Replies to post #1160 on AMD strictly moderated (AMD)

[pgerassi]

bmw:

You must know that Intel TDP is using some low power use programs. Prime95 easily beats Intel's TDP rating on their CPUs. For dual and later quad core, you need to run 2 to 4 copies. Typically using the same methods AMD uses to rate their CPUs maximum power draw, Sum(Imax*Vmax)) over all supplies yields a figure 33% higher than Intel's TDP rating. Then of course you need to add in that used by the NB for the FSB switch and memory interfaces. I didn't include that used by the FV-DIMM buffers over the underlying DRAM.

And you forget to add in the additional latencies when adding FB-DIMMs into each channel plus the cost of switching in the NBs and MCHs. Adding one FB-DIMM to a channel increases latency more than one HT hop. Then there is the fact that 4 cores compete for one FSB whereas in the Opteron that traffic goes through the far higher bandwidth of the on die crossbars (XBAR and SRQ) clocked at CPU speed versus chipset speed (2.8GHz versus 333MHz). Lastly the effective cache sizes because Opterons use exclusive caches and can use the remote copies is more like 9MB for the quad 90nm and 17MB in the quad 65nm dual socket. Because Intel uses inclusive cache, they need that at every socket.

As to remote uses, that is mitigated by intelligent use of processor/memory affinity. The occasional need for remote memory is typically used in synchronization and coordination. And that's a small percentage of overall memory requests. If it wasn't, caches would be completely ineffective. Latency is the king in most applications even server based ones. Here the advantage is all AMD.

At the typical quad socket level each FB-DIMM channel (of 4) would have to have 4 FB-DIMMs on it compared to 2 DDR2 DIMMs in each of 8 channels of AMD (AMD can have 4 DDR2 DIMMs per channel). That amount of FB-DIMMs will make latencies increase faster for Woodcrest than that of Opteron widening the present latency gulf.

So high BW needing solutions go to AMD, Low latency needing solutions go AMD. Higher IPC needing solutions also go to AMD. The only thing still going for Intel is complacency (and corruption).

Pete