Bij Ace's Hardware is deel 6 van de serie "Secrets of High Performance CPUs" gepost. Dit deel heeft de titel "The Future of x86 Performance" meegekregen. Er wordt dit keer uitgebreid ingegaan op verschillende methodes en technieken die gebruikt kunnen worden om x86 CPU's sneller te maken, zoals caching, reorder buffers, branch prediction en pipelining. Ook wordt uitgelegd waarom een CPU op papier veel sneller kan zijn dan in werkelijkheid, en natuurlijk wat voor trucs we in de toekomst van de CPU designers kunnen verwachten:
The future of x86 performance is not a brainiac or very wide superscalar processor. It is very unlikely that we will ever see an x86 CPU which has the ability to decode and execute 4 x86 instructions per clockcycle. It is clear that the old x86 instructions set reduces the average amount of ILP and that x86 CPU designers will focus on clockspeed and average memory latency.
The most significant improvements in performance will come from bigger, faster, more intelligent and efficient cache systems and other techniques which will dramatically reduce the average memory latency. Not long ago, a CPU designer said to me: "it is incredible how much performance increase you see from integrating the memory controller in the CPU die, considering it does not consume much die space." Yes, large and complex reorder buffers might be able to partially hide a bit of the memory latency, but a relatively simple integrated memory controller will reduce the latency of the main memory by tens of cycles (the chipset will be no longer middleman between the CPU and RAM).
Lees de vijf (vrij technische) pagina's hier. Mocht dit allemaal abacadabra voor je zijn raad ik je aan om bij deel 1 te beginnen.