The Register® — Biting the hand that feeds IT

Feeds

Magic Quadrant for Enterprise Backup/Recovery

Skaugen said that the Nehalem-EX chips would allow server makers to create two-socket servers that support up to 512GB of main memory, nearly three times as much as AMD can do using 8GB DIMMs with the Magny-Cours Opteron 6100s announced yesterday. Intel will be able to support 1TB of main memory in a four-socket configuration, while the controller inside the Opteron 6100 only allows a four-socket machine using these chips to address a maximum of 512GB.

Skaugen rubbed it in a little that Intel's Nehalem-EX partners had over 50 new products in rack, tower, and blade form factors, and that it had 75 per cent more four-socket designs than with any prior server chip launch in its history. A dozen OEM partners have 15 different servers in the works that will span eight or more processor sockets, and apparently some are pushing their designs up to 16, 32, or 64 sockets.

The big bad box at the Nehalem-EX launch, of course, was the Altix UV massively parallel supercomputer, which El Reg told you all about last November. The Altix UV machines allow for up to 2,048 cores (that's 256 sockets and 128 two-socket blades) to be lashed together in a shared memory system suitable for running HPC codes. The shared global memory is not the same as a more tightly coupled symmetric multiprocessing (SMP) or non-uniform memory access (NUMA) cluster used in general purpose servers for running applications and databases. But that said, the Altix UVs are very powerful machines indeed and are intended to scale to petaflops of performance.

The Boxboro chipset that Intel is shipping as a companion to the Nehalem-EX chips supports configurations with two, four, or eight sockets gluelessly. If you want more sockets than that, you have to create your own chipsets, as HP, IBM, Silicon Graphics, and Bull have done for sure and others will no doubt follow.

But you can't just plug any old Nehalem-EX chip into any old configuration. That would be too simple, and Intel likes to charge premiums for features, like most capitalists. Take a gander at the feeds and speeds of the Nehalem-EX lineup:

Intel Nehalem EX Table

The Intel Nehalem-EX Xeon 7500 and 6500 processors

The first thing you will notice is that there are two different families of Nehalem-EX processors. The Xeon 7500s are aimed at general-purpose workloads and offer the most socket expandability. All of these chips can be used in two-socket or four-socket boxes, and some of them can be used in eight-socket or larger machines, too. The Xeon 6500s are cut-down versions of the chips that only work in two-socket boxes and that are specially tuned for the HPC market. These chips, explained Skaugen, were optimized to have the highest bytes per floating point operation ratio while minimizing the amount of node-to-node communication among the processors in the complex.

The top-end X7560 part has eight-cores spinning at 2.26GHz, has 24MB of L3 cache on the chip, and is rated at 130 watts using Intel's thermal design point (TDP) scale. The chip supports Turbo Boost, which allows for a core to have it cycle time jacked up if other cores are shut down when they're not being used, and it also supports Intel's HyperThreading simultaneous multithreading, which virtualizes the physical pipeline in the chip so it looks like two virtual pipelines to a system's operating system and its applications. In best-case scenarios, HT can boost performance of applications by around 30 per cent. In 1,000-unit trays, the per-chip price for the X7560 is a whopping $3,692. That is exactly what Intel charged for a dual-core Montvale Itanium 2 with 24MB of L3 cache.

The X7550 drops the clocks down to 2GHz, chops the L3 cache down to 18MB, and the price comes down to $2,729, which is exactly what Intel was charging for its top-bin six-core Dunnington Xeon X7460 processor running at 2.66GHz with 16MB of L3 cache. The next part down, the X7542, jacks the clocks up to 2.66GHz, drops the cache down to 18MB, cuts out HyperThreading, and reduces the core count down to six from eight; the price drops down to $1,980.

For that same $1,980 you can get a standard 105 watt part, the E7540, running at 2GHz with six cores and that same 18MB cache. If you are willing to take lower clock speeds, you can get even cheaper standard parts, the E7530 and E7520, which cost $1,391 and $856, respectively. Intel has also cooked up two low-voltage parts, the L7555 and L7545, running at 1.86GHz and rated at 95 watts, which have eight and six cores, respectively. These are reasonably pricey chips that will no doubt be used inside Nehalem-EX blade servers where a premium is expected in exchange for extra density.

