BRIGHT SIDE OF NEWS About | Advertise | Contact BSN USER Login
| Register
SUBSCRIBE Newsletter | RSS Feeds
Thursday, March 11, 2010
Email this to a friend.
Your friend's e-mail:
Your Name:
Your e-mail:
Message subject:

IDF Spring 2009: Intel Larrabee to feature 1.7 billion transistors?



In the wake of a recession, Intel turned IDF Spring into a small, local focused event, but still at least one Western-world news office got in - and deserve "pic of the week" award with no questions asked.

Pat Gelsinger showed up the wafer, showcasing LRB [Larrabee, all of major Intel projects have a three-letter short, similar to airport naming] chip for the first time in general public. If we compare the picture to some previous wafer comparisons and Mr. Gelsinger's fingers, we would put the die slightly above 600mm2. PC Perspective agreed with the estimate and went as far to count that there are 64 dies on a single wafer. 64 dies on a 300mm2 45nm wafer… can't be cheap to make LRB even with 100% yield, that's for sure.

Intel's SVP does his trademark move: shows the Intel-built wafer to the world. This time, one with Larrabee dies.According to our sources, first Larrabee chips will be built using 45nm High-K Hafnium-loving process with a move to 32nm a year after. Now, if Bloomfield die [Core i7] is 263mm2 in size and is consisted out of 731 million transistors, that means that Larrabee's 600mm2 die would end somewhere in the 1.65-1.75B transistor range.

If you're not impressed at the specs of this chip, with its 12 P55C-based cores + Vector units, think again. ATI is testing 40nm manufacturing with RV740, a $99 chip featuring 826 million transistors. Intel's "test" has more than double that. However, the entrance into the world of cGPU or GPGPU won't be that easy. From the hardware side, ATI is currently shipping 959M parts [4890] and 1.92B with their dual-GPU cards [4850X2/4870X2], nVidia ships 1.4B part for almost a year now [GTX260/275/280/285], or 2.8 billion transistor part if you count their dual-GPU part - GeForce GTX295.

Recently, we learned the targeted die-size for GT300 and saw that nVidia isn't changing itself, and you can expect that the next-gen DX11 part will fit somewhere between Larrabee and GTX295. Yes, with just one die. Performance wise, look for strong accent on GPGPU applications and not-so-much accent on games, with LRB performing in the range of mainstream cards…

We cannot estimate the current cost of the Larrabee-based graphics card, since LRB chip features 1024-bit internal and 512-bit external crossbar bus with 1Tbps internal bandwidth. Externally, if Larrabee connects to 1024-2048 MB of GDDR5 memory at 1.0 GHz clock [4.0 GT/s], we would see a card with 256 GB/s, gaining the title of the world's highest bandwidth graphics product.

The ssage of a 512-bit interface will result in slightly more complex PCB [much less than GDDR3-powered GTX200 series], so that will not radically increase the complexity and the price. Thus, we estimate that LRB probably cannot sell for anything less than $350-450 given the cost of BoM [Bill of Materials], even in 2010. That might change, though - if Qimonda really stops being around, prices of GDDR5 memory might skyrocket.

However, you should never underestimate Intel and its marketing and sales muscle. The company is famous for its "buy a CPU, get a motherboard" tactic on System Integrators, so think twice if you think AMD or nVidia can walk all over Larrabee. We end with a message to Jen-Hsun, nVidia's CEO: Larrabee is no longer a PowerPoint presentation, what do you plan to do now… open a can of whoop-ass?



© 2009 - 2010 Bright Side Of News*, All rights reserved.



Related articles:

Tags:

Share and enjoy :)

  • Digg
  • del.icio.us
  • StumbleUpon
  • TwitThis
  • Reddit
  • Furl
  • Google
  • Technorati
  • Sphinn
  • Mixx
  • Facebook
  • LinkedIn
  • Slashdot
  • Newsvine
  • Ma.gnolia
  • BlinkList
  • connotea
  • Fark
  • MisterWong
  • Netvouz
  • PlugIM
  • Propeller
  • Simpy
  • SphereIt
  • Spurl
  • ThisNext
  • YahooMyWeb
  • co.mments
  • Live
  • MySpace
  • Yahoo! Buzz


Comments:

Place your bets? by: General Lee D. Mented on 4/10/2009
Once graphics transitioned from special purpose fixed-function devices to "GPU" cores that are Turing-Complete, this became a war of fab process. One Intel is in a very strong position to fight.

