Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
AMD Confirms RX 480 At $199 USD, Other APU & Polaris Announcements
- Thread starter lowendlinux
- Start date
- Sort by reaction score
Source? Please provide data that suggests the R9 Nano has better perf/watt than Maxwell.
It doesn't...
Its close, but it only makes it competitive and certainly not a "star".
About efficiency of Nano in 2016 and in 2017 we have situation like this:
The Dawn of War 3 is actually one of best optimized games I have seen lately for both architectures(both companies). Look at performance of GTX 780 Ti. Look at Kepler in general. Absolutely tremendously optimized, almost on par with Doom in Vulkan.
Im sure you guys are intelligent enough to calculate the efficiency of the GTX 980 Ti(240W) vs R9 Nano(185W) based on those benchmarks, from current suite of apps.
As for compute: there is enough compute benchmarks in the wild that I don't even bother to post them...
Why do you claim something first: that Nvidia has higher efficiency than AMD for very long time, I reiterate that is not true, and then you ask me to prove it. You prove it first. Based on current suite of benchmarks. Based on todays performance of GPUs when software matured. Otherwise you show plain stupidity if you base your overlook on hardware just based on launch reviews. Software matures, as we could see in previous debate "Nvidia CUDA vs AMD OpenCL performance in Blender".
Of course I used benchmarks that are showing AMD in good light. There is a lot of benchmarks that are Nvidia optimized, and showing worse efficiency. But claiming that Nvidia has higher efficiency for very long time(actually they have higher efficiency for only 14/16 nm round, R9 390X was on par with GTX 980 Ti in compute, both used similar amounts of power, yet nobody here cares about it...).
The Dawn of War 3 is actually one of best optimized games I have seen lately for both architectures(both companies). Look at performance of GTX 780 Ti. Look at Kepler in general. Absolutely tremendously optimized, almost on par with Doom in Vulkan.
Im sure you guys are intelligent enough to calculate the efficiency of the GTX 980 Ti(240W) vs R9 Nano(185W) based on those benchmarks, from current suite of apps.
As for compute: there is enough compute benchmarks in the wild that I don't even bother to post them...
Why do you claim something first: that Nvidia has higher efficiency than AMD for very long time, I reiterate that is not true, and then you ask me to prove it. You prove it first. Based on current suite of benchmarks. Based on todays performance of GPUs when software matured. Otherwise you show plain stupidity if you base your overlook on hardware just based on launch reviews. Software matures, as we could see in previous debate "Nvidia CUDA vs AMD OpenCL performance in Blender".
Of course I used benchmarks that are showing AMD in good light. There is a lot of benchmarks that are Nvidia optimized, and showing worse efficiency. But claiming that Nvidia has higher efficiency for very long time(actually they have higher efficiency for only 14/16 nm round, R9 390X was on par with GTX 980 Ti in compute, both used similar amounts of power, yet nobody here cares about it...).
About efficiency of Nano in 2016 and in 2017 we have situation like this:
The Dawn of War 3 is actually one of best optimized games I have seen lately for both architectures(both companies). Look at performance of GTX 780 Ti. Look at Kepler in general. Absolutely tremendously optimized, almost on par with Doom in Vulkan.
Im sure you guys are intelligent enough to calculate the efficiency of the GTX 980 Ti(240W) vs R9 Nano(185W) based on those benchmarks, from current suite of apps.
As for compute: there is enough compute benchmarks in the wild that I don't even bother to post them...
Why do you claim something first: that Nvidia has higher efficiency than AMD for very long time, I reiterate that is not true, and then you ask me to prove it. You prove it first. Based on current suite of benchmarks. Based on todays performance of GPUs when software matured. Otherwise you show plain stupidity if you base your overlook on hardware just based on launch reviews. Software matures, as we could see in previous debate "Nvidia CUDA vs AMD OpenCL performance in Blender".
