Intel vs AMD chip designs
http://news.com.com/2061-10791_3-6047412.html.
Very interesting, intel makes a higher profit by keeping the memory controller off its precision fab CPU, gets more CPUs per die than AMD. The opteron NUMA (Non uniform memory architecture) is tricky, but even its slowest 3 hop memory fetch is faster than the normal 1 hop memory fetch on the intel chipset.
The tradeoff is smaller L2 cache, some of the transistor budget for that is used for the integrated memory controller on the opterons. Oh, and intel makes more profit. Thats a tradeoff too, on another budget. Yours to thiers.
Who wins? Right now opteron. When memory sizes (RAM) shoot to 2-4 Terabytes for a server? The ratio of cache to RAM will continue to expand, and the tradeoff intel is making could very well not show well as RAM densisty and sizes take off again. Your hit ratio goes down as apps use more memory, unless you can keep expanding cache.
It also depends on the app - small apps - cache. Massive RAM simulations, like I have spent (is it twenty?) years of my life writing and architecting - integrated memory controller. And the benchmarks prove the theory on this one.
Whats funny to me, is this is clearly a piece of Intel PR, and yet, even with thier hand on the pen, makes me prefer the AMD CPUS. But my benchmarks had me biased already. So did my air conditioning bill, and the fact that intel rigged thier compilers to produce "not as optimal" code when producing code for AMD chips. (see slashdot for details).
Very interesting, intel makes a higher profit by keeping the memory controller off its precision fab CPU, gets more CPUs per die than AMD. The opteron NUMA (Non uniform memory architecture) is tricky, but even its slowest 3 hop memory fetch is faster than the normal 1 hop memory fetch on the intel chipset.
The tradeoff is smaller L2 cache, some of the transistor budget for that is used for the integrated memory controller on the opterons. Oh, and intel makes more profit. Thats a tradeoff too, on another budget. Yours to thiers.
Who wins? Right now opteron. When memory sizes (RAM) shoot to 2-4 Terabytes for a server? The ratio of cache to RAM will continue to expand, and the tradeoff intel is making could very well not show well as RAM densisty and sizes take off again. Your hit ratio goes down as apps use more memory, unless you can keep expanding cache.
It also depends on the app - small apps - cache. Massive RAM simulations, like I have spent (is it twenty?) years of my life writing and architecting - integrated memory controller. And the benchmarks prove the theory on this one.
Whats funny to me, is this is clearly a piece of Intel PR, and yet, even with thier hand on the pen, makes me prefer the AMD CPUS. But my benchmarks had me biased already. So did my air conditioning bill, and the fact that intel rigged thier compilers to produce "not as optimal" code when producing code for AMD chips. (see slashdot for details).
Comments