Discussion in 'CPUs' started by Boomstick777, Apr 10, 2015.
Interesting bit of news
Godavari? What about Carrizo?
Carrizo is for mobile,Godavari is a refreshed Kaveri for desktop. From what I gathered the major focus of Carrizo was a reduction in total die area,and a decrease in power consumption,however this has come with the penalty of lower clockspeeds at higher voltages(something to do with the high density libraries IIRC). Hence I suspect a desktop version might have not been worth the effort.
Can anyone decipher the following and set it out in laymans terms what it actually means? Lol
Basically AMD is moving to a more traditional CPU design instead of the CMT-like one they had with Bulldozer and its derivatives. Hence instead of two cores with one FPU its back to one core with one FPU.
Some more info:
I think it's a shame, since that was one of the few things differentiating it from Intel's lineup.
I'll be surprised if changing from CMT to SMT fixes all their problems. Surely if an 8 module (16 int 8 fpu) CPU from the current range is getting destroyed by a quad core then it would be better to make your modules more efficient than throw more fpus at it? (Presumably to 16/16?)
Was the problem that int workloads aren't as ubiquitous as AMD were gambling they were? That would be a cockup wouldn't it. Or could it be compilers still don't properly optimise for modules?
IIRC you can't optimise for modules you just optimise for threads/cores. AMD's module approach was along the same lines as intel's Hyperthreading approach, give a CPU the ability to handle more threads than it truly can. It's just AMD's was a better implementation (using worse cores).
If the new AMD CPU has ditched the failed module approach and they will be launching true octocores with Sand bridge grade IPC or higher then I'm in
A "source" (read: post on a forum) indicated that they're targeting IPC between Ivy Bridge and Haswell. Of course, Skylake will be out by then.
If this chip is anything like these rumors, with the IPC being close to IB / Haswell, then im trading my 4770k in for one
That's not true AMD's implementation was (and still is) much poorer in regular usage such as lightly threaded scenarios because most the processor goes unused, whereas with Hyperthreading even if you only have 4 main threads you still have 90%+ of the processor providing the grunt to execute them.
AMD's cores (modules) weren't really worse than Intel's because if you compare an FX8350 and 2500K they perform about the same speed under ideal conditions when all of the processor is being utilised (eg. encoding), the problem is that AMD chose to split their cores straight down the middle leaving them short-handed in less than ideal conditions.
Considering how much CPU performance has gone up since SB (which honestly doesn't seem much), AMD could do pretty well here. I guess it will come down to price and power consumption in the end. Still probably won't get me to upgrade but it should be some decent competition to Intel if it does well.
So they will be selling true hex/octo cores with better IPC than IB-E?
I think you missunderstood, I was saying that the module design approach is better than the Hyperthreading design approach, not that any of the processors made using the module approach were better than any of the processors made using the Hyperthreading approach.
I.E if you made an 8 core FX CPU using the cores from a haswell i5 then it would destroy both the FX-9590 (because it has better IPC) and depending how many cores are used it would match or destroy the 4790K (because it has a better multithreading implementation).
It's not going to matter if Skylake's out by then if AMD are launching an 8 core with that IPC.
I really can't see it, but if they do manage that, it'd change everything over night. AMD would be literally blazing it.
If you read back before BD was launched the CMT approach was an attempt to try and reduce the amount of die area required for multi-threaded software since AMD was at a process node disadvantage to Intel. However,as indicated by another poster in lightly threaded situations,it lead to poor utilisation of the modules - BD was probably more a server CPU shoehorned for consumer workloads. But due to problems with the 32nm process,it meant delays and missed clock targets too. Its probably why Intel has done tick-tock to try alleviate a double whammy of issues,although SB was both technically a tick and a tock in one.
The use of SMT by Intel is more a throughput and core utilisation move. Intel uses more complex larger cores,and by replicating some of the front end and using clever branch prediction and scheduling it means that it minimises the percentage of the core which is left idle.
Actually was wrong on SB - the 32NM process was already being used for some Intel CPUs.
Clarkdale was 32nm wasn't it?
My main concern is that AMD is still going to launch a new chip on a new unproven process.
If the process does not pan out - it might screw them over with regards to clockspeeds.
Its basically what happened to Kaveri - it was released a year late with lower than predicted clockspeeds.
If it is true then the new AMD flagship would be a direct competitor for the 5960X. Of course it should have slightly lower IPC but it shouldn't have the £860 price tag either, by comparison it will have higher IPC and 33% more cores than the 4930K which still goes for >£400 where stock can be found.
Separate names with a comma.