49 posts • joined 19 Mar 2011
Re: Not supercomputers - not even close
They don't get put in supercomputers. The last paragraph is referring to something completely different (the SPARC-based Post-FX10 processor) which *is* a supercomputer processor, but the GS21 chips mentioned in the first several paragraphs are legacy 31-bit processors for mainframes; high compute performance is not (and hasn't been for decades) a major concern for these systems. The focus is on reliability and long-term binary compatibility, not compute performance.
Not supercomputers - not even close
These are proprietary mainframe processors, not supercomputer chips. There's a very, very big difference there...
It's Solaris 11. OpenSolaris is a dead project and the source press release doesn't mention it anywhere.
BS2000 was originally Siemens (well, originally, it was RCA - but that was a LONG time ago) but was shifted into the Fujitsu-Siemens joint venture, which was acquired by Fujitsu a few years ago.
The Soviet EVM mainframes stopped being sold in the mid 90's due to being phenomenally slow compared to western machines.
They still are, technically - Fujitsu sells 390-compatible mainframes in Germany and Japan (the Business Server and GS21 lines, respectively.) The GS21 runs an MVS-compatible OS called MSP, while the Business Server family uses a custom, non-IBM-related operating system (BS2000).
Re: can't resist
IPF is listed under Intel, not HP. The last IPF system on the list - NASA's Columbia Altix cluster - remained in the list until June 2012.
The SPARCs in the list aren't Oracle, and are only vaguely SPARC - they're clusters of non-SMP-capable Fujitsu SPARC64 parts with custom vector extensions.
Re: Big memory
Oracle has never claimed that M6-32's can be linked that way, and neither their HotChips presentation nor their M6 marketing materials claim that it can. "Bixby supports 96 sockets" doesn't translate to "you can glue together arbitrary Bixby servers to make a 96-socket machine" any more than "Boxboro-MC supports 8 sockets" means you can put 4 rx2800's into a single system.
Re: Big memory
Bixby is an internal interconnect ASIC. While Bixby itself scales to 96 sockets, this isn't a NUMAlink-type system and a customer can't just plug multiple M6-32 systems into each other with it.
That being said, I expect M6-64 and M6-96 will be coming within the next few months.
Re: Just Imagine
I've owned an N810, an N900, and an N9, which I used as my primary phone for over a year.
The N810 and N900 are fun but very much not fit for general consumption. The OS has a lot of gotchas, is not particularly finger-friendly in places, and occasionally has unexplained crashes of core components a la Windows Mobile. That being said, for a power user who's familiar with Linux, the N900 kicks ass.
The N9 (which runs Meego) is both frighteningly generic and quite limited - far from the wide-open Linux phone I think most people imagine. It can only load signed firmware, you can't downgrade your firmware, and only runs signed binaries unless you do some hacks (Inception). It also has a number of bizarre bugs - for instance, the camera often saves images with a temporary filename and file extension, which means they don't show up in the gallery. The browser really sucks (slow *and* unstable is a bad combination, and it doesn't even do text reflow).
Overall, I think Meego wouldn't have been Nokia's savior without a lot more work. Windows Phone may be limited in some ways, but it's far, far beyond Meego in actual readiness.
He made it clear that that's the package size, not the die size.
"An efficient x86 emulator could move workloads onto Power iron, and that was the idea behind the QuickTransit buy, supposedly, but Big Blue has not really done much with it."
Except... uh... exactly what you're suggesting?
Why is it that when the Republicans blocked this, everyone was commenting on the article talking about how evil, evil, evil the Republicans were for it - but now CISPA has apparently become the devil? The change between the comments on this page and on http://www.theregister.co.uk/2012/08/02/senate_blocks_cybersecurity_bill/ is fascinating...
List prices have now been posted. $53k for the lowest-end 2-socket system seems... optimistic, compared to what IBM is selling 2s Power systems for.
CPU's look decent, though. I'm interested to see how the M5-16 is positioned relative to the T5-8, since core performance on the T5-8 should be slightly higher.