For a long time it's been pretty obvious that the GPU isheading for collision with the CPU. If your graphics chip becomes complex enough, it can do your general purpose calculations. Branches being added to shaders and integer support were some of the last real bottlenecks to this. From the other side, if you can't ramp clock speed on CPUs and have to go massively parallel, graphics is one of the few drivers of performance demand that can scale with more cores gracefully. AMD and ATI figured this out from separate ends of the problem and joined forces. Nvidia seems ignorant of it. Intel seems to have just been in a "we'll cross that bridge when we come to it" position, and lo and behold, a bridge has appeared!

I honestly thought Nvidia would buy VIA's x86 component. But apparently not. I don't know how they're going to navigate the patent minefield that is x86 hardware. And as much as people are shouting the second coming of ARM I honestly don't see it gaining enough mass for x86 to fall over. Nothing has been able to dethrone it, despite numerous architecture issues. The only real way is through virtualization, which is counterintuitive to the goals of the ARM camp.

But in reality, virtualization is the key to the future of using lower power devices. It is the only solution to the "nobody wants to type on a cellphone" problem and the bridge to wearable computing. Until we see WRAITH (Widely Redundant Application Instanced Thread Handling) added to the underlying OS stack people are going to be continued to be vexed by the problem of cellphones and netbooks bloating with features that eat power and cost more, file collections that get out of hand between multiple devices and are a pain to sync, and a desktop market that has no choice but to try reinvent itself as the "home server" but can't then justify expensive graphics, sound, LCDs, cases with pretty lights, or the cost over just an embedded platform with a lot of SATA connectors in a box.

The solution is to virtualize the system such that the base platform is on the device the user carries most (cellphone), and dynamically incorporates more resources on the fly when available. This can mean things like running an app on a server and streaming its window via RDMA, or in reverse using the moble device as kind of like a login key that the desktop consults when it boots up so it knows what settings to load. Features like per-application state suspend and hibernate will be critical to handle issues like if my OS kernel is on my phone and I'm using my desktop to play a game on its local resources, if I suddenly get up and walk out of network range what happens to the game thread? What happens to my desktop session? Ideally we want the desktop to transition to the phone but the game to enter a hibernate state to disk so that I can reconnect to it later where I was, but someone else can use the desktop in the meantime. This is going to be a big challenge to do and requires redundant threading, presence detect, and hardware virtualization features on low-resource devices.

But it's a key step. We can't move to the next stage where my storage is in my pocket, the processor is on my arm, and the software-configurable multimode radio is head-mounted without dealing with the issue of "How do we handle the addition and subtraction of computing resources on the fly?" It has to be transparent and efficient.

I think Intel has figured this out and Larabee is more just another signpost on the road for them. What they want is x86 in everything, and that means it has to scale down (Atom) and go parallel (Larabee). The stuff they're doing with vPro and talked about at IDF where the chipset just presents the presence of a GPU like an instruction set extension (with no drivers or OS dependence) are hints to the puzzle.
Leave a comment:

Author:

Title:

Comment:


Enter the code shown above:

(Note: If you cannot read the numbers in the above
image, reload the page to generate a new one.)




Highlight
  • Super Micro comes out with a 12-Core ready Motherboard
  • Fashion: The land of cotton is going nanotech
  • Intel's Six Core, 12-Thread, 32nm Core i7 980X; Obsecenly fast in threaded Apps.
  • PC Power and Cooling: A Practice in Perfection
  • PC Power and Cooling: A Practice in Perfection
New servers, new features

Welcome to BSN*!

We have just completed the server switch and brought three new server to power Bright Side of News*. We will be introducing new features to the site over the course of next few days, so stay tuned.

For those wanting to know, all three servers are powered by Intel Xeon processors at 2.6 GHz each.

Best regards,

The BSN* team

© 2009 - 2010 Bright Side Of News*, All rights reserved.