AMD CPU speculation... and expert conjecture

Page 465 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Status
Not open for further replies.
The more gamers keep finding out that the Xbox one can only do 720P when the PS4 keeps pumping 1080P for less the more Microsoft is going to lose.

99% of gamers don't care about that. Hell it wouldn't make a difference anyway, HDTV's are 1360/1366x768 native anyway. All but the most expensive use an internal downscaler to convert the 1080p signal into its native resolution.

Console gamers only care about what titles are available on which consoles, it's why console manufactures fight over exclusive releases whenever possible. Nintendo doesn't sell units before their "superior" they sell them because there is a huge consumer base wanting to play Zelda, Metroid and Mario. With producers now pushing to make games across multiple platforms, the decision between them comes down to what home theater system you want and Sony is just better at that then MS. PS3 was the best bluray player ever made.
 

juggernautxtr

Honorable
Dec 21, 2013
101
0
10,680
consoles bah humbug!!!!!! worthless overpriced blue ray player.... bought a ps3 and built my machine around the same time, bought them bullet storm for ps 3 and for my pc, they saw it on pc and i hardly got touch my computer till they went to bed.
 


I liked a few games, but mostly I use my PS3 to watch Bluray's / DVD's and play the old PSX games like Alundra.

Best Blu-ray player ever made.
 

jdwii

Splendid


Obviously you're not getting those numbers from someplace like IGN/Or major web sites that review games and read the comments. Console gamers really do care about graphics at times they may say they don't but i know they do..And that is the reason why the Xbox one is losing compared to the PS4,Even in the United states which amazes me.
 


Stop and think a minute. Most console "gamers" don't read reviews, don't visit forums and don't even care. They couldn't tell you the difference between 720 or 1080, have no idea who AMD or Nvidia are and still associate Intel with the blue pipe guys. They aren't enthusiasts and have zero care what happens in the computer world, google is a place they go to find out the latest cat pictures or sports scores. These are the guys who buy a new version of Madden every single year. They would use RCA cables to hook up their system if given the option, otherwise they are bribing me with pizza and beer to come over and setup their new console or home theater.

They don't care because they don't know.

PS4 is selling because it has a more attractive game library and is a more recognized media platform. The aforementioned "gamers" remember the PS3 as the all-in-one DVD/BD player and associate the PS4 with that memory. That btw is why MS is trying so hard to market XBONE as a media platform, it appeals to a much larger crowd and is remembered better then technical details.
 
Uh, it does though. Having to shuffle everything over the PCIe bus shouldn't make it slower than doing everything on the CPU.

Yeah, an APU with something like FX 8350 CPU with 280x GPU would be a lot faster than FX 8350 dCPU and 280x dGPU, but it's still going to be faster than FX 8350 without anything on the CPU.

You're also assuming that the only bus to connect GPU and CPU that will ever exist for HEDT market is PCIe 3.0.

PCIe 4.0 is going to offer almost 32GB/s bandwidth in a 16x slot. DDR3 2133 is around 17GB/s.

Here's the issue with current GPU design: as a result of the separate memory pools, you have to physically copy all the GPU data across the PCI-E bus, which is VERY time consuming. Hence the VERY large and VERY fast amounts of VRAM on GPU's now: You do NOT want to do multiple reads to main memory if you can avoid it. Any dGPU is going to have this bottleneck. This is the SAME EXACT REASON why CPU's have cache: even getting access the main memory is a significant system bottleneck. All forms of onboard cache are simply ways to avoid the delay it takes to stuff data over a bus.

And the point that was missed here: OpenCL works with any form of Intel, AMD, or NVIDIA GPU. HSA works only with AMD. Which one is more likely to be implemented by developers running on strict deadlines? I'm not arguing HSA isn't faster (and, like the FX lineup, in some subset of tasks, significantly faster), I'm arguing its far less likely to be adopted because its currently tied to a single vendor.

A vendor which saw its GPU shipments drop 10% last quarter, counting sales of APU's:

http://fudzilla.com/home/item/33997-nvidia-intel-gain-gpu-market-share

Most worrying is their collapse in the laptop market, which is where AMD should be selling these like hotcakes.
 


As I've said before, due to architectural differences, you need to rip apart the graphical backend anyway. The reason the 360 was the lead development platform last time was due to the 9 month headstart, rather then DirectX.

Fact is, the XB1 is overpriced, and the sales numbers prove it. Packaging the kinnect, which everyone has basically abandoned at this point, was a major, major mistake. As long as Sony itself doesn't collapse (which is QUITE possible given its finances), the PS4's won this round.

Changing topics:

Again, Every major etailer and reteailer are price gouging people to make that extra buck due to the miners.

