AMD CPU speculation... and expert conjecture

[BeastLeeX]

oh i agree. im just saying. im looking forward to a powerful cpu. amd needs a powerhorse on there shoulder. no wonder there losing money, first bought ati. than stop competing with intel. now gonna mostly do arms chips. i hope they dont scrap there performance series. cuz im gonna be pissed.

Hopefully AMD can get this cpu on par with Sandy in single threaded, if they do, then I will happily buy the FX-8x20 of Steamroller, and hopefully more will upgrade from their old Phenom's also.

 

[gamerk316]

so your saying programmers won't know what threads to schedule to a locked core? could be the main problem right here if programmers aren't tought what their thread actually does.

noob, really simple example here: Say you hard lock your heavy workload thread to core 7 of an 8 core system, since that core almost never gets used. For the sake of argument, you give this thread "above-normal" priority, to ensure it gets control of that core. Now, some OTHER program goes along, schedules its heavy workload thread to core 7 of the 8 core system, since that core is almost never used, and gives its thread "high priority". Guess what? Your first applications thread will almost never run, and it can't move to another core because you locked it to just that one core. Guess what? You just tanked your applications performance. This is why outside of platforms where you have a guarantee of what resources are available for use (consoles, and most all integrated platforms), you do NOT hard lock threads to cores.

You are making the same exact arguments that were made back in the 70's in regards to manually setting variables in the CPU registers. Most programming languages still support some form of the "register" keyword, which allows the developer to manually specify what register to put a variable in. The idea at the time was that some heavily used variables would be better off locked in a register [say, the control variable of a FOR loop], to avoid doing a lot of costly memory reads. However, that also reduces the number of registers available for the rest of the system to use, and if you do a lot of processing, performance could suffer. FYI: No one would ever consider doing this anymore, because in every instance, the compiler is smarter then you are when it comes to putting variables in registers. Same thing applies to threads: The OS has a FAR better understanding then you do of the entire system state, and will do a better job then you in regards to thread allocation.

The ONLY instance I'd ever modify the number of cores a thread could run on:
A: I know I have to high-workload, totally independent threads. In this case, I'd take steps to try and ensure they go on different cores [and even this is risky if not handled really carefully].
B: The CPU cache is split between different cores, and as a result, I would try and limit threads to the cores that have access to the same cache.

And gamerk, don't dismiss 64bits that easily in multi-threading. You'll have a bigger address space to actually allocate resources for threading. In the calculation part, you'll be able to have more variables in memory at a given time to consume. You have to look at it from that POV. I know it won't be a huge deal, but it will be like the transition from 16bits to 32bits. All in all, it's not a game changer, but it adds to the party, hahaha.

The overhead, in terms of Address Space usage, is trivial, unless you are coping large amounts of data across the thread boundary. I point out, if you need to replicate that much data across threads, you probably shouldn't be using a separate thread in the first place.

 

[sarinaide]

well i hope that steamroller will be 20% over my phemon

http://www.anandtech.com/bench/Product/102?vs=697

I think by the time Steamroller comes out it will be like 50-60% faster than your 955BE. The average differential between a 1100T to FX 8150 to FX 8350 is around 20%, every workload the 8350 crushes its previous flagships.

 

[truegenius]

then how about increasing priority of heavy/critical threads upto high priority ?
for example skyrim (iirc , this was the game that we discussed in awhile that it love being locked to 1 core only )

i think/imo that k10 is 10% better than piledriver in per ghz in single thread and 40% slower in multithreaded (1 module vs 2 core) at same clock speed
so i think that in single threaded tasks at same clock, steamy can reach upto 1st gen core series (lga1156, Nehalem) level only.

 

[BuddiLuva]

Hopefully AMD can get this cpu on par with Sandy in single threaded, if they do, then I will happily buy the FX-8x20 of Steamroller, and hopefully more will upgrade from their old Phenom's also.

Count me in.

 

[sarinaide]

I would expect Steamroller is going to continue the trend, maintain or lower power window, ramp up clockspeed and with a far more advanced Front End and added FPU's along with the 40% RAM and Cache Latency reductions will see sizeable IPC gains. If AMD got around that Sandy/Ivy territory it will be a pretty damn good effort considering the resources at disposal.

I have just about every AMD chip, the 8350 is definitely a step up from the 1100T where the 8150 was more a side step newer tech same performance. The 8350 in every workload tested is considerably faster than its prior parts and this is a ES chip so only running at 3.3ghz. I will be getting a 6300 soon to test relative to the 6200, from what is seen it to is a step up at times rivaling x6's.