Itanium is a direct evolution of PA-RISC, and there's a strong commonality from PA2 instructions to Itanium instructions - which is why execution of PA2 binaries on Itanium is pretty fast.
The core design, on the other hand, is fairly different - although some design factors like HP's tendency to like processors with huge chunks of fast SRAM and small cores survived. But the PA-8K design was probably nearing the end of its productive lifespan anyway, and at the time a lot of Itanium's design changes (OoO sacrificed in favor of wider issue) seemed like reasonable choices.
Given that Itanium roundly trounced PA-8K in benchmarks from the beginning, I have a hard time qualifying the move from the PA-8K design as a bad idea. That being said, IPF is pretty obviously dead as a doornail at this point.
Re: A Bit Lame
IBM consistently _does_ publish benchmarks, and yes, they're faster than Xeons - especially for multithreaded or cache-sensitive workloads.
It's plenty clear.
"It is still not entirely clear if the M4 chip is another name for a future Sparc64 processor from Fujitsu, as most of us expect."
It's Oracle. And not particularly high-spec.
Re: What commie iron?
The Mainland has a system with custom processors and custom interconnects in the top 100 - the ShenWei SW1600-based "Bluelight" machine. The cores themselves are supposedly distant derivatives of early Alpha cores, fabbed at 65nm, with 16 to a chip and clocked at 1GHz.
That being said, I don't see anything with it that's inherently more appealing than Fujitsu's pseudo-vector-processor system.
No. Just... no.
"For the M3, you get a total of eight threads running at 3GHz, or a combined 24GHz of aggregate clocks, and for the M4 you get a total of 32 threads running at 4.5GHz, for a combined 144GHz of aggregate clocks."
Not how SMT works. If you really want to do this "aggregate clock" thing, the numbers you're looking for are 12GHz and 72GHz..
Increasingly skeptical of M4 being Oracle
The new Hot Chips program, posted a few days ago, mentions a 16-core "SPARC64 X" for the "next-generation UNIX servers." This would seem to imply that either M4 is actually a Fujitsu processor, or that Fujitsu's going to keep making its own CPU's for its own Solaris servers without Oracle on board.
Re: Cell fail?
Those are PPC, but not Cell-based. No SPE's involved.
What's so terrible about Itanium? It's faster than one of its two major competitors and continues to sell.
Re: Eddie Edwards
"Except the ARM mentioned has out of order..."
So did the P4.
So did the P4.
Yep. And the P4 issued 3 instructions per cycle, rather than 2.
"usual vague references to the microcode decoder"
Almost every processor - including many ARMs - have microcode. Performance impact is negligible.
"faster RAM interface"
Judging by ARM memory performance thus far, I'm extremely skeptical.
Nonstop doesn't run on the BL860c i2 itself - as far as I know, it only runs on NonStop-branded blades with ServerNet cards.
What makes you think that the new M-series part is SPARC64? I don't recall them mentioning that in the announcement event, and it seems just as likely that it's a clock-boosted (4-5GHz, perhaps) 8-core 28nm T4.
Anonymous Coward is right. You seem to like posting any kind of meaningless numbers that make Oracle look good. What are you trying to accomplish?
Bada doesn't run on Linux, but on a licensed microkernel called Nucleus. There's no technical reason that it couldn't be a Linux-based platform, but right now, it isn't one.
The T-series processors are for 1-4 socket systems, which is a space that x64 does just fine in. Oracle/Fujitsu do have the SPARC64, which scales to 64 processors, but it's not exactly the fastest chip in the industry.
Mainframes and high-end RISC servers have supposedly been on the verge of death by x86 since the Pentium Pro. At this point, repeating those claims is just comical.
Re: IBM "compatible"?
The Bull and NEC systems are certainly not IBM-compatible, as they're Itanium systems emulating the old 36-bit GE/Honeywell platforms and running GCOS or ACOS.
