For the longest time, Intel’s relationship with graphics has been troublesome. Intel’s first foray into discrete graphics, the i740, was a near total failure. Subsequently, the i740 was integrated into the chipset. At the time, chipsets were pad limited and a modest amount of die area was ‘free’. Actual graphics performance was perceived as an afterthought, and the unused area (so called white space) in the chipset was available for the GPU, but rarely a millimeter more. The imperative was to provide a free graphics solution, without adding to the cost of the platform. The investment in software and drivers was similarly limited, leading to products that were good for multi-media but wholly inappropriate for 3D graphics and games.
Over time, as graphics became more programmable and closer to actual computing, Intel’s interest grew. Simultaneously, Moore’s Law suggested that eventually GPUs would be integrated with the processor itself. This culminated with a bold announcement in 2007 that graphics would be a first class citizen in the Intel architecture. No longer would graphics languish at Intel, but the company would aggressively invest resources in software and hardware development and begin to allocate die area to create more powerful GPUs. The culmination of this effort in 2011 was the Sandy Bridge graphics architecture (also known as Gen 6), a DirectX 10 GPU that was integrated into the same 32nm silicon as Intel’s CPUs. The Gen 6 GPU occupied nearly 40mm2, more than a single CPU core and a far cry from the earlier white space only GPUs.
The graphics performance for Sandy Bridge was quite good in a relative sense, easily exceeding the stated goal of a 10× improvement, and reaching 25× for some workloads. In many cases, the Sandy Bridge GPU achieved better performance than low-end discrete graphics cards, while consuming less power to boot. To justify the added expense of a well-designed GPU, two versions were productized; a higher-end GT2 version and a low-cost GT1 variant. The idea was that the GT1 would continue to be available for ‘free’ and provide baseline functionality, while the GT2 would cost a little extra for better performance. This set the stage for Intel to increase performance for premium products, while reducing cost for others.
Intel’s ‘tick tock’ strategy avoids changing both the CPU architecture and the process technology at the same time to reduce risk and complexity. Ivy Bridge is the successor to Sandy Bridge, manufactured on Intel’s 22nm FinFET process. The main focus for Ivy Bridge was on migrating to the new manufacturing technology, rather than updating the CPU architecture. However, GPUs are much easier to design and validate than CPUs. APIs like DirectX, OpenGL and OpenCL are significantly simpler for backwards compatibility than x86. Moreover, the hardware is managed by a graphics driver which essentially includes a just-in-time compiler, so fixing bugs or taking advantage of new features is much easier.
While Sandy Bridge’s Gen 6 graphics were solid, one catch was programmability. The graphics industry was moving towards standards such as OpenCL and DirectX 11 that expose GPUs for general purpose computation, rather than 3D rendering only. However, the design for Gen 6 was finalized before the standards were available, in part due to Intel’s particularly rigorous validation process and lengthy design cycles. Additionally, AMD’s first mainstream integrated GPU in Llano was substantially higher performance than the Gen 6 graphics in Sandy Bridge, while providing support for modern APIs. Overall Sandy Bridge was a tremendous step forward for Intel, but 6 months after launch it was overshadowed by the competition.
Intel’s graphics team decided that it was justified to change the GPU architecture for Ivy Bridge, while simultaneously transitioning to a new process technology. The Gen 6 GPU in Sandy Bridge was already tightly integrated and shared the last level cache (LLC) with the CPU cores. The new Ivy Bridge graphics architecture inherits this system infrastructure and significantly improves programmability while advancing performance. The performance target for the Gen 7 graphics was roughly a factor of 2×, compared to the previous generation. Since this is a throughput-focused GPU, the performance largely comes from more parallel execution, rather than any increases in frequency. Ivy Bridge is Intel’s first truly programmable GPU and should be the highest performing integrated graphics at launch.
Discuss (35 comments)