems like a good trade to me.
Yep. It would be nice to see a bit more detail about that. E.g. does each core actually execute instructions from multiple threads at the same time, or does it just switch very quickly between threads (like Intel’s next Itanium implementation).
The clocks (1 GHz?) are very disappointing
Yes, but why is that? I mean, no-one is expecting Prescott style GHz madness, but Pentium M, Athlon and PPC achieve well over 2GHz with moderately sized pipelines. Does the Niagara perhaps have a classic four-stage pipeline design?
Google is your friend:
http://www.aceshardware.com/read.jsp?id=65000292
http://www.sun.com/processors/throughput/faqs.html
http://www.theregister.co.uk/2005/09/07/sun_niagara_details/
http://www.sun.com/processors/throughput/datasheet.html
http://www.mdronline.com/watch/watch_abstract.asp?Volname=Issue~*~@…
I work at Sun.
So, uh, quiet down, please?
Secondly – I’m not going to bother arguing, but you obviously don’t understand Sun’s focus with this chipset. It’s designed for Solaris and for systems that are heavily multithreaded – it’s entirely designed for parallelism – not high operating freqeuncy.
They haven’t yet started shipping these, but they are letting some customers test ealry version of them.
At my company we’ve gotten the chance to try one. The machine was a little bit flaky, which is understandable since it’s not a finished product, but the performance was impressive. We could do things that before required a lot more CPUs and also bigger machines.
You’re right about it not being very fast for a task that requires a single thread to do all the work, but that’s not an issue for us. So for what it’s designed to do, it’s probably about as good as Sun says it is.
We’ve been using multi-CPU sun servers for a long time and if the Niagra means that you can do more with one CPU then that’s good news. Sun could have provided the same performance with their bigger servers, but this way they can offer good performance at a much lower price, which is something they start to realise is important.
Since have seen absolutely nothing of them, this statement
seems perhaps undersupported.
> Your counter statement is also under-supported.
The clocks (1 GHz?) are very disappointing,
> go to Intel Dev Forum slides. There is 1 slide that shows the impact of cache miss. the 1GH performs much better than the 3GHz.
and indicate that the coordination with the TI fab isn’t working so well.
> need more supporting facts. The statement is far too general.
Ditto for the decision to ship 6 core parts to “make the best use of chip yields” —
which might be better translated as “OK, yields suck, but we can turn off the broken cores and ship it anyway”.
> This is a common manufacturing technique. In the chip industry, the output will more for lower frequency, as the yield not so good at high frequency. In the disk industry, you can ask Seagate for some info. It’s the same technique.
And remember that the zillion-core tradeoff is only a win for some
workloads. Even for servers, having a single high-performance thread
is important for many tasks (locked database updates, for example, or
the final link of large C programs). The Niagra (which is an in-order
execution design) is going to perform something like a 733MHz P3 on
these tasks.
> Niagara is not for everyone. But for most people, it’s performance is good. I’ve seen various benchmarks shown by Sun. If you are a customer of Sun, you can ask Sun rep to show the internal results.
Anyway, I’d plug your ears to the spin until these things ship.
> If you are customer, you would appreciate the early info for DC planning purpose.
It’s a really neat idea, but it’s a huge risk for Sun.
> agree, if Niagara is the only chip Sun has. For other workloads, there are other chips. Sun is not 1 chip company.
> In general, you should check with Sun (again, if you are customer, you can get the internal info) before making such statements.
Will BeOS run on these?
Will BeOS run on these?
I’m not sure if this is a serious question– but I’ll answer it anyway.
The Niagara chips are SPARC chips, so the short answer to your question is: no.
There is also a long answer. If you by BeOS mean Haiku, then it is possible that someone will be willing to a) buy a Niagara-based server (costing tens of thousands of Euros, probably) and b) port Haiku to it when it’s ready. If that happens, then yes, BeOS will run.
Sun is usually nice to porters, and gives away or sells at heavily discounted prices, to developers intent on doing a port of their software, to their hardware or OS (Solaris). I am keeping it vague (“hardware”, “OS”, both etc.) as I know of multiple examples of this kind of cooperation.
However, a single-chip 1U Niagara server might probably not cost “tens of thousands” of euro. Ten thousand, maybe, 20.000 or more, I don’t think. Just based on past experience, Sun has never ever priced any of their 1U servers more than about US $10.000.
For example: Sun Fire V20z, EXTRA LARGE configuration, US $5.795. Sun Fire V210 LARGE config, US $8.845. V120 LARGE is only US $2.845. V100 LARGE is US $2.195.
My guess: an entry level, 6-core 1U server (“T1000”?) will cost about US $4000-5000. With a little Sun discount (rather easy to get), suddenly a Niagara machine seems very accessible.
My guess: an entry level, 6-core 1U server (“T1000”?) will cost about US $4000-5000. With a little Sun discount (rather easy to get), suddenly a Niagara machine seems very accessible.
Even so, their first chips will be for blades and U1 and U2 – the fact remains, these chips are not SMP capable, that doesn’t occur until next year when Rock is released – this is mainly being targeted at the telco market, which has actually had a bit of a rebound over the last few years with the possibilities of 3G finally being relised and passed on to consumers at reasonable prices rather than the astronomical costs there were at the beginning.
I think the interesting thing will be whether later on, as the size of the core decreases more, and thus left with the extra space, will they soup the core up to a UltraSPARC III? with that being said, with the SPARC V from Fujitsu, will that eventually replace the existing UltraSPARC processor? and as for Opteron, will they eventually replace their SPARC stations with wall to wall Opteron workstations?
I think there are alot of still unanswered questions, which need to he addressed by SUN so that customers know where SUN is heading in atleast the next 5-7 years.
The information is available, if you sign an NDA (I did).
The information is available, if you sign an NDA (I did).
Which is kinda useless; have you *EVER* tried to work through the maze of beaucracy that is SUN Microsystems? sure, it isn’t as bad as say Microsoft or IBM, but non the less, it still is rather cronic.
Try HP!
HP is a company?
Going by their red tape, they’re more like a government department in a communist country where by more processes are added as to provide employment in the public service
Even so, their first chips will be for blades and U1 and U2 – the fact remains, these chips are not SMP capable, that doesn’t occur until next year when Rock is released – this is mainly being targeted at the telco market, which has actually had a bit of a rebound over the last few years with the possibilities of 3G finally being relised and passed on to consumers at reasonable prices rather than the astronomical costs there were at the beginning.
I didn’t ever say they wouldn’t be for blades and U1 and U2. And I never said these chips are SMP capable.
I agree with everything you said, I just don’t understand why you started with “even so”, like you had a coutner-argument.
If Haiku was to be ported to SPARC (how likely this is? I don’t know, I won’t speculate), it would probably run perfectly well on a Niagara machine. BeOS is “pervasively multithreaded”, it’s API in fact encourages multithreading. Niagara seems like the perfect match for it.
The only problem would be if these Niagara machines didn’t have a framebuffer. Umm… yeah, that’s likely.
Sorry about that; I tend to be rather crappy in the english department
It would be interesting to see, however, a developer desktop/workstation – but then again, these are 4/6/8 UltraSPARC II cores – not exactly world enlighting performance machines in regards to workstation/desktop performance.
You are right, we probably won’t see a workstation based on (this) Niagara. In fact, I am guessing that Sun won’t make too many SPARC based workstations for now, instead churning out Opteron workstations, like the Ultra 20. Which, by the way, is a damn good machine, and well-priced, too.
Definately; it will be also interesting what their two other Opteron workstations will look like once they’ve updated and upgraded it to the new funky Andy style workstations.
I’d say that its only a matter of time before they mothball the SPARC workstations – its just playing the waiting game for software vendors to start providing x86 versions of their software on Solaris – once that has done, the sales should decrease to a point where SUN will say, “it isn’t viable enough for us to continue to develop SPARC workstations any longer”.
What I hope though, is that they, along with others adopt EFI in place of the current BIOS – if there was ever a source of problems on the x86 platform, it would be the BIOS and the numerous attempts to provide a unified power and configuration management standard.
it will be also interesting what their two other Opteron workstations will look like once they’ve updated and upgraded it to the new funky Andy style workstations
Well, I guess we can predict that pretty accurately, at this point:
compare the new servers:
http://news.com.com/2300-1010-5845883.html
with the new workstation
http://www.sun.com/desktop/workstation/ultra20/index.jsp
From here it’s easy to extrapolate what the new look will be. It will be a good look!