Supply and demand. There is an over-demand and lack of supply, so prices are going to go up. Gouging is simply evidence the free market is working as intended; its the natural outcome of supply shortages, just like inflation is the result of economic growth, and so on and so forth. Complain all you want, but that's how our economy is DESIGNED to work.
 

the apus' dt share went up due to the introduction of the fm2+ motherboards. those sales were almost all "retail".
in laptops, amd's share went down due to them not bundling radeons with haswell laptops as much as geforce (kepler) gpus. in both laptop apus, chipsets and laptop and discreet gpus, amd's share drop is due to them being late with new apus (kaveri is mia until later) and no new laptop discreet gpus. they were also late with introducing the new rebadged gpus in both laptop and desktop. those were "bulk" losses.
i think amd will see a spike in high end gfx due to mining craze this quarter. apu shipments will also go up (in dt) thanks to kaveri.

about the new exc link from jdwii: i really, really hope that beema and mullins and their successors don't come back to glofo. glofo has a proven track record of failures with low power apus. unless tsmc allocates their 20nm to others. that'd be unfortunate for amd. hopefully gm107 being 28nm is only for nvidia.
 

juanrga

Distinguished
BANNED
Mar 19, 2013
5,278
0
17,790


I merely applied basic logic laws. You did make a generic statement about HSA, which is quoted above, "doesn't add significant performance benefits over what OpenCL already offers." To show that your claim is wrong I did only need to quote one example invalidating it.

There are other examples where HSA shows benefits over plain OCL. They have been presented during talks of HSA. For instance the face detection HSA algorithm mentioned in my article about Kaveri. I only give the overall performance of HSA, but AMD offered a split into the contribution made only by the CPU, only by the GPU (OCL) and the overall HSA (CPU+GPU).
 
AMD’s Radeon Dual Graphics: Looking at Laptops
Dual Graphics is finally not a drawback...
http://semiaccurate.com/2014/02/21/amds-radeon-dual-graphics-looking-laptops/
mm nice... can't wait to see how well mobile kaveris perform in dual gfx...

glofo please don't fab low power x86 apus and socs! pretty please! make those arm socs instead!
 


Notice how the guy ignored the fact that the improvements for the Dual Graphics solution is not working in DX9 (dual graphics is NOT XFire, remember that). And even the frame pacing was said to be DX11 only (for now... right...).

For all the love I have for APUs, I have to agree with palladin that in notebook space, adding a dGPU to it is very stupid. APUs are for small form factors, but AMD is not doing anything to push that approach. For the added price of a dGPU, they could even use full sized DIMMs! I wonder how that would work, haha.

And gamerk, it is very clear that HSA won't make things cheaper, but the idea behind HSA from AMD is to make it as seamless as possible to keep the overhead (price and code) low. That's why we need to sit and wait till more programs come out with the HSA sticker on them. If they don't come out, then AMD couldn't deliver on their promises (leaving Intel conspiracy theories on the side, haha). The benefits from it are very real, from what LibreOffice showed, so if Adobe and Oracle push their HSA software soon, the rest of the companies might like HSA as a viable development addition.

Cheers!
 

logainofhades

Titan
Moderator


Vishera AM3+ FX chips till 2015??? That doesn't give me a very good feeling at all. Hopefully they don't cripple excavator for FM2+. :-/
 

he mentoned those in the desktop's dual gfx article
http://semiaccurate.com/2014/02/20/amds-radeon-dual-graphics-desktop/
amd may never address dx 9 support or delay it a lot, considering the age.

i disagree. laptops (higher end ones with dgfx) can benefit a lot from discreet gfx with apus if handled properly. dual gfx, gpu switching and gpu power-down - these three need to work smoothly. users should be able to choose between gpus or dual gfx. for example, the igpu works to save power while the dgfx or dual gfx works in "performance" mode or gaming on external display. upgradable dgfx would be amazing but i don't think laptop oems will use it for amd laptops. these need a well-working software component which amd seems to have failed to deliver since 2011.

 

juanrga

Distinguished
BANNED
Mar 19, 2013
5,278
0
17,790


The official roadmap for 2015 was given in this thread several months ago. Rumors say that FX ends in 2015 and AMD will migrate to all APU for 2016: codename Basilisk.

If you pay attention to roadmap, Carrizo is 65W APU (Kaveri is 95W). I don't expect huge performance gains over Kaveri due to the reduction in TDP.
 

truegenius

Distinguished
BANNED


:kaola:

they do care about resolution
even mobiles are having 1080p display and moving to 2k displays (which is useless, we don't need insane ppi), thus resolution matters

even i don't like to play games on my PC at 720p on my 768p monitor because i notice noticeable difference in quality
 

8350rocks

Distinguished


No...

PCI-E 3.0 had no effect on MB prices, manufacturers increased the price with a new socket introduction, just as they always do.

They will be high priced because the raw computing power is exceptionally higher.

