What is the architectural structure, memory hierarchy, and processing unit organization of older generation GPUs compared to modern ones? How do the evolution of shader units and differences in rasterization stages impact performance and energy efficiency? Do you think these differences could still give older cards an advantage in certain applications? Share your thoughts and experiences.
How do the architectures of older generation graphics cards compare to modern GPUs?
👁️ 112 views💬 2 replies❤️ 0 likes
2 Replies
Bro, I used to run a GTX 560 and HD 5870 around 2012-2013 for older indie games and 3D apps below 1080p all the time. Back then, GPUs had fixed-function rasterizers and separate pixel/vertex units, with shader cores split between limited "pixel shader" and "vertex shader" blocks. Modern GPUs, though, have merged all those units into a single unified shader pool, and the rasterizer now mostly acts like a post-process stage—letting the same hardware handle geometry and compute tasks more efficiently at the same time. Power-wise, old cards sometimes even edge out newer ones per watt thanks to lower clock speeds and fewer transistors.
From experience, the cases where old cards still shine are low power draw and apps built on pre-DirectX 11/12 APIs. A friend, for example, ran an old CAD program at low res on an HD 5870 because the software couldn’t fully leverage newer cards’ overhead, and the legacy rasterizer gave enough performance with lower energy costs. Still, new-gen GPUs crush it in shader count and bandwidth, dominating both raw performance and power efficiency—old cards only have a tiny niche left where they can compete.
Older generation GPUs had a much more limited architecture compared to today's AMD RDNA or NVIDIA Ampere GPUs, especially those based on pre-GCN VLIW shader units. Back then, shader units typically processed a single instruction packet, and rasterization was handled by a separate rasterizer block with a limited memory hierarchy (usually DDR2/3). Modern GPUs, on the other hand, feature unified shader units and reduced memory latency through layers like the register file, L1/L2 cache, leading to significant jumps in both performance and power efficiency for the same workload.
I think the only real advantages older cards have now are their ability to still deliver reasonable FPS in legacy DirectX 9/10 games or low-resolution retro applications, or their low power consumption when running simple shaders in CAD programs or basic scientific calculations. In niche scenarios, like offloading basic physics calculations to a GPU in an old game server, older, cheaper cards can still be cost-effective. Personally, I still use my old GTX 560 Ti to smoothly run older RPGs at 1080p, but when you're talking about a modern RTX 4070, the difference in teraflops per watt is no longer up for debate.