Generally speaking, the Xeon 6500 processors are cheaper than their Xeon 7500 counterparts because they have some features and functions turned off, as El Reg predicted they would last fall. This is in keeping with the general philosophy that HPC shops are super-stingy and will not pay one extra penny for a feature they don't want and will never use.

The Nehalem-EX processors are implemented in 45 nanometer processes and have 2.3 billion transistors. ®

Agentless Backup is Not a Myth

Anonymous Coward

Ah, the mainframe guy still trying to justify his existence!

It is the same as the TCP/IP principle - you know that the IP layer will have errors, so you compensate for them in the TCP layer. You know a 'pc' will not be resilient, so you cluster/grid/RAC/virtual farm 2 or more of them to compensate.

So you get the similar performance for an order of magnitude reduction in cost.

1
0

sigh again again

Again your post is again full of various strange statements so it is kind of hard to know where to begin.

First on the spec_int part.

Well I'd rather wait with the SPEC_INT comparisons, Intel is reporting base numbers only. Which makes it kind of an invalid comparison, cause the specint numbers for Nehalem-EX could be significant better. So we'll just wait and see those ones.

As for

http://www.intel.com/performance/server/xeon_mp/summary.htm?iid=perf_server+mp_sum

Now what caught my eye is these benchmarks:

SPEC_JBB2005

Fujitsu PRIMEQUEST* 1800E 64 cores 8 sockets -> 3,321,826 BOBS/

IBM System x* 3850 X5 32 cores 4 sockets -> 2,012,730 BOBS.

This you have to compare to

IBM Power 750 Express 32 cores 4 sockets -> 2,478,929 BOBS

Which would mean that a 3.0 GHz power 750 would be faster, than the x3850. Now as for price then a half filled x3850 costs half that of a half filled power 750.

And SAP 2 Tier.

Fujitsu PRIMEQUEST* 1800E 64 cores 8 sockets -> 16.000 users.

IBM System x* 3850 X5 32 cores 4 sockets -> 10.450 users

This you have to compare to

IBM Power 750 Express 32 cores 4 sockets -> 15.600 users

So to match POWER7 you need a big boy 12U PrimeQuest 1800E, Now I don't have the price of that box, but other PrimeQuest boxes I have seen are in the same price range as a power 780.

But AC's comments on RAS features on Nehalem EX are relevant. These are the same reasons why people buy SD/MX000 and POWER 7XX boxes.

Again you get what you pay for.

Just admit it, lowest clocked POWER7 is as fast as the highest clocked Nehalem-EX and when the number of sockets grow.. then POWER7 is a sure winner.

And I don't know where you get your 256 socket x86 Solaris boxes from.. Well.. Sure SGI makes some but it's not like they support Solaris on that machine. Sure compile it yourself.. if Oracle hasn't cut support for something you need. As they are doing on other platforms right now.

And MC based boxes.. lets see some numbers rather than your mad ramblings.

// Jesper

1
0

tough choice

$3,157 for an 8-core 1.86Ghz 95W chip from Intel for 4 socket systems

$1,165 for a 12-core 2.2Ghz 80W chip from AMD for 4 socket systems

tough choice...

2
1

More from The Register

SCO vs. IBM battle resumes over ownership of Unix
Zombie lawsuit back and wants to suck the brains out of Linux
 breaking news
You don't need phone lines or cable for ANYTHING, says Dish
The satellite-dish man can sort you out with phone and broadband over the air too
 breaking news
What's HP got under wraps? Looks awfully flash and tape shaped
What happens in Vegas won't stay there - we've got the details
AMD lifts the veil on Opteron, ARM chip plans for 2014
Hey Samsung, let's see your chippery handle 64GB of RAM
Microsoft borks botnet takedown in Citadel snafu
Stupid Redmond kicked over our honeypots, wail white hats
IBM's $1bn layoffs latest: Now axe swings in US, Canada - reports
Union claims 121 storage bods canned after dismal sales
NetApp musters muscular cluster bluster for ONTAP busters
Storage array OS overhauled to juggle more nodes, go down on you, er, less
HP adds 'Haswell' Xeon E3s to entry ProLiant servers
Gussies up MicroServer for SMBs, adds baby switches