Of course I used benchmarks that are showing AMD in good light. There is a lot of benchmarks that are Nvidia optimized, and showing worse efficiency. But claiming that Nvidia has higher efficiency for very long time(actually they have higher efficiency for only 14/16 nm round, R9 390X was on par with GTX 980 Ti in compute, both used similar amounts of power, yet nobody here cares about it...).
Its pretty easy to pick and choose benchmarks that prove your point. I provided data that summarized many cases, you hand selected cases that only prove your point. This is easy enough to do. I can find tests that show a GTX 980 is faster than the R9 Nano.
Here we have the power consumption of the GTX 980 is lower than the nano.
and a bonus compute result
See how easy that was.
Yeah I don't know how you can claim that the R9 Nano is the absolute star of the 28nm generation. Sure, it might be the most efficient AMD GPU of that generation, but as I said, NVIDIA has been clobbering AMD on overall perf/watt for 2 full generations now (Maxwell and Pascal are both generally more efficient than their AMD counterparts) as those charts from Stacc demonstrate.
So lets see something here:
Witcher 3 - 2015 game.
Far Cry 4 - 2014 game.
Project Cars - 2015 game.
In that period, Nano was on par or behind GTX 980.
Lets see my post.
Doom - 2016 game.
Sniper Elite 4 - 2017 game.
Warhammer 40000: Dawn of War 3 - 2017 game.
At this point R9 Nano is on par with GTX 980 Ti, and far ahead in front of GTX 980.
Koyoot is misleading everybody by picking latest benchmarks!
This is funny actually. Everybody completely missed most important part of my post, which was about software maturity.
Im done. Talking to you guys is like talking to a wall.
Witcher 3 - 2015 game.
Far Cry 4 - 2014 game.
Project Cars - 2015 game.
In that period, Nano was on par or behind GTX 980.
Lets see my post.
Doom - 2016 game.
Sniper Elite 4 - 2017 game.
Warhammer 40000: Dawn of War 3 - 2017 game.
At this point R9 Nano is on par with GTX 980 Ti, and far ahead in front of GTX 980.
Koyoot is misleading everybody by picking latest benchmarks!
This is funny actually. Everybody completely missed most important part of my post, which was about software maturity.
Im done. Talking to you guys is like talking to a wall.
You are the one who is missing the point. I made a general statement and provided generalized data that summarized the performance of various cards. You picked out very specific cases favorable to your point. My point is that I can play this game too.
For instance here is a brand new game that shows the GTX 980 beating a Nano.
LOL.
So you look at those benchmarks and you see nothing wrong?
Maybe you don't know about problems with Fiji drivers in some games at releases, which results in GPU performing on the same level as much slower GPUs?
This is the downside of HBM1 memory. Requires Driver optimization, to get all out of it.
Its absolutely hilarious that you look at those benchmarks of R9 Nano performing on par with RX 480 and think: "yes, I see nothing wrong with that".
So you look at those benchmarks and you see nothing wrong?
Maybe you don't know about problems with Fiji drivers in some games at releases, which results in GPU performing on the same level as much slower GPUs?
This is the downside of HBM1 memory. Requires Driver optimization, to get all out of it.
Its absolutely hilarious that you look at those benchmarks of R9 Nano performing on par with RX 480 and think: "yes, I see nothing wrong with that"
.LOL.
So you look at those benchmarks and you see nothing wrong?
Maybe you don't know about problems with Fiji drivers in some games at releases, which results in GPU performing on the same level as much slower GPUs?
This is the downside of HBM1 memory. Requires Driver optimization, to get all out of it.
Its absolutely hilarious that you look at those benchmarks of R9 Nano performing on par with RX 480 and think: "yes, I see nothing wrong with that"
I look at these benchmarks and think that if I owned an AMD card and wanted to play this game (especially at higher resolutions than shown here) I would probably be disappointed.