 

[BeastLeeX]

Radeon HD 8000 series is now confirmed to be released in Q4

http://wccftech.com/amd-confirms-radeon-hd-8000-sea-island-delay-hd-7000-series-remain-primary-focus/

 

[JAYDEEJOHN]

Radeon HD 8000 series is now confirmed to be released in Q4

http://wccftech.com/amd-confirms-radeon-hd-8000-sea-island-delay-hd-7000-series-remain-primary-focus/

http://forum.beyond3d.com/showpost.php?p=1703845&postcount=906
Now, we could play ping pong or he said she said, but lets just slow it down now.
Dont believe everything you read, and today more than ever, certain sites seek hits

 

[BeastLeeX]

http://forum.beyond3d.com/showpost.php?p=1703845&postcount=906
Now, we could play ping pong or he said she said, but lets just slow it down now.
Dont believe everything you read, and today more than ever, certain sites seek hits

I realize that, but i doubted they would release right now anyway. There is also a lot of speculation and rumors already confirming this, so I guess you can believe what you want, and we shall see how things turn out!

 

[JAYDEEJOHN]

I understand, and Im not picking on you.

My best guess would be, look into when the 20nm node would be ready, take that time from when its ready from the 7xxx series release, divide by two, and youll be close.

 

[m32]

Aren't the next gen Nvidia cards rumored for Q4 also? I think I believe it. I think there are demands for other 28nm products, and pressure on card makers to make as much money on the current gen products, to push'em back. This is especially true on AMD end since they have alot of produces they want to get out in a years' time.

 

[JAYDEEJOHN]

Heres something to look at
Looking further ahead, Chang said his company's already seen enough clients and demand for the upcoming 20nm manufacturing process, which should have a more significant financial contribution in 2014. The exec also predicted that at TSMC, its 20nm production will see a bigger growth rate between 2014 and 2015 than its 28nm counterpart did between 2012 and 2013 -- the former should eventually nab close to 90 percent of the market, said Chang.

http://www.engadget.com/2013/01/18/tsmc-28nm-process-2013/
So, if 20nm will see a greater gross if usage in 2014 year ending than 28nm for 2012, it appears it will be very ready early next year.
Now, in the past, AMD has jumped on early process dev/first to market.
If this takes place say 2nd qtr 2014, and we dont see the 8xxx series til qtr 4 this year, thats only a short period for skus to sell, even for a refresh

 

[noob2222]

noob, really simple example here:

Go back and read what I wrote. I was referring to an update to the windows scheduler for programmers where windows wont put 2 locked threads on the same core if that ever happens. Obviously it wont happen because programmers don't want the ability to hand schedule anything even if its done in a way that improves performance.

EDIT: Keep it civil, guys...

 

[BeastLeeX]

Heres something to look at
Looking further ahead, Chang said his company's already seen enough clients and demand for the upcoming 20nm manufacturing process, which should have a more significant financial contribution in 2014. The exec also predicted that at TSMC, its 20nm production will see a bigger growth rate between 2014 and 2015 than its 28nm counterpart did between 2012 and 2013 -- the former should eventually nab close to 90 percent of the market, said Chang.

http://www.engadget.com/2013/01/18/tsmc-28nm-process-2013/
So, if 20nm will see a greater gross if usage in 2014 year ending than 28nm for 2012, it appears it will be very ready early next year.
Now, in the past, AMD has jumped on early process dev/first to market.
If this takes place say 2nd qtr 2014, and we dont see the 8xxx series til qtr 4 this year, thats only a short period for skus to sell, even for a refresh

Could mean 20nm for Excavator (If they actually make it, which haven't seen any signs of it other then on roadmaps)

 

[JAYDEEJOHN]

In my link, it said much about the newer SR chips coming on 28nm.
I think with all the market change, the restructuring at AMD, plus the node availibility as well as maturation of them has a few things in the way for a perfect roadmap, and excavator and how it fits into all this is being defined along the way

 

[esrever]

Could mean 20nm for Excavator (If they actually make it, which haven't seen any signs of it other then on roadmaps)

cpu != gpu. AMD's 28nm GPUs have been shipping for a year now. Their 28nm CPUs haven't even hit the market yet. Could be a long time until excavator.

 

[iceclock]

well see jaydee. amds doing bad financially. best thing for them might be arm processors and low powered mobile processors.

 

[JAYDEEJOHN]

It all depends on how the SoC market blooms, the uses and needs of HSA, gpgpu, gpus themselves on SoC etc etc, all things I can think of being used for larger things like PCs or special devices like consoles etc.
At the lower nodes, and with better perf and much better power controls, it also allows AMD to go bigger as well, since itll be their gpus/igps used, easier and better designs put forth to costumers etc, as we saw with the consoles, where Im sure everyone had an iron in the fire for their wins

 

[BeastLeeX]

cpu != gpu. AMD's 28nm GPUs have been shipping for a year now. Their 28nm CPUs haven't even hit the market yet. Could be a long time until excavator.

I'm just speculating the best thing that could happen for AMD right now, we are not even sure if they will create Excavator, but if they do, it could be on 20nm, which would be awesome.

 

[mayankleoboy1]

http://www.engadget.com/2013/01/18/tsmc-28nm-process-2013/
So, if 20nm will see a greater gross if usage in 2014 year ending than 28nm for 2012, it appears it will be very ready early next year.
Now, in the past, AMD has jumped on early process dev/first to market.
If this takes place say 2nd qtr 2014, and we dont see the 8xxx series til qtr 4 this year, thats only a short period for skus to sell, even for a refresh

The only problem now is that Apple is in competiton with AMD+Nvidia+Qualcomm.
And we all know how Apple does its business. Meaning : Buy all/most wafers from TSMC , use hwhatever it can (and they use lot) , and only after then distribute the left over wafers to others.
At the 28nm era, Apple was using Samsung, so no competition for TSMC users. AMD could be the MVP for TSMC. But not anymore.

 

[JAYDEEJOHN]

The only problem now is that Apple is in competiton with AMD+Nvidia+Qualcomm.
And we all know how Apple does its business. Meaning : Buy all/most wafers from TSMC , use hwhatever it can (and they use lot) , and only after then distribute the left over wafers to others.
At the 28nm era, Apple was using Samsung, so no competition for TSMC users. AMD could be the MVP for TSMC. But not anymore.

Read the link, it speaks to how much growth is expected, almost triple.
This leaves room for others, as we also see Apples demand in the market slow as well.
The bigger question here is, do these numbers for 2014-15 include AMD apus?
They do for this year, which will be less than 20nm

 

[de5_Roy]

amd's hard-earned money is being well-spent by glofo (except amd doesn't seem to be benefitting from it).
"GlobalFoundries Teams Up with Cyclos to Speed Up ARM Cortex-A15 Designs.
Customers of GlobalFoundries Will Be Able to Add Resonant Clock Mesh to 28nm ARM Chips"
http://www.xbitlabs.com/news/cpu/display/20130206061859_GlobalFoundries_Teams_Up_with_Cyclos_to_Speed_Up_ARM_Cortex_A15_Designs.html
"GlobalFoundries: 14nm-XM Process Tech Two Times More Power-Efficient Than 28nm.
Preliminary Results Show Significant Declines in Power Consumption of 14nm-XM Chips"
http://www.xbitlabs.com/news/other/display/20130205234907_GlobalFoundries_14nm_XM_Process_Tech_Two_Times_More_Power_Efficient_Than_28nm.html
moar on 14nm which are not from intel. i think that the keyword here is 'common platform'.
"Samsung shows off 14nm test wafer
Common Platform 2013: Hybrid 20nm/14nm process is the next big thing at CP"
http://semiaccurate.com/2013/02/07/samsung-shows-off-14nm-test-wafer/
"IBM shows off 14nm wafers and flexible 28nm ones
Common Platform 2013: Flexy FD-SOI, FinFETs, fun fun fun"
http://semiaccurate.com/2013/02/05/ibm-shows-off-14nm-wafers-and-flexible-28nm-ones/

btw, what happened to rcm technology in piledriver? it seemed to me that pd's marginally better performance/watt came mostly from maturity of 32nm process, minor arch and bios firmware tweaks and implementing better [strike]throttling[/strike] thermal management in the cpus. iirc cyclos requested toms to remove their text on rcm from the fx8350 review. why'd they do that if rcm was well implemented in the fx cpus? or was it because pd turned out to be less rcm-friendly and in the end it was mostly a marketing gimmick?

 

[mayankleoboy1]

IIRC, RCM wasnt there in PD at all. It was supposed to be present in SR ?

 

[sarinaide]

RCM went AWOL, most reviews omitted RCM discussion simply because it wasn't ever presents. My guess was to cut Vishera's cost pricing since it was picking up the pieces Zambezi left.

 

[BuddiLuva]

Correct me if I'm wrong, but hasn't AMD struck a deal to supply the next gen consoles with AMD CPUs and GPUs?