Intel Details Nehalem CPU and Larrabee GPU

Vigile writes "Intel previewed the information set to be released at IDF next month including details on a wide array of technology for server, workstation, desktop and graphics chips. The upcoming Tukwila chip will replace the current Itanium lineup with about twice the performance at a cost of 2 billion transistors and Dunnington is a hexa-core processor using existing Core 2 architecture. Details of Nehalem, Intel's next desktop CPU core that includes an integrated memory controller, show a return of HyperThreading-like SMT, a new SSE 4.2 extension and modular design that features optional integrated graphics on the CPU as well. Could Intel beat AMD in its own "Fusion" plans? Finally, Larrabee, the GPU technology Intel is building, was verified to support OpenGL and DirectX upon release and Intel provided information on a new extension called Advanced Vector Extension (AVX) for SSE that would improve graphics performance on the many-core architecture."

Comments

Re:Nehalem? Larrabee?

[TechyImmigrant]

They are code names, not product names.

Intel has a rich collection of silly code names.

Re:Gflargen and Blackeblae

[iknownuttin]

Haven't they heard of numbers?

You can't trademark numbers. When AMD started releasing "x86" numbered processors, Intel filed suit and lost. The judge stated that you can't trademark numbers. It's such an old case, this is what I found in the last 10 minutes regarding Intel and trademarking numbers.

I'm tired and too lazy to find the actual lawsuit.

More Integrated Garbage?

[immcintosh]

So, this Larrabee, will it be another example of integrated graphics that "supports" all the standards while being too slow to be useful in any practical situation, even basic desktop acceleration (Composite / Aero)? If so, I've gotta wonder why they even bother rather than saving some cash and just making a solid 2D accelerator that would be for all intents and purposes functionally identical.

Re:More Integrated Garbage?

[Kamokazi]

No, far, far, from integrated garbage. Larrabee will actually have uses as a supercomputer CPU:

"It was clear from Gelsinger's public statements at IDF and from Intel's prior closed-door presentations that the company intends to see the Larrabee architecture find uses in the supercomputing market, but it wasn't so clear that this new many-core architecture would ever see the light of day as an enthusiast GPU. This lack of clarity prompted me to speculate that Larrabee might never yield a GPU product, and others went so far as to report "Larrabee is GPGPU-only" as fact.

Subsequent to my IDF coverage, however, I was contacted by a few people who have more intimate knowledge of the project than I. These folks assured me that Intel definitely intends to release a straight-up enthusiast GPU part based on the Larrabee architecture. So while Intel won't publicly talk about any actual products that will arise from the project, it's clear that a GPU aimed at real-time 3D rendering for games will be among the first public fruits of Larrabee, with non-graphics products following later.

As for what type of GPU Larrabee will be, it's probably going to have important similarities to we're seeing out of NVIDIA with the G80. Contrary to what's implied in this Inquirer article, GPU-accelerated raster graphics are here to stay for the foreseeable future, and they won't be replaced by real-time ray-tracing engines. Actually, it's worthwhile to take a moment to look at this issue in more detail."

Shamelessly ripped from:

http://arstechnica.com/articles/paedia/hardware/clearing-up-the-confusion-over-intels-larrabee.ars/2

Re:Nehalem? Larrabee?

[Have Blue]

Most of Intel's codenames are names of real places.

Ummmmm, no

[Sycraft-fu]

First off, new integrated Intel chipsets do just find for desktop acceleration. One of our professors got a laptop with an X3000 chip and it does quite well in Vista. All the eye candy works and is plenty snappy.

However, this will be much faster since it fixes a major problem with integrated graphics: Shared RAM. All integrated Intel chipsets nab system RAM to work. Makes sense, this keeps costs down and that is the whole idea behind them. The problem is it is slow. System RAM is much slower than video RAM. As an example, high end systems might have a theoretical max RAM bandwidth of 10GB/sec if they have the latest DDR3. In reality, it is going to be more along the lines of 5GB/sec in systems that have integrated graphics. A high end graphics card can have 10 TIMES that. The 8800 Ultra has a theoretical bandwidth over 100GB/sec.

Well, in addition to the RAM not being as fast, the GPU has to fight with the CPU for access to it. All in all, it means that RAM access is just not fast for the GPU. That is a major limiting factor in modern graphics. Pushing all those pixels with multiple passes of textures takes some serious memory bandwidth. No problem for a discrete card, of course, it'll have it's own RAM just like any other.

In addition to that, it looks like they are putting some real beefy processing power on this thing.

As such I expect this will perform quite well. Will it do as good as the offerings from nVidia or ATi? Who knows? But this clearly isn't just an integrated chip on a board.

Re:Nehalem? Larrabee?

[TheRaven64]

But before they laughed, I remember a bunch of companies folded up their project tents (sun, mips, I think you are mistaken. MIPS still exists, but SGI stopped using it. HP killed both PA RISC and Alpha, but they co-developed Itanium, so it isn't entirely surprising. Sun kept developing chips, and currently hold the performance-per-watt crown for a lot of common web-server tasks.

Re:Gflargen and Blackeblae

[tzot]

Pentium, Pentium II, Pentium III and Pentium 4 made it clear which one was newer (although the shift to arabic numerals was a little inconsistent).

Someone sent an email to the Intel board of directors, allegedly from CIA, beginning with "Dear Sirs: it has come to our attention that you label your products with arabic numerals."

It took them a while to get that it was a joke.

Re:Why the brick wall?

[djohnsto]

Because power generally increases at a rate of frequency^3 (that's cubed). Adding more cores generally increases power linearly.

For example. Let's start with a single-core Core 2 @ 2GHz. Let's say it uses 10 W (not sure what the actual number is).

Running it at twice the frequency results in a (2^3) = 8X power increase. So, we can either have a single-core 4 GHz Core 2 at 80W, or we can have a quad-core 2GHz Core 2 at 40W. Which one makes more sense?

Re:Why the brick wall?

[TheSync]

1) We've hit the "Power Wall", power is expensive, but transistors are "free". That is, we can put more transistors on a chip than we have the power to turn on.

2) We also have hit the "Memory Wall", modern microprocessors can take 200 clocks to access DRAM, but even floating-point multiplies may take only four clock cycles.

3) Because of this, processor performance gain has slowed dramatically. In 2006, performance is a factor of three below the traditional doubling every 18 months that occurred between 1986 and 2002.

To understand where we are, and why the only way to go now is parallelism versus clock speed increase, see The Landscape of Parallel Computing ReseView from Berkeley.

Re:The Giant is awakened

[Hal_Porter]

I think AMD will do OK. Once Dell and the like get used to using CPUs from multiple sources they will probably survive. And a small company like AMD probably has an edge in terms of shorter design cycles and the ability to pick niches. AMD64 was a brilliant hack in retrospect that gave people most of the features of Itanium they wanted (64 bit, more registers) and none that they didn't (and expensive single source CPU with crap integer performance). Meanwhile Intel got hopeless bogged down trying to sell people Itaniums that they didn't want.

AMD and they have other clever stuff in the pipeline. E.g.

http://www.tech.co.uk/computing/upgrades-and-peripherals/motherboards-and-processors/news/amd-plots-16-core-super-cpu-for-2009?articleid=1754617439

What's more, with that longer instruction pipeline in mind, it will be interesting to see how Bulldozer pulls off improved single-threaded performance. Rumours are currently circulating that Bulldozer may be capable of thread-fusing or using multiple cores to compute a single thread. Thread fusing is one of the holy grails of PC processing. If Bulldozer is indeed capable of such a feat, the future could be very bright indeed for AMD.