It will be interesting to see how long they will maintain it. Sun has been developingthis tech for several years. Some of their top scientists were involved in this development, so it’s probably not simple to reverse engineer. But expect Intel developing something half-assed and present it as “the same thing”, like with their multiprocessor Xeons.
Anyway, Sun should have another go at blades. They would be all over the IBM blades that consume an insane amount of energy, and put a giant strain on the computer room’s cooling system.
Or you could cluster 32 VIA boxes for the same price. But don’t expect to do that with Solaris just yet. Maybe OpenSolaris, if someone knows how to do this then it will be done.
I originally understood from the Afara stories that the threading was fine grained, FG MTA ie every 4 clocks, always 4 threads each get 1 issue into the pipline, and no other options. That means that the pipeline really simplifies with possible removal of hazard detection logic, register forwarding and it also 1/4s all the latencies the threads see. This is the form I like as it allows the silicon to keep speeding up (but maybe not at TI).
Now this article suggests the other form of threading, a thread keeps running until it hits a wall ie cache miss and then swaps out to let another get a try. This one I don’t like as it means that each thread forces the pipeline to have all the same complexity that a single threaded processor has but at least can carry on with occasional swaps.
I’ll have to go do some digging.
Any way the real issue with Niagara and also Raza’s MIPS 8 core 4way threaded cores is the memory model. That should be threaded too using massive interleaving DRAMs with a scheme to spread memory requests over memory banks to avoid hitting only 1 bank. With RLDRAM its actually possible to issue DRAM requests every 2.5ns but I don’t think anybody else has figured to use it yet. If DRAM with fully random bank selections can hit 2.5ns (best case) issue rates, that actually faster than L2 SRAM cache and it changes everything.
transputer guy
I don’t think Niagara would be the best option for Haiku esp if an 8 node 4 way cpu goes for $10K, probably much less but nothing like say $100 that typical modest Athlons go for. Since BeOS runs fine on a 300MHz P2, why ask for the moon.
On the other hand I will be describing an N node 4 way theaded (FG MTA) in a few days at cpa2005 that only runs at 300MHz in FPGA and for a slower 200MHz or so version costs about $1 per node (in volume) of FPGA resources. The core is really similar to Niagara in many respects but mostly by accident. The MMU actually is intended to use 3.3ns issue DRAM requests so each node more or less see no significant memory access delays other than a few instruction slots for loads,stores. Each 300 MHz node compares to an xp300 with out most of the complexity of x86. An ASIC version could probably hit 1GHz with suitable funding.
For BeOS/Haiku I would love to see it boot on an N node processor array but thats a quite along ways off.
transputer guy
I would like to know more on the implementation of these cores in FPGA. Which FPGA; how many cores, etc. Some juicy details, pleae :o)
All the details at the cpa2005 conference are at
http://www.wotug.org/cpa2005/
The abstracts are up now, the full pdf papers (and slides) some time after its over sep 21.
and micron.com or rldram.com for the memory details
johnjakson at usa_dot_com
transputer guy
What makes solaris so much better at thread scheduling than Linux, FreeBSD, Dragonfly, etc? Why can’t these operating systems compete favorably on Niagara systems?
NewOS (from which Haiku was forked, to make it more like BeOS) may see a 64-bit UltraSPARC port at some point, according to the NewOS homepage.
I’ve got the same hardware Travis is or was planning to use, a Sun Blade 100, but mine’s unfortunately stuck on some bad NVRAM. (Suits me for buying used hardware on eBay.) I’ve been googling like crazy for a fix, but none of the OpenBoot voodoo I’ve tried have had any effect, and Sun’s website isn’t helping.
If anyone’s got experience with changing the battery and/or NVRAM of a Sun Blade 100 I’d be very grateful for some pointers in the right direction. My email is jonas@(see username).com.
I have replaced the NVRAM/rtc in a couple of Sony News workstations. I know that the same NVRAM IC works for most if not all Sun workstations as well.
If you are sure it’s the NVRAM, then open the chassis of your Blade 100 and check the make and model of your NVRAM – it’s veryeasy as it’s the only “fat” IC on the motherboard. It’s a rather tall PDIP (plastic dual-inline package).