Intel Larrabee

» see intel.com as it lived, on the wayback machine

~ the obit ~

Every graphics card of 2008 was built the same way. Fixed-function silicon handled the parts of rendering that always happen and programmable shader units handled the parts that vary. The driver's job was to translate what a game asked for into whatever that specific hardware could actually do. Larrabee proposed throwing the fixed-function half out. It was to be dozens of small in-order x86 cores, derived from the original Pentium, each with a 512-bit vector unit that chewed sixteen single precision floats at a time, with scatter and gather instructions and a mask register to make the vector unit tolerable to program. The rasterizer would be a program running on those cores rather than a circuit.

The argument for it was flexibility, and Intel's SIGGRAPH 2008 paper made the argument concretely by listing things that were awkward or impossible on the GPUs of the day: reading from a render target, order-independent transparency, irregular shadow mapping, real-time ray tracing. If the whole pipeline were software you could rewrite any stage of it, and a game could look different from every other game because it was not renting the same fixed pipeline. Michael Abrash and Tom Forsyth were among the people writing the 3D stack, at RAD Game Tools, which told the graphics world that this was a serious attempt rather than a research slide. The chip was named for Mount Larrabee, or possibly the state park, in Washington.

The numbers never showed up. At NVISION in 2008 an Nvidia employee dismissed the SIGGRAPH paper as marketing and quoted an analyst calling the architecture a GPU from 2006. In June 2009 Intel claimed prototypes were level with Nvidia's GeForce GTX 285, a card that had already been on sale for months. On 17 November 2009 Justin Rattner demonstrated an overclocked Larrabee topping one teraFLOPS at the Supercomputing conference, which he presented as a first for a single chip, and which was roughly one fifth of what a shipping graphics board could do. Seventeen days later Intel said the first generation would not be released as a consumer product.

I was finishing university and starting work while this was being argued about, and Larrabee is the first hardware I can remember having firm opinions about without ever having seen one. That is what it was, for two years: a paper, and an enormous quantity of forum argument over whether general-purpose cores could do graphics well enough. I wanted it to work, mostly because a graphics card you could program like a processor sounded like it would make the rest of my life easier. Nothing was ever for sale.

~ the rap sheet ~

~ the pitch ~

The SIGGRAPH paper included cycle-accurate simulations with the memory and texture limits modeled, and it published how many one gigahertz Larrabee cores would be needed to hold sixty frames per second at 1600 by 1200 in real games. Roughly twenty-five for Gears of War without antialiasing. Ten for Half-Life 2: Episode Two with four times antialiasing. Performance scaled almost linearly out to thirty-two cores and was still at ninety percent of linear at forty-eight, which was the strongest thing the paper had to say, because linear scaling means you can buy your way out of a deficit by adding cores.

At the Intel Developer Forum in September 2009 the public demonstration was a ray traced build of Enemy Territory: Quake Wars, running in real time with a water surface reflecting the ships and vehicles around it correctly. Ray tracing was the emotional argument rather than the commercial one. It was the thing a fully programmable renderer could do that the fixed pipeline could not, and it was also the thing nobody was shipping games with.

~ the cracks ~

The schedule was the whole problem. A 2007 report in PC Watch had the first Larrabee arriving as a 32-core part on a 45 nanometer process in late 2009, with a 48-core shrink to follow. The silicon slipped, and an architecture whose entire performance case rests on core count cannot absorb a delay, because the competition spends that time adding units to a design that already works. In-order cores and a software rasterizer both trade peak efficiency for flexibility, and that trade only pays if you have the transistors and the clocks to cover it. Intel's official reason for the December 2009 reset was delays in hardware and software development, which for once was probably the actual reason.

On 25 May 2010 the Technology@Intel blog confirmed it. Larrabee would not be a GPU. It would be a high performance computing product competing with Nvidia's Tesla line, which meant the graphics half of the project, the part that had made it interesting to everyone outside supercomputing, was over.

~ the long tail ~

The hardware became Xeon Phi. Knights Ferry was the prototype card and Knights Corner the 22 nanometer production part. It lived for years as a coprocessor in machines whose purchasers cared about floating point and not at all about triangles. The instruction set had the longer afterlife: the wide vector work done for Larrabee is the direct ancestor of AVX-512, which shipped in ordinary server and desktop processors and is still there.

Intel did eventually ship discrete graphics, as Arc in 2022, twelve years later and built along entirely conventional lines with fixed-function hardware where everyone else puts it. And the argument Larrabee lost has been quietly half conceded by the winners, because GPUs since then have become far more programmable and general than the ones that beat it. Larrabee was wrong about the timing and about how much you can throw away at once. It was not obviously wrong about the direction.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »