Intel's cache gigantism results in 2bn transistor Tukwila What with being the keeper of Moore's Law and all that, Intel found its rightful place as the first revealer of a 2bn transistor chip this week at the sausage fest that is the International Solid State Circuits Conference in San Francisco. The chip goes by the name Tukwila, and it's a four-core version of Itanium due out in …
Seeing as everyone made such a fuss about Power6, when it only delivered a 10% real performance improvement over Power5+, I'm quite looking forward to Tukzilla as I expect about 50% real performance gain. But then my main problem is not the speed of the cores, it's getting the data to the cores to keep them all spinning, and as usual it's the SAN disk that is the biggest sluggard in the whole stack. Tukzilla has an even wider pipe to feed the cores so it gives me even more arguments for some nice SSDs. ;)
Given the law of diminishing returns, if they genuinely achieve a 50% performance increase at Matt Bryant mentions, it'll be pretty impressive!
If anyone can think of any setup that has ever gotten above a 40% increase by doubling the resources I'd be interested to know, as even to cut overheads to 10% would be pretty difficult
Two cores, same megahertz, twice the throughput, wouldn't that be what one would expect? And, since it's a smaller process, it is only 25% more power instead of 100% more power - so that's an improvement.
Of course, it would be nice if it was a redesign that did more work per cycle, but instead it saves energy.
SHUSH!!!!! There's no way the beancounters will sign off for the SSDs without me waving a nice "projected performance improvement" figure in front of their eyes! ;)
PS: If anyone from HP is reading this - the MSA2k is all very nice but when am I actually gonna get SSDs for EVA?
Folks,
I might be a little bit of a retard but then i think thats precisely what was done by Mr. Sun "Java" Microsystems when they pushed in two cores of a US III and came up with a USIV chip. Ah, not only that, the same was continues with US IV+. Come to think of it. Double the performance , with double the cores. How does it hit the Oracle Licensing. With every new Itanium that HP (or Intel?) comes up with, Larry Ellison gets a sposored vacation to Mars.. Wowie....
On the lighter side...
How many Itanium researchers does it take to change the light bulb?
None.. There aint no electricity. Larry sold them the Oracle licenses.
Long Live the Itanic.
Yes, you are a bit of a retard! Sorry, couldn't resist. AT teh time, HP had dual-core PA-RISC when SUN went dual-core, and HP had the mx2 module that allowed two Itanium CPUs to fit in one socket (not real dual-core, but it gave the same end result). So what was the difference? Here's a simple way to think of it - seven years ago I worked on V2500s with 32 CPUs in a box about half a rack in size (and that's with no internal disks!), which required dedicated wall connections to ring mains. Just a few weeks ago I worked on an rx7640 with eight CPU slots but dual-core CPUs, and I had a system that could out perfrom the old V2500 by a wide margin. Pretty soon I could take that rx7640 and upgrade it to run quad-core Tukzillas, then I will have my 32 cores (and those are real cores, not multi-thread vertical timesharing mini-RISC weenies) in a box roughly a quarter of a rack in size, powered off standard rack PDUs rather than dedicated rings, and with it's own internal disks and devices. The difference to Mr SUN? Well, the Itanium route has increased performance and has a path for future development with OS compatibility built in, all great news for consolidation, whilst the SPARC option is dying a slow and painful death with SUN pushing a disruptive and non-compatible shift to one of four chips (Xeon, Opteron, SPARC64 or Niagara), each with their own limitations, and none of which can give the performance or longterm future of the Itanium, and with a vague promise that they'll try and deliver Rock in the future.
Unless you run large distributed jobs with lots of communication (in which case you should consider getting a large SMP anyway) I suggest you make price/performance benchmarks of your applications before jumping on the new Itanium.
Perhaps for the same X amount of money you can get way more nodes with slower cpus, but all together they will provide your cluster with higher throughput than a smaller number of nodes with top-shelf IA-64 on board.
-marek
--
clusteradmin.net :: a blog about building and administering clusters
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
