There seems to be a lot of conflicting articles, posts and emails everywhere.

With memory performance I didn’t aim specificly on bandwidth. I meant the overall speed of memory usage, whether this is increased due to more or better l2 cache or just faster memory. A modal application and OS relies a lot more on memory operations than on floating point operations. Offcourse specific applications use floating points a lot more.

In my opinion the architecture doesn’t really matter, but the layer on top. Although I think it is great that there has been some movement in the market.

I didn’t mean the windows ‘kernel’ but rather the not-architecture-specific stuff on top like for instance the desktop and the explorer which will be rewritten in .net. Microsoft old plans to do almost everything in .net were just not practical enough.. and they took too long like the rest of their .net inovation.

There’s been some confusion around the TFLOPS theoretically available in the PS3. Sony quotes a complete system throughput of 2 teraflop. however, the 1.8 teraflop are the GFX card, not the cell… cell is supposed to have 218 gigaflop; and even this number is ridiculous, because it’s obvious that with such a limited architecture, there’s not much that’ll run anywhere near that many flops of usefull instructions on the cell.

Doubling memory bandwidth doesn’t add more performance than doubling flops (as example, take the well known benchmark superPi); and anyways, the cell has been architected with local memory and the like to try and circumvent the bandwith problems somewhat. Obviously you need an architecture without screaming bottlenecks if you want good performance; however, memory bandwidth isn’t the bottleneck many people think it is because cache sizes have increased significantly (and in cells case, you have the local memory).

If I understand correctly, then the linux-on-cell isn’t sony’s thing, but IBM’s, which makes sense as it’s natural for IBM to want to use these things in scientific computing and the like in future machines.

I think the cell is the best invention since sliced bread (well…) – but not because it’ll blow us away performance wise, but because it finally marks a real step in an architechtural shift that’s been a long time coming.

Oh and btw, microsoft is using .net a lot less in longhorn internals than planned, the speculation being because of performance problems. .net will remain the api of choice however, it’s just that the actual code cruching internally with be “unsafe”.