backtop


Print 23 comment(s) - last by mlau.. on Feb 8 at 1:28 PM


Intel Tukwila Processor  (Source: BBC News)
Intel Tukwila promises twice the performance of current Itanium 9100 CPUs

Intel has announced that its next generation Tukwila processors have two billion transistors on the die. The first Tukwila processors will be in the Itanium family and the bulk of the transistors will be used for on-die memory.

An Intel senior fellow, Justin Rattner, told eWeek that the Itanium Tukwila processor will have 30MB of on-die cache and support up to eight instructional threads.

While Intel has been ramping up its 45nm processors with a retool of some of its fab plants to build new Penryn processors, the Tukwila processor uses the older 65nm process for its design. Intel says it expects the Tukwila processors to hit the market by the end of 2008 and the first processors will run at a 2GHz clock speed.

The new Tukwila processors will also be the first of Intel’s Itanium processors to use the new QuickPath interconnect technology, which is an integrated memory controller. The integrated memory controller has been used by Intel rival AMD on its x86 processors for years.

Rattner tells eWeek that the new Itanium Tukwila processor will offer twice the performance of the current 9100 series processor and the Tukwila will operate inside a 130-watt thermal envelope. The core architecture of the Tukwila processor is said to remain the same as the current generation of processors though Tukwila will offer more features.

Intel’s Tukwila processor will compete in the same market as Sun’s Rock processor which sports 16 processor cores.



Comments     Threshold


This article is over a month old, voting and posting comments is disabled

The Human Brain
By masher2 (blog) on 2/7/08, Rating: 0
RE: The Human Brain
By ZaethDekar on 2/7/2008 12:47:33 PM , Rating: 5
Then you add in the contacts that let you change your vision, shock your brain to remember more information, and then get some artificial limbs that look real.... and you have a borg!

Brings a new meaning to intel inside.


RE: The Human Brain
By amanojaku on 2/7/2008 12:49:35 PM , Rating: 2
And we'll still kick the PC's @$$ in FPS games.


RE: The Human Brain
By Etsp on 2/7/2008 12:55:27 PM , Rating: 5
Only in games where the PC doesn't cheat.


RE: The Human Brain
By surt on 2/7/2008 12:57:23 PM , Rating: 5
And last time I tried, it took only 100 million transistors operating at 1ghz to simulate one neuron in real time.

A neuron and a transistor are not at all computationally equivalent.


RE: The Human Brain
By masher2 (blog) on 2/7/2008 1:50:55 PM , Rating: 1
The researchers who performed a real-time mouse-scale cortical model simulated some 8 million neurons with Blue Gene/L...which works out to only about 6,000 transistors per neuron. Of course, if one were to design an ASIC to model a neuron, rather than general-purpose processors, one could cut that down by probably an order or magnitude or more.

Still, my point wasn't to suggest that in 18 years a CPU would be as powerful as a human brain, but just that by then we'll be just starting to reach its level of complexity.


RE: The Human Brain
By hlper on 2/7/2008 3:31:46 PM , Rating: 2
It's easy to get excited about the possibilities, but you have to also consider that a single neuron can have 10,000 synaptic connections across vastly differing parts of the brain (basal ganglia neurons are amazing in their diverse connections across completely different sensory modalities). There are also numerous neurotransmitter systems (more receptor types and new molecules to bind them are constantly being discovered) that can tell a single neuron to increase and decrease firing simultaneously, as well as increase or decrease the strength of connections through gene transcription.

Aside from neurons, there are also about a trillion glial cells to consider in the brain, which support critical neuronal function and science is only beginning to understand their role in cognition.

Who is to say we will ever be able to model such a complex, and plastic system. Certainly, anything we do in the next couple of decades is unlikely to come close to a human brain.

By the way, I am a Neuroscience graduate student because I think this stuff is truly awe inspiring. I am currently studying molecular contributors to development of schizophrenia.


RE: The Human Brain
By SeeManRun on 2/7/2008 1:45:59 PM , Rating: 5
quote:
The average brain has around 100B neurons. At the current rate of growth, CPUs will exceed that number in transistor count by around the year 2024.


But a transistor can only be 1 or 0, and a neuron can do so much more!
http://en.wikipedia.org/wiki/Neurons


RE: The Human Brain
By masher2 (blog) on 2/7/2008 4:00:39 PM , Rating: 1
Work has actually been done on tristate and more complex ternary logic circuits. Potentially we may one day see a computer using such multi-valued circuits in place of the boolean we use today.


Woo!
By ZaethDekar on 2/7/2008 12:27:21 PM , Rating: 1
It will be very interesting how this turns out. I've always wondered why they only went 2 mb cache instead of 10-20. Now 30?




RE: Woo!
By boogle on 2/7/2008 1:07:06 PM , Rating: 4
Speed / Performance. This CPU in a way has 8 cores, so at 4MB per core, that would be 32MB. So you could say this new CPU doesn't have that much more cache than upcoming quadcores.

Either way the amount of cache is decided by how much it would cost to implement it, offset by how much the cache would improve performance. With standard x86 applications, the amount of necessary cache is relatively small - the prime reason L1 cache hasn't really increased in size for a loonnnggg time. Functions are designed to fit inside the L1 cache, increasing it's size won't really improve performance that much. L2 cache only really needs to fit the size of the .exe (which is say, 2MB or so) and regularly called functions from said executable.

This is an Itanium so it's designed for a very specific kind of server. Applications will be designed specifically for this server, therefore the extra cache can and will be used.

More cache helps, but you end up with diminishing returns if you go too big too fast.


RE: Woo!
By imperator3733 on 2/7/2008 1:14:29 PM , Rating: 3
Tukwila is a quad core, not an octal core. It has something like HyperThreading which makes it look like it has 8 cores.

Also, L1 cache hasn't increased much because the larger you make the L1, the slower it gets. L1 is designed to be very fast and hold only the most recently/commonly used stuff.


Yesss...
By TimberJon on 2/7/2008 12:33:39 PM , Rating: 2
I'll take just 10 MB of on-die cache. Hopefully the 30MB is not just "Technically 30" or affected by other restrictions.

Processor speeds/models? Is there a list yet?
Need a good strong nexgen gfx design/CAD processor.




RE: Yesss...
By Cobra Commander on 2/7/2008 12:39:32 PM , Rating: 2
what CAD are you using on IA64 (Itanium) architecture?


RE: Yesss...
By Amiga500 on 2/7/2008 1:20:15 PM , Rating: 2
Good question... ???

Pity our clusters don't have these things inside!


what a waste
By goku on 2/7/08, Rating: 0
RE: what a waste
By murphyslabrat on 2/7/2008 3:20:47 PM , Rating: 4
if( you_knew_what_this_was ==TRUE)
Who the hell is going to overlcock an Itanium?!?!?
else
Sorry, don't read the above.


RE: what a waste
By mlau on 2/8/2008 1:28:15 PM , Rating: 2
This thing kicks ass in scientific calculations with large working
sets (think large, multidimensional matrices). The RAM interface is
still agonizingly slow and has huge latencies compared to a SRAM
clocked at core frequencies.

You know, there's more to computers than windows, directx and crysis...


How do you fit that into the box?
By murphyslabrat on 2/7/2008 1:40:36 PM , Rating: 2
SO, if my numbers are correct -2000*(143/291)-, there it should be a ~983 mm^2 die. This translates, roughly, into a 30mmx33mm die. That's nearly as large as the entire CPU package that us desktop users are used to.




By IntelUser2000 on 2/7/2008 9:56:25 PM , Rating: 2
Tukwila is on Intel's 65nm process and has a die size of 699m2. The 30MB L3 is arranged to the following hierarchy and size.

512KB Instruction L2
256KB Data L2
6MB dedicated L3 cache per core
2MB Directory cache
34GB memory bandwidth
96GB QPI bandwidth


Why 30mb of cache?
By Comdrpopnfresh on 2/7/2008 8:16:52 PM , Rating: 2
If it has an onboard memory controller, why does it need so much cache? I thought that was an advantage of onboard memory dies.




RE: Why 30mb of cache?
By masher2 (blog) on 2/7/2008 9:50:37 PM , Rating: 2
The EPIC architecture of Itanium benefits much more from a large cache than does Core (or any other desktop CPU for that matter).


How is it distributed?
By imperator3733 on 2/7/2008 1:09:48 PM , Rating: 2
How is the cache distributed? I heard somewhere that Tukwila has 6MB L3/core. That would leave 6MB non-L3 cache. If that were the case, I would expect that each core have 1MB L2i and 512KB L2d. Anyone know if this is right?




"This week I got an iPhone. This weekend I got four chargers so I can keep it charged everywhere I go and a land line so I can actually make phone calls." -- Facebook CEO Mark Zuckerberg

Related Articles













botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki