AMD ''Vishera'' FX-Series CPU Specifications Confirmed

Status
Not open for further replies.

samwelaye

Distinguished
Apr 24, 2010
284
0
18,790
you know... at only 95 watts, and a 4ghz base frequency, and also improved IPC on piledriver, the FX4320 could be quite the potent value cpu. would quickly make the i3's obsolete at any rate. heres hoping they deliver!
 
[citation][nom]oddlyinsane[/nom]Still...its not the best thing around.[/citation]

You might be surprised. Disable one core per module to eliminate resource sharing (letting the still active core of each module get a significant performance per Hz boost while cutting power consumption by 30-40%) and overclocking the CPU and CPU/NB freuqneyc (increases the L3 cache frequency) on the FX-8120 and the FX-8150 can make them able to compete with the K edition i5s. Doing the same with this FX-8350 might let it be able to top Intel in lightly threaded performance until Haswell launches and even compete with the quad core LGA 1155 i7s in highly threaded performance with this *mod* and beat them without it.

It might actually be one of the best things around if used properly.
 
[citation][nom]samwelaye[/nom]you know... at only 95 watts, and a 4ghz base frequency, and also improved IPC on piledriver, the FX4320 could be quite the potent value cpu. would quickly make the i3's obsolete at any rate. heres hoping they deliver![/citation]

It might actually get Intel to update the i3s with Ivy if they intend to do that at all until Haswell comes out. Ivy i3s with raised frequencies relative to the Sandy i3s could make a good middle-ground between the i5s and the older i3s that can compete with AMD's quad core Vishera and Phenom II CPUs better.
 

idroid

Honorable
Aug 18, 2012
1,525
0
11,960
[citation][nom]blazorthon[/nom]You might be surprised. Disable one core per module to eliminate resource sharing (letting the still active core of each module get a significant performance per Hz boost while cutting power consumption by 30-40%) and overclocking the CPU and CPU/NB freuqneyc (increases the L3 cache frequency) on the FX-8120 and the FX-8150 can make them able to compete with the K edition i5s. Doing the same with this FX-8350 might let it be able to top Intel in lightly threaded performance until Haswell launches and even compete with the quad core LGA 1155 i7s in highly threaded performance with this *mod* and beat them without it.It might actually be one of the best things around if used properly.[/citation]
javascript: void(0);
Is this something that's confirmed or just BS? i believe we all want the source of this statement.
 
[citation][nom]idroid[/nom]Is this something that's confirmed or just BS? i believe we all want the source of this statement.[/citation]

http://techreport.com/articles.x/21865

That was with mere thread-scheduling improvement rather than disabling of cores, had a frequency disadvantage to the normal core configuration, and it stil managed a 10-20% improvement (usually closer to 20%). The CPU/NB frequency controls, among other things, the L3 cache frequency (defaults to 2.2GHz). Increasing it to the CPU frequency would make it a full-speed cache and that helps a lot in performance per Hz of the CPU frequency too. This is not BS.

Also, with a 30-40% power consumption drop (you can't expect halving the core count to not reduce power consumption greatly, can you), overclocking headroom increases (and it was already pretty high). These CPUs can probably break 6GHz on air. Admittedly, that is an estimation rather than provable fact right now, but my estimations are usually pretty good and considering that the current FX 81xx CPUs can already break 5GHz on lower end Air coolers in the same configuration (they can reach for 4.5GHz on their stock coolers with all eight of their cores, what do you expect), chances are good that these CPUs which have higher thermal headroom at a given frequency (and don't need as high voltages to reach said frequencies) can hit a higher frequency.

Another thing to keep in mind is that they already have higher performance per Hz than the FX-81xx CPUs did.
 

Jarmo

Distinguished
Jan 28, 2009
136
0
18,680
If I were putting together a new machine soon, I'd probably pick the 4x4GHz @95W one.
Too bad there still doesn't seem to be a compelling reason to dump my near 4 yrs old Phenom II. 4 cores at 3GHz is still good enough.

Damn, 4 yrs and if the GTX280 hadn't been such a dud, it could still run new games at max settings as originally purchased. Progress really isn't racing on like it used to...
 

tomfreak

Distinguished
May 18, 2011
1,334
0
19,280
And why does the lowend APU are not dual graphics capable? This is the area where people are budget constrain and would love to have better performance/price.
 

Hazle

Distinguished
crossing my fingers. if the 4320 performs close to an IB i5 at the same clock, then i got nothing to complain. basically just as long as we can finally put the Phenom II x4 955BE/960T to rest, and not rely on recommending it for the best AMD gaming build anymore. they're getting a little too old to recommend nowadays as great as they are.
 

tomfreak

Distinguished
May 18, 2011
1,334
0
19,280
[citation][nom]Jarmo[/nom]If I were putting together a new machine soon, I'd probably pick the 4x4GHz @95W one.Too bad there still doesn't seem to be a compelling reason to dump my near 4 yrs old Phenom II. 4 cores at 3GHz is still good enough.Damn, 4 yrs and if the GTX280 hadn't been such a dud, it could still run new games at max settings as originally purchased. Progress really isn't racing on like it used to...[/citation]it isnt, back then u can replace a $50 CPU and still get a decent performance jump.
 


AMD probably beat Phenom II significantly, but I highly doubt that they can fight with the i5s without the mods that I've mentioned. Chances are that it will be one or two more generations before they can do without them and still meet or beat the Sandy and Ivy Bridge i5s and even then, the Ivy i5s can have their IHS removed and the crap paste switched out with better paste (making them actually better than Sandy at overclocking) and give them a run for the money. Disabling one core per module and upping that CPU/NB frequency is a necessity to truly meet or beat the i5s with overclocking, at least for now.

EDIT: Even with the mods that I've mentioned, AMD needs a higher frequency to compete with Intel. There's no trick that I'mn aware of left to get AMD on-par with Intel in performance per Hz right now. Maybe getting top-end memory as well as the mod and overclocking the CPU frequency and the CPU/NB frequency will help, but even then, it's unlikely that it will match Intel. Maybe with another generation, AMD can do that, but if they can keep frequencies this high, they might not need to match Intel in CPU performance per Hz per core anyway.
 


Try Sleeping Dogs and such at maximum settings for 1080p and we'll see if it really can handle today's games with maxed out settings (it can't even come close because even no AA would probably be too much for it, let alone maxed out AA and other such quality enhancements). Heck, many games today are far too much for any graphics setup when you play them on their absolute maximum settings.
 

tomfreak

Distinguished
May 18, 2011
1,334
0
19,280
u can blame that on console ports. Back then in PentiumII/III/AthlonXP to P4Athlon64 era. a gaming rig would only last at most 2-3yrs b4 failing to max everything. Now u can do it with a 4-5yrs old rig.
 


They'll probably work like the current FX CPUs do in that regard, but it's usually recommended that you use them in an AM3+ socket.
 
Wow Blaze you got a red arrow troll on your tail. Did you piss someone off?

For everyone else, those "mods" he's talking about, ~I~ can confirm they work as advertised. You can further enhance the speed with AMD PSCheck and thread affinity (windows).

Time to explain Window's schizophrenic task scheduling engine. Threads are dynamically scheduled to whatever resource is free, those threads are forcibly interrupted several times per second to allow other tasks to get a chance at running. That is how task swapping works, your PC is constantly switching between a few hundred threads even if the thread is in wait state and only gets 1 cycle. Now what does windows do with your thread when it's time to switch it back? You'd figure it would put it back where it got it but no, windows will assign the thread to whichever "core" is least free, so you thread can jump around between all cores on the CPU.

That is why "turbo boost" rarely seems to work, rather then one core at 100% and three at 0% you have four cores at 25%. We use processor affinity to force the thread to stay on one or two CPU's that we chose, this way we can ensure the other modules will clock down / turn off to conserve TDP. That in turn free's up power to boost the two modules we are using to higher frequencies.

My most successful attempt as with my laptop AMD A8-3550MX. It's a 2.0 Ghz four core APU (K10.5), by down clocking three cores to 800Mhz I can get one core to run stably at 3.0Ghz. I can even get two cores to run stably at 3.0Ghz but as the CPU shares the same cooling system as the GPU, I can't be doing any gaming. You can use this same methodology to get any AMD CPU to run better, including bulldozer. Disabling every odd numbered core (1.3.5.7) yields a four "core" CPU that is a bit slim on the ALU's though doesn't have any arbitration issues with cache / instruction decoder. You could also only disable the last odd core leaving you with three full modules and one dedicated core module. Run like normally and when your going to lead a lightly threaded single task program, schedule it on the last core to get maximum performance.

Now I haven't played with PD, I don't know how effective their tweaks were with the instruction decoder / cacheing system. At worst you can do the same as you do with BD to pump out more performance.
 

mmstick

Honorable
Apr 5, 2012
56
0
10,630
If you seriously think disabling one of the ALUs in a module is a good idea, I think you seriously need help. Disabling an ALU will effectively halve your total integer ouput. Much as the url in abitom's post demonstrates, performance is lost, not gained. There are plenty of benchmarks out there that put the 'disabling half your ALUs yields performance increase' myth to rest.

AMD didn't probably beat Phenom II significantly, it DID beat Phenom II significantly, as much as many people would like to disagree, the truth can't be changed. I own three Phenom II X6s clocked to 3.5-3.8GHz each. They are all 3x slower than my FX-8120 @ 4Ghz with stock voltage at x264 encoding, they are twice as slow in compiling source code in Linux using make -j9 than Phenom II X6s with make -j7, and if you compile your own source code with -march=bdver1 you are looking at potentially 2-3x faster computations in these programs, though depends on what kind of calculations this program is doing.

There are two primary problems with the current generation of FX. The first is that most compilers software makers are using do not support AMD at all, if you use an AMD processor, it might as well be running in Core 2 Duo mode, instruction set-wise. There are a few cases where you may find someone is actually properly compiling their software, aka x264 encoder, but you'll find all benchmark software not on Linux is compiled with Intel C+ which cripples performance to abnormally low results. The second is the cache issue, which hopefully has been addressed with Piledriver. I don't see why AMD wouldn't fix it since they are aware of its existence.

But finally, a lot of AMDs problems is Windows, no joke. Install Linux and you'll never have problems, even old Phenom IIs may receive double the performance due to software actually being compiled correctly. Windows programmers just suck at programming.

I foresee great success with Piledriver.
 
Yet another page detailing performance comparison between 2m/4c, 4m/4c and 4m/8c

http://www.xtremesystems.org/forum [...] eaded-Perf

Interesting, cos of the number of benchies shown

Not really relative. Those are all multi-threaded tests and thus gain from having as many CPU targets as possible. Disabling cores would generally be a bad idea in those scenarios.

What me and Blaze are talking about is single threaded or lightly threaded (one main thread and a few secondary threads) applications that do not make use of more then two cores, aka "Games". The above tweaks allow you to give maximum resources to a single thread while also allowing you to clock it much higher then what would normally be allowed by the TDP window. It sacrificed multi-threaded performance for a large increase in single threaded performance, which is the one area that everyone is complaining about. BD is already good at heavily threaded apps, people are angry because they can't get higher "fps" in some program and want to know why.

If you seriously think disabling one of the ALUs in a module is a good idea, I think you seriously need help. Disabling an ALU will effectively halve your total integer ouput. Much as the url in abitom's post demonstrates, performance is lost, not gained. There are plenty of benchmarks out there that put the 'disabling half your ALUs yields performance increase' myth to rest.

This makes absolutely no sense ... you can't "disable ALUs" ... only cores. Each BD core has two ALU's and two AGUs inside it, those units process instructions fed to a single register stack. Phenom II / Intel Core each have three generic ALU's (combined ALU / AGU) per core, the Phenom II feeds one register stack per group while the Core with HT can feed three with two register stacks. CPU performance does not increase linearly with increased ALUs as x86 is not a multiprocessing friendly instruction set, going from two to three ALU's provides a small if any performance increase, especially if those ALU's also pull AGU duty when required. Intel got around this by using HT to allow two streamss of instructions to operate on three ALU's and thus each stream would get 1.5 ALU's worth of performance, which is more efficient then trying to keep three ALU's busy with one instruction stream. AMD did it's own thing and decided to make two ALU's per instruction stream and make a real core vs an abstracted core. Back end L2 caching and instruction decoding arbitration limit the viability of their design, the two cores fight each other for access to the shared L2 cache and then only when their not waiting for the decoder to give them instructions.

By disabling one "core" per module you effectively have 50% of the theoretical processing power yet you also remove a limiter to the first core. So while total theoretical processing power goes down, actual realized processing power goes up. One 8xxx CPU has eight "cores" for a total of 16 ALU and 16 AGUs, there are very few consumer orientated tasks that require that much horizontal computing power, yet there are many tasks that could benefit from four "cores". Thus you can manipulate your processing power to best suite your needs, as wide or as narrow as you desire.

Welcome to my work, try not to go crazy...
 

ojas

Distinguished
Feb 25, 2011
2,924
0
20,810
Well, i won't claim to know the "truth" about bulldozer's performance, since i don't own one.

However, i'd like to say this: At least compare stuff at the same clock rate. If you want to compare a bulldozer cpu to a previous generation AMD cpu, at least use the same clock rate and core/module count. Don't say stuff like "OMG bulldozer at 4GHz beats a Phenom II at 3.5 GHz!!!!".

And see...from what some of you are saying, bulldozer is probably an enthusiast friendly CPU, in the sense that you can tweak it and play around with it to "unlock" more performance. But the point is that people shouldn't have to do that. After all, most people wouldn't be willing to invest that much time in research, a lot won't be capable of pulling it off, and for others time may equal money.

So AMD's stuff has to work out of the box. That's what piledriver has to do.

Anyway, i think this stuff's going to launch after windows 8 for a good first impression. After all, it would suck if AMD has to say "it'll do better in win 8" again.
 

mmstick

Honorable
Apr 5, 2012
56
0
10,630


No, you should always compare the peak performance of both processors. Processors are not all created equally. FX's design allows higher frequencies at lower voltages, there should be no amount of handicaps given in comparing.

Also, it's not that hard. If someone is too lazy to do a little bit of research it is their fault. And no, if they tried they would be capable, 4Ghz is a VERY low frequency that all FX processors are capable of. They should have been binned at 4Ghz at start, to be honest, as you can even underclock below stock voltage and 4GHz is still prime stable.

As well, if someone can't be harassed to overclock their processor, I don't think they actually care about making their processor 30% faster, as apparently having a fast processor is not productive for them. It honestly only takes a few clicks with AMD OverDrive or a series of key presses in BIOS, I only need 30 seconds to configure a FX processor for maximum performance. Once you have done this, you have the overclocking profile for life. This saves time AND money for those people that are affected by processor frequencies, like me.
 
Status
Not open for further replies.