You are still missing the point. I don't really care about which card is better than which or for what reasons. What I do care about is that you use better metrics to back up your claims instead of hand picking results that are favorable to AMD and then calling their cards the best.
GDDR6 coming up soon (well, not so soon):
http://www.anandtech.com/show/11297...gddr6-memory-for-graphics-cards-in-early-2018
http://www.anandtech.com/show/11297...gddr6-memory-for-graphics-cards-in-early-2018
Well probably sooner than an updated Mac Pro.
from article though.
"...
keep in mind that while it took GDDR5 a relatively short amount of time to replace GDDR4 on high-end graphics cards in 2008 – 2009, it then took the memory standard years to replace GDDR3 on mainstream adapters. ..."
That extended life for GDDR3 did not have to complete with iGPUs competitively creeping up the low end of the market.
It won't be one years, but a 3-4 years is probably more so reselling old designs than. If HBMv2 completely fails and stalls on new design wins, then there would be longer transition ( it will be easier to keep GDDR5 prices high), but GDDR5 also has pricing pressure from a cohort too if HBMv2 can ramp on volume. Bulk VRAM ( > 12 GBs) will probably be cheaper with GDDR6 than through HBMv2 (and a v3 if it arrives).
HBM's trick of stacking RAM dies higher I don't think is linear in terms of costs. It saves implementation footprint, but that comes at a cost of implementation and TDP dissipation capacity. HBM as a cache probably scales way better than trying to match main RAM sizes from several years ago for high workload systems.
AMD will probably leverage this on some implementations but won't be first vendor out of the gate. They still have several of their "current generation" to get out the door before the end of the year. That means next generation stuff probably not in early 2018.
Yes. I do not get it why Intels processors are so expensive and a good old fashion competition will give us more cores at cheaper prices. I assume no one will complain to that. I also do not understand why Intel has let Nvidia run away with compute the way as it has happened. Perhaps I do not fully grasp why GPU is superior to CPU except for the price performance ratio. Now that AMD seem to have some good stuff in the CPU market, I hope we see some competition. I believe that if CPU cores were cheap the incitament to use GPU for compute would be far less.What do you mean by Core war, and Cheap Cores?
8 Cores you can buy right now for the price of 6 core.
GPUs offer a staggering amount of processing power. CPU performance is still measured in the hundreds of GFLOPs, while GPUs have passed the 10 TFLOPs mark. So, you're looking at something in the ballpark of 10-30x more processing power on the GPU side of things. CPUs will never be able to compete at any parallel processing task.
The important caveat here is that GPUs are only very good at a select number of simpler parallel compute tasks. Traditional x86 CPUs have the benefit of ubiquity and can handle more complex compute tasks.
I'd disagree about the simple versus complex part of this, GPUs can and do run extremely complex computing tasks. Do you have a specific example of a task that a GPU can't handle but a CPU can?
Sure, lets take photoshop for example. This seems like something that would be conducive to GPU computations since we are working with image data. Some filters are GPU accelerated, but very intensive processes such as photomerging or exporting photos still require the CPU.
GPU's tend to have much more limited data structures and very poor single threaded performance. These restrictions can make it much harder to code for. Essentially they are designed to do simple operations on a large number of pixels. Anything that can roughly be formatted in this way could potentially be accelerated by a GPU. They can accelerate video because they often have fixed function hardware designed for this purpose. Or another example is computing neural networks, which is essentially just a bunch of very simple operations that are executed in parallel to compute an output for a given input.
Okay sure, that's why I said parallel programming (i.e. doing stuff on pixels in a photo) rather than single-threaded programming (photo exports probably falls into this category, since it's generally writing stuff out to disk). My point was that the per-pixel operations in photo or video applications can still be extremely complex, it's just that the inherently parallel nature of that task makes it a good fit for a GPU.
GPUs are good at code without inter-dependent state and conditional flow structures. CPUs are better when control flow is complex, and parallelism is weak.
Think of a GPU as a collection of tiny, underpowered CPUs.