The Germans and Friends will not let that happen=P Now I think folks ought to pay more attention to what Intel says sometimes, AMD's own fault. During the Lawsuit AMD brought against Intel here's what I saw. AMD is now the one fifth the size of Intel after ATI was absorbed. Yet they made one tenth of the possible profits even after selling between one forth and one fifth the total processors sold. They sold almost all of their product. Sounds like AMD needs to worry more about AMD than Intel.
Intel's showing 4GHz air cooled with no tricks I hope, shows a wall for this stepping. Both Intel and AMD have getting better as each new stepping is shipped.
Last edited by Donnie27; 08-05-2007 at 09:47 AM.
Originally Posted by Movieman
qft!Posted by duploxxx
I am sure JF is relaxed and smiling these days with there intended launch schedule. SNB Xeon servers on the other hand....
Posted by gallag
there yo go bringing intel into a amd thread again lol, if that was someone droping a dig at amd you would be crying like a girl.
SSE4 won't do anything for FP SIMD as almost all the instructions are for
integer operations/miscellaneous. Yes, cache matters, even the FPU deficient K6-3 was faster than K6-2 in games. Most benchmarks you see for games are video card limited.
To clear things up, Core 2 has 2 floating point units that do the most used FP operations FADD/FMUL as does K7/K8. The other units are FSTORE/FLOAD/FMisc(trigonometric functions like sine/cosine and SQRT) , Fdiv. K7/K8 floating point units can do 1 SSE/SSE2 operation every other clock cycle, while Core 2 has 1 cycle throughput. They are not symmetric, ie, each unit is designed to do FMUL or FADD, but not both. Max (Core 2) floating point performance is 8 SP FLOPS/cycle, and 4 DP FLOPS cycle but the fetch bandwidth isn't enough. Sciencemark gives max 4.33 SP/FLOPS/cycle and 2.2 DP/FLOPS/cycle. If K7/K8 can only fetch one 128 bit SSE operation per clock cycle (as can Core 2) but only execute in 2 cycles, than max throughput is 2 SP FLOPS/cycle, if K8 could fetch 2 128 bit SSE operations per clock, consisting of only FADD+FMUL than it could execute 4 SP FLOPS/cycle and 2 DP FLOPS/cycle. Both K8 and Core 2 have idle floating point units, as only a maximum of 1 (or just slightly over with a fused op?) 128bit SSE op can be fetched in 1 clock cycle. I assume that with increasing IPC heat generation becomes a major issue as more functional units are in use.
Conroe's extra cache really helps in large memory footprint pure number crunch apps, i.e, I'm interested in 4 meg Conroe because it's about 20-50% faster clock for clock than Conroe 2M in SMP F@H projects. IMC is good in branchy large memory footprint applications like games but Conroe's L2 cache is so much faster and bigger than K8's + the improved BPU and prefetching that K8 can't keep up.
http://www.hkepc.com/bbs/hwdb.php?ti...iew&rid=837360
"In addition, Intel former generation 65-nanometer products at the highest vein about 3GHz, clock more than 3GHz, power there will be significant growth, and 45-nm products up to a maximum of about 4GHz."
Seems they will need 4Ghz since Penryn is next to no faster than Conroe(5-6% on averarage,in 128bit test in coolarer forum there is no improvement at all!).
Generally speaking ,on average,the difference is because the 2MB of additional L2 and higher FSB
Look here:
http://www.xtremesystems.org/forums/...&postcount=299
Look at other scores there.Are you saying those are fakes ,or what?
Look at hkepc review,on average(with that 115% better test) it's only 9% faster in 46 tests.That is pretty massive amount of tests,so it represents a lot of scenarios.Without that 115% better result,its massive 7% faster than equally clocked Conore,amazing indeed.
Let's snip the pointless benches. Snip pcmark05 the antiquated "de facto of old benchmarking", ScienceMark the paid AMD bench http://www.sharkyforums.com/archive/.../t-165292.html, Sandra, pointless synthetic, Photo gallery+Office because CPU power is really needed for it (not), 3Dmarks (gpu benches obviously). Now put a sticker number on the realish benches that matter. Decent gains considering it's not a new arch ala Nehalem and some great gains with SSE4.
Last edited by red; 08-05-2007 at 02:22 PM.
Gaming benches looks good. With about a 10% in averge improvement. With HL2 being the topscore of 30%.
Nice!
Crunching for Comrades and the Common good of the People.
Well,you should wait for reviews before you can judge K10.Those numbers means nothing.15% in what?On average?In SSE(not a chance ).Useless.
And K8 is on average 15-20% slower clock/clock than Conroe.With those useless Sun pdf we can conclude K10 matches Conroe on average.We will see what reviews say ,not the pdf from Sun.
How do you define "SSE"? You expect encoding time to be halved because of "doubled SSE performance"? http://www.anandtech.com/cpuchipsets...oc.aspx?i=2808 Merom has 3x SSE units, doesn't mean 3x performance.
So how do you define "SSE"?
While we are still waiting for the first K10 benches we now know what K10's direct competition can do.
3.7GHz at 1.1V (even if it IS cherry picked) is dang sweet. 4+GHz on high-end air seems very likely.
I would say 5-10% IPC improvement from a tweaked "shrink" is a lovely bonus.
A 5 - 10% improvement across the board is impressive for a die shrink + minor uarch improvements. I'm not sure why Informal is trying to downplay the improvements, it's not like we were expecting Netburst -> C2D all over again.
Intel claimed 5 - 10% higher IPC, with SSE4 getting bigger boosts, and these tests confirm that.
Now that Intel has shown it's cards, it's time to see if AMD has an ace up it's sleeve in K10. I'm hopeful the latest rumours of 3GHz+ Phenoms are true, these would bring true competition to Penryn. Ahh, exciting times!
Bookmarks