Nehalem-EP......BLOOMFIELD

[savantu]

I think what you will find is different apps seeing some healthy gains, others seeing none or little to nothing. On the whole I suspect healthy overall but not overwhelming single threaded jumps in performance (similar to K8 to Phenom). There are examples where, clock for clock (IPC), Phenom showed moderate to very slight gains, and in other apps nice 15-20% improvements.

In this case, SP1M -- there is info on the web that Intel buffed up loop prediction, as SP1M is inherently recursive I would postulate that the benefit we see here is coming some from that architectural improvement. Speculation on my part though....

Just looking at the SP1M, roughly 13% gain clock for clock -- which is healthy for single threaded performance ... not quite a jaw dropping as the Conroe leap over Netburst, but still healthy.

Jack

You know very well that Core and especially Penryn ( with its improved latencies for some instructions ) are the pinnacle of single thread performance.
With the advanced prefetchers , very large and extremly low latency L2 they keep the large number of execution units busy and offer superb single thread performance.To improve on that with a completely different cache structure ( inferior IMO for single threaded apps ) and a CPU designed for scalability and multithreading is nothing short of astounding.

People expect Nehalem to improve dramatically over Penryn , but they fail to realize how hard that is as the Core/Penryn uarch simply loves single threaded apps.

AMD K10 has a full 3 pages of improvements over K8 and can't touch Core even with a 10 foot pole in single threaded apps.Imagine what work went into Nehalem to offer even that measly 2% in Cinebench.