Savings made by moving to a 22nm process aren't as generous as previous drops
Frustration mounts as Nvidia mulls over the future of its GPU architecture. ExtremeTech has revealed that slides presented by Nvidia at the International Trade Partner Conference (ITPC) forum in November last year openly criticised Taiwan Semiconductor Manufacturing Company (TSMC), who are responsible for manufacturing their GPU cores.
The issues, as you'd expect, are about money, and allegedly stem from the lack of responsibility and clear communication on TSMC's part. One particular slide throws down the gauntlet with a passive aggressive list titled "What partners do?", outlining Nvidia's expectations from their partners. To quote one of the points made, "[partners should] understand each other's REAL needs, capabilities and constraints" (emphasis theirs).
Nvidia's projections for normalised transistor cost are the crux of this debacle. The graph that was shown placed 20nm and 14nm manufacturing processes as providing negligible reduction in transistor cost. Nvidia attributes this to increases in wafer prices, the complex process, and the fact that good yields don't quite cut it anymore. This explains why the GK104 architecture (as used by the GTX 680) is already focusing on efficiency instead of ramping up the transistor count.
Nvidia also provided potential solutions to the issues ahead, including better collaboration and waste reduction, creating trust between partners, sharing costs, and acting as a "virtual IDM". An IDM (or Integrated Device Manufacturer) is a company which fabricates their own IC designs.
What does this mean for future graphics cards? The transistor count is unlikely to increase significantly in future designs, and may even drop to keep die size and costs down. Software will have to take the lead to provide performance increases and added efficiency to prevent complicating the architecture. This may even cause an anomaly in the prediction made by Moore's Law, despite GPUs regularly placing ahead of the curve due to early adoption of new manufacturing processes.
TSMC is the world's largest independent semiconductor foundry, and provide services to AMD, Nvidia, Qualcomm , Broadcom, and VIA to name a few. TSMC are yet to make a public comment about Nvidia's concerns, but have coincidently just announced plans to introduce Chip-on-Wafer-on-Substrate (CoWoS) technology to reduce the cost of larger dies.