Well that was quick … I’m sure there’s some legal issues behind this move, but then again Intel can afford to foot the lawyer’s bill 😉
Intel hired Raja Koduri, who resigned as head of AMD’s Radeon Technologies Group (RTG), earlier this week. Koduri has been made Senior Vice President and Chief Architect of Intel’s future discrete GPUs. That’s right, Intel has renewed its dreams to power high-end graphics cards that compete with AMD and NVIDIA. Intel’s last attempt at a discrete GPU was “Larrabee,” which evolved into a super-scalar multi-core processor for HPC applications under the Xeon Phi line.
This development heralds two major theories. One, that Intel’s collaboration with AMD RTG on graphics IP could only go further from here, and what is a multi-chip module of Intel and AMD IP now, could in the future become a true heterogeneous die of Intel’s and AMD’s IP. Two, that the consolidation of AMD’s graphics assets and IP into a monolithic entity as RTG, could make it easier to sell it lock, stock, and barrel, possibly to Intel.
Intel places great faith in Koduri’s ability to either develop a major product line from scratch, or to effect tectonic shifts in the industry, such as Apple’s transition to the x86 architecture for its Mac product line.
Intel could have one of three approaches to build a discrete GPU from scratch. The first and most obvious one would be to scale up its current gen 9.5 architecture. The trouble is, that Intel’s SIMD parallelism is more transistor-heavy than even NVIDIA. It takes roughly 400-500 million transistors (coarse estimate) on the “Kaby Lake” die to build a GPU with 24 execution units. With 10 billion transistors, we’re looking at around 480 execution units, if not more.
More at TPU
