On the first of this month, Intel told the computing world about Larrabee, its upcoming GPU architecture. This reporter got in on the last conference call of the day with the assistance of Nick Knupffer at Intel Global Communications after multiple cockups by Hill & Knowlton, who were abviously vying for the title of the world's most inept PR agency. Here is what we learned, and what they would not say.

Rumors about Larrabee have been floating around for at least two years. Intel's team sitting around their conference microphone said they had been working together on Larrabee for more like four years. Intel has hired a highly qualified group to pull this one together. We are all awaiting their final results.

So what in the heck is a “Larrabee”? Right now it is an architectural design document for a powerful graphical chip with a set of enhancements to existing APIs. All is promised to be explained in a speech that will be given during Siggraph.

We do know officially that the first products will target discrete graphics applications, support DirectX and OpenGL, and run existing games and programs. In addition Larrabee will run scientific and engineering applications, plus medical applications that use high-level graphics with ray-tracing for rasterziation.

Intel is sending mixed signals on who the buyers of Larrabee's architecture will be. On the one hand they are touting this as an integration from some X86 CPU functions into a more powerful GPU. That sounds like something that would go onto motherboards being used in laptops and low- to mid-priced desktop computers. There are rumours that one of the first Intel products to see Larrabee will be high-end graphical servers for large gaming installations.

We heard rumours that the Intel CPU used for modeling their X86 functions in Larrabee was a very specialized Pentium processor which was modified by US Department of Defense. Then, Pat Gelsinger, an Intel spokesperson and close personal friend of our publisher, Mike Magee, said in a July interview that Larrabee was based on Intel's P54C (i.e. pre-MMX) architecture. Pat was further quoted as saying that in the future, Larrabee would replace the CPU. Later, that claim was toned down a bit by another Intel representative.

Larrabee's architecture pipeline is a short execution pipeline with a fully coherent cache structure. The architecture has a wide, vector processing unit (VPU), multi-threading, 64-bit extensions, and sophisticated pre-fetching.

Larrabee takes advantage of X86 task scheduling which is performed entirely with software, rather than with fixed function logic. Therefore, rendering pipelines and other complex software systems can adjust their resource scheduling based on each workload’s unique computing demand.

Intel's engineers said their native programming model supports a variety of highly parallel applications including irregular data structures. This enables future development of graphics APIs, rapid innovation of new graphics algorithms, and true general purpose computation on the graphics processor with established PC software development tools.

The Larrabee architecture supports four execution threads per core with separate register sets per thread. It allows efficient in-order pipeline, but can handle more complex out-of-order pipelines when running highly parallel applications.

The Larrabee architecture uses 512 bits in each direction of ring-networking so agents can communicate in low latency manner resulting in faster communication between cores. There is 256K of shared L1 data cache. Intel is optimistic about Larrabee scaling from a few X86-based cores to 16, 32 or even higher.

Okay, we have given you the official version of the Larrabee.

Now for our ideas on how all this will work out for you computing folks. To begin with today's Larrabee is only an architectural rendering (a.k.a. Slideware), not a shipping graphics board. Every time anyone directed a question to engineering about “what does this do, or how does that really work?” the corporate answer was: “that is a product question and today we are talking about the new architecture” or some such simplistic put-off-to-the-future response.

Intel's first generation of Larrabee gaming developer boards are unofficially on schedule for fall-to-winter delivery this year. That means if you are a real game developer you already have your name in the hat to get one, don't you? If not, what are you waiting for? Another question: How long are you willing to wait for delivery?

We have been watching PC computers since the days of CP/M and Z-80 processors. We have yet to see a product smoothly scale from initial slideware presentations to a shipping product.

As every code developer knows, CPUs are designed to get maximum performance from a stream of instructions, which operates on diverse data (such as integers and floating-point calculations), perform random memory accesses, and branching. CPU architects are always working to extract more parallelism of instructions; that is, to launch as many instructions as possible in parallel.