@juanrga - are you really arguing that the TITAN HPC would be any more effective at exascale computing with APUs? FLOPS are FLOPS. That same number of computations can be done regardless of format of the hardware...why? Because FLOPS is a measured performance metric, and is registered by calculating FLoating point Operations Per Second...meaning the results can be calculated...not speculated. Your argument holds no water, and there are no APUs in TITAN. Now what?
 

logainofhades

Titan
Moderator


Last map I saw said 2014, that FX would end, not 2015. PCI-E 3.0 didn't change prices much at all. My P67 extreme4 was about the same price as my Z77 extreme4. The Z87 extreme4 is priced about what my P67 Extreme4 was back when I bought it.
 

blackkstar

Honorable
Sep 30, 2012
468
0
10,780
@gamerK, the point of HSA is to avoid having to copy memory.

https://en.wikipedia.org/wiki/Graphics_Core_Next#Architecture

GCN can address unified memory. It exists in GCN since 7970 came out but it's disabled. The point I'm getting at is while having the GPU read from an address over PCIe bus is very far from optimal, it has the potential to still be faster overall given that the amount of throughput on the GPU can offset the latency.

That is the point I have been getting at. Would it work for real time applications? Probably not too well. Would it work for massive HPC tasks or large scientific calculations which can take days or even weeks? Absolutely. Could it be refined or be on a different bus or something to make it more effective in real time situations? Absolutely.

As I've mentioned before, HSA needs big wins like this to really entrench itself. Imagine if CUDA's only effective uses were encoding video and only consumer oriented things like that. It would not have been as big at all. Also imagine if CUDA only worked on GPUs as powerful as mid range Nvidia, say GTX 750 range. That's essentially what AMD is doing by limiting themselves strictly to APUs. Last I checked, Quardo/Tesla/etc were very lucrative markets for Nvidia.
 

jdwii

Splendid




Yeah i thought i was the only one who thought that statement was a bit off. I read and talk to console fans they argue all the time about what system has better graphics even if they don't understand why although 1080P vs 720P is very easy to understand. If console gamers did not care about graphics they would of never upgraded from their old consoles. For example if an xbox user owns a Xbox 1 and a Xbox 360 what version of titanfall are they going to get, its the same game but 1 has better graphics(which IMHO still look like crap).
Actually console fans even buy game informer and what not and they do talk about 720P vs 1080P mess. I remember hearing from some nintendo fan bragging about how their system actually has more 1080P 60FPS games on it.
 

jdwii

Splendid


I don't know he even keeps talking about Intel discrete cards? I never knew they made gaming cards unless he is just talking about the server place which is not 100% of the market. PC gaming is on the rise..With 50% of the developers wanting to make games for that platform and indie gaming becoming bigger and bigger.
 

Cazalan

Distinguished
Sep 4, 2011
2,672
0
20,810


I see you're still confused by Intel's HPC plans. They are not making APUs for HPC. And Phi is only 1/2 of Intel's HPC solution. They have "big core" CPUs, and "small core" CPUs. Note there is no discontinuation of Xeon. It remains an integral part which they will continue to improve upon.

image.php


And they will continue to use HSSIO (High Speed Serial IO) lanes as an interface. These can be configured as PCIe or QPI for local connectivity or as Infiniband as a host fabric interface. There is no getting around that the thousands of CPUs have to get linked together to work on large data sets.

The biggest change for HPC is the local memory bandwidths that HMC/HBM provides. Effectively making the RAM as fast as L3 cache and at much lower power. The rest is coming from die shrinks and standard cell libraries better tuned for low power. Eventually logic dies will be stacked but that looks to require new materials to compensate for the thermals.
 


I am going to wait and see. There is honestly no real reason for any console to exist if we look at the fact that PCs have had more power than current gen consoles for 2 years.

I think it will be an interesting battle but I only am glad that now they all use x86 hardware thus making PC porting easier and that also means PC versions wont have nearly as many bugs to work out.



I built a HTPC from extra parts (Asus M5a78 mobo, Athlon 2 X2, 4GB DDR3 XMS, HD5450) and threw in a Blu-Ray. Not only did it only cost about $300 I also can do more than just watch discs.

Plus I have the TV port all the sound from it and the 360 into a decent 2.1 system hooked up to the PC.



I like the TBD GFLOPS.

And of course this is only one side. It is all a different sector as well and people are saying Intel will give up it's discrete GPUs, well KL is not a GPU. It is based on Larrabee which was to be Intels dGPU but they canned that and now it is just a very fast coprocessor for HPC solutions.

If they could get RAM to be as fast as L3 it would be a plus as I can see Intel making it act like L3 which it stores commands for faster access. Of course L3 might also speed up and I doubt RAM will ever be as fast as cache but it still wouldn't hurt the HPC space to have that speed.

Of course we are talking about years down the road. We wont see this till we are old fogies.
 

Cazalan

Distinguished
Sep 4, 2011
2,672
0
20,810


A 16 lane stack of HMC 1.0 spec is 20-30GB/s in each direction. SB-E gets about 32GB/s write and 42GB/s read from L3.

It's not quite L3 cache speed but in the ball park. That is with 10-15Gb/s SerDes channels. Which can be doubled again with 28Gb/s SerDes, or with more links.
 
Status
Not open for further replies.