AMD Gains 4.8% Market Share In Q1 2016

H

HardOCP News

[H] News
Joined
Dec 31, 1969
Messages
0
For the last nine months, AMD has brought a singular focus back to GPUs through the Radeon Technologies Group. During that time, the company has made significant investments in hardware, marketing, and software for the graphics lineup leading to four straight quarters of market share growth.

According to Mercury Research, AMD gained three points of unit volume share in Q1 2016. The recent issuance of both Mercury Research and Jon Peddie Research market share data for Q2 2016 has confirmed AMD has posted the fourth consecutive quarter of desktop discrete GPU share growth, driven by AMD’s strongest quarter of channel GPU sales since 2015 and the commencement of shipping of the next generation Polaris GPUs.

In total discrete graphics, AMD gained 4.8 share points to 34.2% of market by unit volume (based on Mercury Research). In desktop discrete, a subset of total discrete, AMD saw a 7.3 share point increase, rising to 29.9% market share. This is another positive testament AMD’s strategy is working as the company drives forward towards “Vega” offerings for the enthusiast GPU market, which AMD expects to bring to market in 2017 to complement our current generation of “Polaris” products.

AMD believes it is well positioned to continue this trend in market share gains with the recently launched Radeon RX 480, 470, and 460 GPUs that bring leadership performance and features to the nearly 85% of enthusiasts who buy a GPU priced between $100 and $300.
 
I'm glad AMD is doing ok. Monopolies are bad for innovation and someone needs to keep Nvidia, and Intel on their toes.

I'm a tech slut, so I'll put my money out for whoever makes a good product. In the last however many years, it has been Intel and Nviia, but I'll gladly buy ATI/AMD if they have a prduct that fits my budget and requirements. I never got the whole fanboi thing.
 
pcgeekesq said:
Of course, it's unit volume, not revenue volume. I wouldn't be surprised if they actually lost share when measured in $$$.
Click to expand...


That is the combined desktop and notebook, notebook went up ~8% desktop whet up 4% (JPR, ~5% Mercury) over 4 quarters......

1% a quarter is not a turn around. Just market equalization of having channel volume. If we look back 4 q's ago there were not new desktop and notebook GPU's from AMD against nV's new offerings, so yeah they will lose quickly at that point).

notebooks picked up a bit, but this is not surprising as notebook segment fluctuation is great and retention rate is low (if ya want a portable device you are stuck with certain things).

I"m suspecting notebook turn around was AMD getting left over stock out the door by lower margins.
 
I don't see what Polaris had to do with sales given that this is Q1 and Q2 data. If anything I think this was merely the R9 300 series gaining some of the traction that it deserved at that point. During this sales period, Nvidia had no answer to the R9 380, 380x, 390, and 390x. These were all the fastest GPUs at their respective price points back then. Nvidia was winning at the sub-$150 price point (GTX 950) and the over $500 price point (980 and higher). And even the 980 was debatable.
 
I'm glad to see it. I run an Intel processor and an AMD graphics card. Competition is important to pressure the companies into improving their products, and to keep prices down.
 
  • Like
Reactions: muxr
like this
Graphics market share is not helping AMD at all. The Graphics and Computing Division lost $70 million in first quarter of 2016 and $81 million in second. AMD must be getting market share by simply selling at a loss or the computing side of the division is totally trashing any improved graphics profits. It like the joke: "You are losing money with every sale. Yes, but I am making up for it by selling in volume!" The other big factor was people waiting for NVidia's newest cards to be released. Can AMD's newest cards compete against NVidia's? We will know by the end of the year, but I expect NVidia to hammer AMD in both sales and profits. My prognostication will be here to see how right or wrong I was.
 
pcgeekesq said:
Of course, it's unit volume, not revenue volume. I wouldn't be surprised if they actually lost share when measured in $$$.
Click to expand...

Its not even that. Its estimated out from their financial numbers. But only Nvidia of the 3 reports it directly.

The Q1 share numbers in case anyone wonders:
AIB-1-PR.PNG

MWPRQ116-1.PNG


But this is quite old news from May and we are close to getting Q2 numbers...
 
Last edited:
Good on AMD. Im sure their are some cavets to nit pick but overall its a more positive sign than a negative one, hard for anyone to dispute unless they are trying to spin this.
 
Arachnotronic got a good financial view on it.

These numbers are misleading.

First of all, they only talk about units, not revenue share. Given that NVIDIA has scaled back its efforts to win non-gaming OEM designs (i.e. MacBook Pro, those Carrizo laptops with iGPU + Radeon, etc.), it's not hard for AMD to increase its share of the low-value dGPU market.

Last quarter, NVIDIA reported that revenue from its gaming GPU biz was up 14% quarter-over-quarter and 18% year-over-year. Last quarter, AMD reported that its total computing and graphics segment was down 5% quarter over quarter and up 15% year-over-year "driven by higher client notebook processor and graphics sales."

If AMD's total GPU +CPU biz was up 15% year-over-year and higher notebook APU sales were part of that growth, and if NVIDIA's gaming GPU sales were up 18% year-over-year, then this means that NVIDIA actually gained revenue share in the gaming dGPU market.

