The poll doesn't specify in which scenario processor x would be faster. In multithreaded applications, I'm confident that Zambezi will destroy anything Intel has on the market at the moment. IPC-wise though? I'm guessing a 2500k will continue to be the best choice for emulation.
My PREDICTIONS based on the details released so far:
Integer IPC per core: much higher than sandy bridge for sure
Floating point IPC per core: the same or slightly lower than sandy bridge
SIMD IPC per core: we have no idea, very little details have been released about how they are handling SSE in the new FPUs, most likely it will be on par or slightly lower than sandy bridge
MIMD IPC (avx) per core: most likely the same as sandy bridge
Multi-threaded integer IPC: MUCH higher than sandy bridge
Multi-threaded floating point IPC: About the same as sandy bridge
Multi-threaded SIMD IPC: About the same as sandy bridge
Multi-threaded MIMD IPC: About the same as sandy bridge
Clock rate: higher for sure, the FX 8170 will run at 3.9GHz and turbo up to 4.5GHz with all cores active
Memory throughput: they have improved the memory controllers over K10 a lot but how will it perform against sandy bridge? We don't know. Most likely on par since the clock rate is 50% higher and that's about how much higher sandy bridge's memory throughput is vs. K10.
Cache performance: better than K10 for sure, not enough details to compare it to sandy bridge though, the size of the L2/L3 cache is much larger, especially L2, but the size of the L1 cache on each core is much lower, it has something to do with the new LD/ST logic
HT (hypertransport) performance: about 25% higher than K10, however that is still lower than sandy bridges QPI performance
Overclocking: much better than sandy bridge. Should be able to laugh at 5GHz thanks to deeper pipelines and smaller die size.
Nothing can be said for sure until it is released though. Please remember though that a single module can accept 1 or 2 threads from the OS. If it accepts 1 thread the execution resources of both integer cores in that module are available to that one thread, resulting in higher IPC. Overall I like the direction they are going in. Bulldozer is the redesign they need. K10 was a dissapointment mainly because even though they made big changes to the backend (scheduling/register rename/execution/load/store/cache) the frontend was left mostly unchanged (brand prediction/prefetch/fetch/decode). The last time AMD did such a complete overhaul of their desktop architecture was K8. CMT seems like an interesting idea for handling threading. Will it be successful? We don't know, but it is certainly very unique. I disagree with AMDs definition of a "core" though, what they are calling a "module" to me looks like a core. How can it be a core if it only contains some of the execution logic and none of the frontend logic?
I also have no doubt in my mind that although it will likely do very well against sandy bridge for a few months ivy bridge is going to laugh at it come early next year.
Gulftown will probably still be faster than Zambezi (6core/12threads v 8cores/4modules).
Wish I could share your optimism NV. Intel wouldn't be delaying SNB-E/IB if it threatened SNB. Even if your right all Intel has to do to steal AMD's thunder is paper launch a faster SNB CPU (like the P3 1GHZ) and/or price cut their current ones. Maybe you can add some XOP instructions to Dolphin to turn the tides.
Quote:Maybe you can add some XOP instructions to Dolphin to turn the tides.
1. That's the compilers responsibility
2. I couldn't. The only language I'm actually any good with is C++. Assembly/machine code is batshit insane for people like me.
3. What possible use could dolphin have for XOP instructions?
Quote:Intel wouldn't be delaying SNB-E/IB if it threatened SNB.
What does SNB stand for? Is that supposed to be sandy bridge?
Also you act as if intel has control over these things and can magically make a cpu available earlier by snapping its fingers. For all we know the 22nm trigate manufacturing is not ready yet. Or perhaps a flaw was found in the cpu or chipset (cougar point anyone?).
Quote:Even if your right all Intel has to do to steal AMD's thunder is paper launch a faster SNB CPU (like the P3 1GHZ) and/or price cut their current ones.
If I'm right bulldozer will compete well against sandy bridge for a few months before ivy bridge crushes it. No increase in clock rate can take that away.
Quote:Wish I could share your optimism NV.
What optimism? I'm saying bulldozer will be able able to compete with a half a year older intel architecture for a few months before being crushed as usual. The data we have so far seems to suggest this but we don't know for sure yet.
Quote:Gulftown will probably still be faster than Zambezi (6core/12threads v 8cores/4modules).
I highly doubt that. 8 cores at a higher clock rate with a higher IPC than K10. The IPC would have to be MUCH lower than gulftown just to be comparable in multithreaded benchmarks. And even if it is on par or slightly slower, that would still be good. Gulftown cpus are $900+, if amd could compete with intels high end for once that would be considered a great success not a disappointment, don't you think?
Internet 101 smileys=joking/sarcasm I wasn't serious about the XOP thing. I better ask for MinGW support so we don't have to deal with more I can haz SSE4a type threads.
SNB=SandyBridge I guess SB (south bridge?) was taken.
IB was delayed due to predicted demand going down. No need to spend extra $ to get it out this year which IMO means Intel doesn't view BD as a threat.
SNB-E delay seems to be out of their control though.
There is a
$600 Gulftown but your arguement is still valid. If at worst a $300 CPU is only about equal to a $600 one then indeed it would be a success (I just want the K7/K8 glory days back). I'm probably taking the rumors too seriously (it sux in SuperPi its a POS
).
Looks like they are planning to compete with high end sandy bridge after all.
