Larrabee and Future Products
X-bit labs: Is there internal deadline for Larrabee launch? It does not seem that the product will be released in 2009.
Nick Knupffer: We have said that the first Larrabee product will launch in the 2009/2010 timeframe.
X-bit labs: Have you taped out the Larrabee GPU already? Is there anything that your engineers, or third-party developers, have to play with already?
Nick Knupffer: We don’t normally give out those kinds of schedules. We are on track for our 2009/2010 launch timeframe.
X-bit labs: What levels of performance (in TFLOPs) should we expect from Larrabee as well as GPUs in the next 2-3 years?
Nick Knupffer: You will have to wait and see. But just like in the CPU space – TFLOPS are not an accurate representation of performance – you will want to look at actual application performance.
X-bit labs: Do you think that full programmability of Larrabee will be utilized by developers despite of the fact that there is no API for it? Perhaps, Intel plans to release its own API for Larrabee?
Nick Knupffer: For the pure C/C++ programmability of the Larrabee Native programming model that we describe in the paper, we believe that the compiler and the ISA is the fundamental building blocks needed to enable these developers. From that they can build any system to harness Larrabee. If by API, you mean release our own "Graphics API", then we remind you again that we have full support for DirectX and OpenGL APIs.
X-bit labs: Do you think that the lack of extreme programmability on ATI and Nvidia GPUs will slowdown adoption of Larrabee’s advanced capabilities by developers?
Nick Knupffer: We haven’t seen this level of developer excitement for a very long time – we will be able to share more closer to product launch.
X-bit labs: What are the reasons that led Intel to decision of creating a radically new GPU design when it comes to Larrabee?
Nick Knupffer: This is a great question and not one many people ask. In actual fact: Graphics has come back to IA, not the other way around.
Let me explain: If you look back to the Pentium processor era, developers ran all their 3D engines on the CPU. They had complete freedom to develop their games and 3D engines as they wished as the CPU was inherently a general purpose machine and as such completely programmable. They could use rasterization, voxels (remember Comanche?), rendered polygons, or whatever method they cared to invent.
Then the first consumer graphics card came out, the 3dfx Voodoo. The Voodoo basically implemented a subset of OpenGL into silicon. Any fixed function silicon device will (or should) always outperform general purpose hardware in both outright performance and power. The games programmers were both delighted at the speed and disappointed in being shackled into using a rigid graphics pipeline. Things developed, DirectX was created and the industry continued to deliver faster and faster rasterization hardware. However – with DirectX 8 and then 9, 10 and now 11 – more and more stages of the graphics pipeline have become more and more programmable.
Today’s graphics cards are still fixed pipeline devices but with some programmable stages. So in effect – the graphics pipeline itself has moved back towards the general purpose architecture of the CPU – and Larrabee is composed of many small general purpose IA cores.