Fujitsu and Hitachi's systems are technically /390 compatible, but as far as I know there's no way to license an IBM operating system for them. Fujitsu's Japanese-market operating system (XSP and MSP) are very similar to VSE and MVS though.
Unisys isn't the only non-IBM mainframe supplier...
Fujitsu, Bull, Hitachi, and NEC all have mainframe businesses. I don't know enough to comment on Hitachi and NEC, but Bull and Fujitsu TS both have solid customer bases in Europe, and Fujitsu and Hitachi evidently have decided that mainframes are worth enough for them to continue fabbing their own mainframe CMOS processors.
"MIPS" in the context of mainframes is a specific benchmark, only directly comparable to other hardware of the same type. z MIPS do not correspond to anything like "millions of instructions per second" on x86, ARM, or anything else of the sort. z MIPS don't even correspond to MIPS on other mainframe platforms, like OS 2200, MCP,, and others.
As far as I know, the Westmere-EX is still only an 8-socket processor - where is all of this talk of 32-socket support coming from?
Missing the point...
The Xeon E7 is a nice chip, and I think it has a real future for four-socket systems. The problem is that the real benefit of using the proprietary systems comes at eight sockets and up, an area where Power (and potentially SPARC and Itanium, assuming they get their act together) is still compelling.
Why should this be used in commercial chips?
The oomph of the SPARC64 VIIIfx comes from a custom-designed HPC-oriented vector instruction set called HPC-ACE. The scalar components are similar to slightly tweaked, cache-starved, low-clocked versions of the existing (slow) SPARC64 VII core. An 8-core SPARC64 VII at 2GHz with a smaller cache doesn't exactly sound like the Holy Grail of commercial computing, so I can't see why they would be "fools" as you said not to commercialize it.
Re: MS and RH
Microsoft and Red Hat had very few (a few percent) total customers on the Itanium platform. Oracle develops the killer apps for all three HP operating systems (Rdb, Oracle, and Tuxedo.) The impact is orders of magnitude greater.
Those systems offer a unique value
Core performance on Power and Itanium has been consistently good or very good for most of the last decade, but that's never been what you pay for. Ever since the Pentium Pro, Intel has offered most of the core performance of RISC platforms at a fraction of the price. What you get with the proprietary UNIX boxes is a fast system that scales far higher than most commercial x64 solutions, with very high reliability and certain features (PowerVM, for instance) that are frequently worth the premium you pay. Judging UNIX systems by list prices is also silly, since most actual purchases involve big discounts.
Core performance on all three major RISC platforms (Itanium, SPARC, Power) will probably make massive leaps in the next few years. Oracle's public performance targets are especially aggressive, although it remains to be seen whether they can make it happen.
Not gonna be Apple
There's no way Apple is going to break binary compatibility with their existing application base, and the Atom really doesn't have close to enough oomph to provide flawless emulation of software designed for a Cortex A8 or A9.
The list price of a PS703 with two 2.4GHz processors (16 cores) and 32GB of RAM is $13k. That isn't impressive, given that you can already get a PS702 (yes, slightly lower density) with two 3GHz processors and 64GB for $16k list.
I wonder how steep the discounts on these are.
Re: Die die die.....
You're missing the point. Power, zArchitecture, and SPARC aren't in consumer parts either, but they're all arguably successful architectures. Itanium is a decent commercial-workload processor, currently hobbled by Intel's insistence on fabbing it on obsolete processes. Poulson should be competitive with Westmere-EX when it gets released, and depending on clock speed might be a bit faster.
GCOS is a Bull product (actually two unrelated ones, GCOS 7 and 8), not a NEC one. NEC ACOS is derived from GCOS (ACOS-2 is based on the defunct GCOS 6, and ACOS-4 is based on GCOS 7) but it is not the same thing.
Itanium is faster (when Poulson comes out, theoretically) and more scalable. Fast x86 chips have been around, and billed as "RISC killers," before. They still stay largely constrained to the low end.
Billl, why do you assume this is an Itanic Killer? Poulson should be similarly fast or a little faster when it comes out.
Larry is insane
This seems suspiciously like a case of "We can't make a compelling hardware platform, so lets hamstring those who can." It may seem like a good idea in the short term, but I suspect most of the customers who are going to be forced to migrate because of this will be extremely wary of doing so to Oracle hardware - DB2 may prove to be the real winner here. If IBM were smart, they'd re-energize Informix and/or DB2 for HP-UX and start trying to poach Oracle customers that don't want to switch OS.
Not that simple
This update isn't as big as I, and at least some NSK customers I've worked with, would like; it would be nice to have an increase in the maximum number of logical processors, for instance. That being said, saying it's "basically nothing" is just wrong. The big I/O and memory bandwidth boost alone is enough to make a difference, and the doubled amount of cores helps too. I doubt that existing IA64 BladeSystem customers will view this as a critical upgrade, but older MIPS and early Itanium NonStop customers are probably looking hard at this.
There's also the advantage that its socket-compatible with the next-gen Poulson processors. If Intel can get these out on time (a big "if" with Itanium), and assuming Power7+ doesn't blow anyone away, these should be the fastest general-purpose processors available when they come out.
That's quite a mouthful, Kebabbert. I'm going to try to address the silliness here bit by bit.
"Awkward it is, that we read about the TurboHercules Mainframe emulator that gives 3.200 MIPS on an 8-way x86 server - that is quite decent performance for a fraction of the price."
Wow! Roughly 6% of the CPU performance of a mainframe, a system designed for moderate CPU performance and high I/O! Truly impressive!
"TurboHercules says their customers are interested, because when their Mainframe crashes, they can switch to an cheap x86 server."
I've spent a fair amount of time with TurboHerc marketing materials, and they don't claim this. They say that a commodity system can be a warm standby failover system. Nobody, even NonStop or Stratus customers, puts all of their eggs in one basket when they can avoid it.
"And the "worlds fastest cpu", z196, is in fact slower than an Intel Nehalem-EX. You need several z196 to match one Intel Nehalem-EX in terms of cpu power."
Floating-point, maybe. Integer, not close. I/O bandwidth, orders of magnitude of difference.
"And about the "slow" SPARC cpus, sure they are slow, says IBM. Let me see, who has the world record in TPC-C? 30 million? Who has several other performance world records, beating the POWER7? It is funny that a 1.6GHz cpu can be fastest in the world on several benchmarks, even faster than the "mighty" POWER7."
TPC-C is a meaningless benchmark, and has been for years. Sun wisely abstained from it, recognizing that it was basically silly. Additionally, I've been hearing some really interesting things about the supposedly invincible T3, although I haven't used one myself...
-Oracle reps taking months to provide a price quote, resulting in customers having to compare Oracle list price to HP/IBM discounted price
-Performance on throughput-oriented multithread workloads being vastly slower than a cheaper AMD Magny-Cours
-Failure by Oracle to provide credible roadmaps for >4socket machines
Anyway, if you insist on using TPC, the fastest non-clustered result is a Power6 machine, followed by a Superdome. If you do some math, it can be assumed that a single Power 795 would be well into the 15mnTPC territory. How many boxes did it take Oracle to get to 30mn again? And how many licenses are they selling, for that number of cores?
IBM seems to be so laser-focused on Oracle's slow SPARC and commodity systems that they're ignoring what is probably the highest-scaling general-purpose server - the HP NonStop. Just because HP is too stupid to market it anywhere outside its happy little high-margin telecom niche doesn't mean it doesn't have potential.
- iPad? More like iFAD: We reveal why Apple ran off to IBM
- +Analysis Microsoft: We're making ONE TRUE WINDOWS to rule us all
- Climate: 'An excuse for tax hikes', scientists 'don't know what they're talking about'
- Analysis Nadella: Apps must run on ALL WINDOWS – PCs, slabs and mobes
- Yorkshire cops fail to grasp principle behind BT Fon Wi-Fi network