BD scales worse than SB actually. Not a positive there...
I think I need to clarify, for clockspeed gains it scales better, Sandy bridge has a kind of dip when you reach what are abnormally high speeds, Bulldozer does not seem to suffer in this respect, its has a near liniar scaling with clocks, which would indicate the architecture has lots of room of clockspeeds if thermals can be kept in check for the future, in terms of per core scaling, no it does not scale as well as obviously it has shared core resources. But think of it this way, if IPC can be improved without increasing size much then the design could all of a sudden be very appealing. the equivalent of 8 thuban cores at 4+ GHZ stock vs 4 ivy bridge 4+ghz cores is something worth noting.
There is no such thing as a "dip". Sure you need higher and higher voltages per clock when you get over a point but performance is unaffected.
I think I need to clarify, for clockspeed gains it scales better, Sandy bridge has a kind of dip when you reach what are abnormally high speeds, Bulldozer does not seem to suffer in this respect, its has a near liniar scaling with clocks, which would indicate the architecture has lots of room of clockspeeds if thermals can be kept in check for the future, in terms of per core scaling, no it does not scale as well as obviously it has shared core resources. But think of it this way, if IPC can be improved without increasing size much then the design could all of a sudden be very appealing. the equivalent of 8 thuban cores at 4+ GHZ stock vs 4 ivy bridge 4+ghz cores is something worth noting.
If they can reduce the cache latencies to 1/5 or less (and increase the clock speed of the L3 cache) of what they currently are, then it may not have been a bad design (although the high cache latency may be inherent to the design; I don't know)
The super high latency cache is probably the largest problem.
bulldozer front end is actually a worst\equal problem,
it can feed 4 interger ops, it would be ok, since one core have 4 ops.... but it is shared with 2 cores...it should actually feed 8 ops!!!
seriously, i still like the idea behing bulldozer...but amd screw it up badly
The idea is that one core is unlikely to to saturate the front-end, so tying it to two cores will increase average utilization.
Does anybody know if the front-end is the main bottleneck in BD? I have a hard time believing it, just because BD seems to be a dog even in ST performance...
It's never a waste to try and innovate. Even if something doesn't work the way you wanted, hoped or expected you can learn from it. It's awesome when everything works right the first time, but still quite valuable when it doesn't.
Intel already concluded a long time ago way before BD that a new uarch built on an entirely new process was a bad idea, nevermind the whole high-clock, low IPC approach was stupid (once again proven by Intel), it was a trainwreck right from the very beginning. More like AMD can't learn from history here.
Only time will tell if it was a waste. I suspect that as they make changes to it, it will prove to be worth the $/Time/Effort.
I think I need to clarify, for clockspeed gains it scales better, Sandy bridge has a kind of dip when you reach what are abnormally high speeds, Bulldozer does not seem to suffer in this respect, its has a near liniar scaling with clocks, which would indicate the architecture has lots of room of clockspeeds if thermals can be kept in check for the future, in terms of per core scaling, no it does not scale as well as obviously it has shared core resources. But think of it this way, if IPC can be improved without increasing size much then the design could all of a sudden be very appealing. the equivalent of 8 thuban cores at 4+ GHZ stock vs 4 ivy bridge 4+ghz cores is something worth noting.
Bulldozer does not seem to suffer in this respect, its has a near liniar scaling with clocks, which would indicate the architecture has lots of room of clockspeeds if thermals can be kept in check for the future,
BD is just a weird arrangement of old tech on a process that doesn't do it any justice, very illogical move for AMD in the long run.
As bad as bulldozer is, it's not as bad as the initial Pentium 4. And later, with Northwood, the P4 was basically competitive with AMD's offerings, even if it was slightly inferior in some ways.
If AMD can improve bulldozer in it's first revision as much as Intel improved the Pentium 4, it will be a very strong CPU.