BACK TO THE FERMI
Nvidia’s latest GeForce 384 series drivers have done something rather unexpected, offering GeForce 400 and 500 series (Fermi based) graphics cards a new lease on life by giving them DX12 support for the first time. It is thought that this is more to provide DX12_0 level compliance simply so they’re fully supported after the Windows 10 Creators Update more than anything else so these parts continue working. So don’t go expecting your five year old GPU to get a free speed boost. Just be happy it’s still supported.
NVIDIA NAVI-GATES THE FUTURE
With large monolithic processors getting harder and harder to manufacture, companies are starting to look towards multi-chip modules (MCMs) to break through the manufacturing R&D wall stalling the progress of chip makers trying to keep up with Moore’s law levels of performance gains every year.
The problem with single large monolithic chips is that just a single defect can ruin the whole chip, and the more die aerial space used per chip the bigger the chance there is of them occurring, reducing yields.
The benefits of MCM become quickly apparent as smaller dies mean higher yields during production (equals cheaper products) and being able to slap several chips together allows you to exceed die space restrictions of current single chip manufacturing, meaning more transistors to play with for the processor as a whole, increasing performance.
AMD has already started down the MCM path to try and offset this manufacturing risk in the form of its Ryzen CPU’s which uses its newly minted Infinity Fabric to provide the interconnects and data signalling required across the physically separated chips on the CPU module.
AMD has long had on its roadmaps (after Vega) a GPU architecture codenamed Navi with the main detail about it being that it brings multi GPU scalability. Only now does this make sense. Essentially, it’ll be multiple smaller GPU chips glued together on the same package using AMDs Infinity Fabric connecting them all together.
Come to think of it, despite AMD’s Zen processors success and performance increase (over 40%! from last gen), I would say AMD’s Infinity Fabric is actually its greatest achievement from Ryzen’s development. It’ll be tying together CPU, APU and GPU MCM modules from AMD for many years to come. Without this strong interconnect foundation, none of which would be possible.
Well apparently Nvidia has started laying the groundwork for its own MCM interconnect. With the likes of its monolithic Tesla V100 coming in at a massive 812mm² (the limits of current single die manufacturing) Nvidia is desperate for more die space, and MCM is the only way forward.
In a whitepaper Nvidia discuss manufacturing easy to make small GPUs modules (GPMs) and integrating them together with high bandwidth power efficient signalling technologies (ah-la Infinity Fabric), offering a product that was 45% faster than the largest currently manufacturable single chip design and within 10% of a theoretical (currently unmanufacturable) monolithic design with the same transistor count. Interestingly it is also 26% faster than a multi GPU system (read SLI) with the same transistor counts.
Fun times ahead. It looks like graphics cards will forever change in the future with multiple GPUs integrated on the one card. It’ll redefine what it means to have a multi-GPU system.