A GPU is much simpler. Its job consists of taking a group of polygons and generating a group of pixels to display on-screen. The polygons and pixels are independent of each other and can be processed by parallel units. That means that a GPU can afford to devote a large part of its die surface area to calculating units which, unlike those of a CPU, will actually be used.

This brings Intel to their convergence of the CPU and GPU in Larrabee.

As shown above, Larrabee's architecture is completely software controlled. It can take advantage of user programmed enhancements to the graphics engine and middleware. Thus, gaming developers can use Larrabee's multi-core x86 design which is particularly suited for physics and real-time ray tracing. The downside could very well be slower operations with traditional rasterized 3D graphics tool kits like DirectX and OpenGL.

However, last fall Intel picked up gaming development tools maker Havok, the authors of the famous Havok physic engine used in top-of-the-line games like BioShock, Oblivion, Half-Life 2, and Halo 2. Havok's tools should keep Larrabee at least up with today's offerings from the other two competitors; Nvidia and AMD/ATi.

The downside for gaming developer's new paradigm of total software control of all the graphical functions is a shortage of in-house programming talent. Intel will need to educate gaming developers who will have to figure out how to write entirely new code for software control of hardware. There are probably not many coders today who are journeymen at writing code for graphic engine internals or modifying middleware that controls multiple-core parallel processes in a combination GPU/CPU. Such lack of experience could lead to slow acceptance of something like Larrabee, which is not past the Alpha stage of hardware development.

This prediction is based on Nvidia's CUDA (Complete Unified Device Architecture) track record. It has been out in the wild for a year and half with a claimed installed base of 60-to-70 million computers. Many game developers, hesitant to implement architecture changes, have been slow to embrace it. A chief complaint about CUDA is its lack of portability across environments.

A major difficulty for Intel to overcome is their past and current reputation for poorly crafted drivers. Their track record for graphics products is infamous for being in the swamp-full-of-alligators category for results.

The Intel engineers said Larrabee's architecture fully supports IEEE standards for single and double precision floating-point arithmetic. Support for these standards is a pre-requisite for many types of tasks including financial applications and medical high-level graphics. Especially problematic are real-time medical graphics transmitted from afar using the public Internet.

Another stumbling block Larrabee must overcome is that Microsoft has made it clear that DirectX 11 doesn't do ray trace rendering. If some gaming developer, or Intel, adds ray tracing to OpenGL, that might create another avenue for Larrabee's enhanced architecture.

There is also the question of how many watts of power will Larrabee use? One rumor is it will be nearly 300 watt TDP. A big jump up from the present 200 watt TDP plateau. Larrabee's speed was questioned because we had heard rumors of Intel targeting 2 Teraflops. Both questions got Intel's standard corporate answer of effectively: no comment at this time because it isn't a shipping product.

Over in the Apple rumor store, there are a lot of comments about Apple discontinuing their relationship with Intel. That does not compute when the major rumor about the next version of Mac OS X is its enhanced ability to handle multi-core processors, including multi-core GPUs. This was another question during the conference call that Intel's engineers deferred to another day saying: we are not talking about a specific product, we are discussing an architecture.

If Intel has chosen GDDR5 memory for their supposed 2 Teraflop speed demon, will that be enough speed? That is a really good question, because AMD/ATi is already at 1.2 Teraflops with their Radeon HD 4870. Claims are the Radeon HD 4870 X2 will break 2 Teraflops. What speed will really be needed come a summer 2009 shipping date for Larrabee boards? Again Intel's engineering team gave their standard ... Okay you can fill in the blanks on this one by now.

Most of you readers know there is a war of words going on in the graphics board manufacturing arena. This weekend it was reported that Nvidia's latest architectural offering has no takers among Taiwan's motherboard manufacturers, and Nvidia was leaving the chipset business. That rumor was immediately denied by Nvidia.

We wish Intel the best of luck with Larrabee. We would like to get a board to test as quickly as they are available.

However, if today you want a high-end graphic board for your computer, at this moment, and in the foreseeable future, we recommend the AMD/ATi Radeon HD 4870 series products. They are at your local computer stores, right now.

How long are you willing to wait for delivery of Larrabee boards and how much extra work are you willing to do to use them? X