Next, AMD is guiding to a ~$200 million increase in revenue for Q3 over Q2 thanks to graphics/semicustom, while NVIDIA is guiding to a $250 million quarter-over-quarter increase in Q3 over Q2. Part of this could be Tegra, but it's probably mostly coming from dGPUs.

So, again in terms of revenue share, NVIDIA should actually gain share on AMD in the coming quarter.
Click to expand...
 
well either gain share or gain margins, it can be either one ;). I am expecting more margins not much change in marketshare either way in Desktop Discrete. If there is a change it will be in favor to AMD in like 1-2 base points.

Forgot to add in there increase in total amount of sales too, Q1 tends to be a weak quarter for total sales.
 
Last edited:
"leadership performance"
Thats a new one, any ideas?
 
AzixTGO said:
this talk about revenue share sounds strange. Why was this never brought up when AMD market volume share was falling? Why does nobody ever talk about revenue share? Maybe because its not as relevant? A company selling more units matters. What revenue they gain from it is just a function of what they chose for their margins. It matters that the company is actually selling more and their plans are actually moving forward. It also matters for the state of the industry in terms of which manufacturer gets attention by developers. For AMD it does not matter as much if nvidia has a higher revenue, for AMD it matters that they are progressing in their business plans.

Revenue share is for "fanboys" to fight each other over. "My company makes more money than yours."

I really should have bought some AMD shares. fml in the butt.
Click to expand...
Just because you sell a lot of units doesn't mean you will have the revenue to continue R&D to compete into the future. Limited revenue means limited R&D.
Regardless of how many units, if they don't start making some real money soon they will have to consider selling of their Graphics division to another company.

Many times companies prop up failing divisions in preparation for a sale.
With stocks going higher and the future looking a lot better, they can ask for more money in the sale, which will go to CPU R&D.
 
Why are you talking about revenue share, I think you might be using the wrong terms cause revenue share is profit sharing among shareholders......

total sales X gross margins = gross profits

So if you have more total sales but less margins, you can still have less profits, and vice versa or they can equalize each other.

Gaining market share at the expense of margins to the point at margins affect gross profits; in that type of scenario doesn't help. Unless you have other products with higher margins that will cover the potential "loss" of the less margin products.

AMD isn't in that situation right now. They don't have higher margin products.
 
Last edited:
WorldExclusive said:
Regardless of how many units, if they don't start making some real money soon they will have to consider selling of their Graphics division to another company.

Many times companies prop up failing divisions in preparation for a sale.
With stocks going higher and the future looking a lot better, they can ask for more money in the sale, which will go to CPU R&D.
Click to expand...

Are you saying AMD wants to sell off their graphics division? I was under the impression that their APU business (GPU+CPU) was a vital part of their income and future investment. Wouldn't selling RTG deprive AMD of that?
 
AzixTGO said:
Selling a lot of units is key to making money for more R&D. Revenue share doesn't even say they do that. Selling lots of units more directly does. Revenue share goes up or down even if it goes up for one company. If there is a general downturn in the market, 20% this quarter could be half of what it was last quarter even if its higher than the 15% revenue share from last quarter.

Propping up in this kind of industry has to be harder. They would be planning to prop up from several years ago...
Click to expand...

R&D for which areas? Mobile and Consoles? Mid-range GPUs?
OK that's where the most money is right?

AMD said they didn't want to be viewed as a budget brand, but their products because of subpar R&D can't command premium pricing.
 
SighTurtle said:
Are you saying AMD wants to sell off their graphics division? I was under the impression that their APU business (GPU+CPU) was a vital part of their income and future investment. Wouldn't selling RTG deprive AMD of that?
Click to expand...

They can sell the dGPU division. Mobile and CPUs are just fine for them.
 
AzixTGO said:
that's revenue sharing. Here I am talking about revenue % of industry value. Assuming that is what this person is talking about


you seem to think the more money you spend the better your R&D is. yet somehow AMDs products in the GPU arena have been more advanced. maybe they just are putting the R&D in different places. For pascal nvidia invested in higher clocks for example. For kepler and maxwell as well they had different strengths.
Click to expand...

Ok, yeah that is the wrong term to use, revenue shares, revenue share, revenue sharing, all the same stuff.

More advanced in what way? Surely not in overall GPU efficiency..... nV has shown gen after gen they can do more with less, maybe just better design choices of chip balancing of overall parts for the type of games coming out that time, timing is everything.

Surely not in power consumption which with Pascal and Polaris the difference is even more then last gen. % wise.
 
AzixTGO said:
How would that work? Is the GPU division not involved in their APUs? To sell the division that has held AMD up all this time would be retarded unless they are absolutely certain zen will hold up. But if they are certain, then they would have no reason to sell their GPU division because they would be making much more money.
Click to expand...


if they do something like that they keep the IP usage rights and any modifications of it there after. I don't think they will do that but we have seen AMD do some strange things in the past.
 
Advanced is the same type of term, nV is way ahead because they are seen as a better products do you agree? I don't see why anyone would be buying nV products if that is not what is perceived if when their products cost more at times.

I see no difference in advancements in silicon vs nV or AMD outside of one can do a hell of a lot more at lower power consumption, if you are talking about ACE's that's hogwash, what happens internally doesn't and shouldn't affect what happens externally. Doesn't matter what happens internally as long as that performance can be utilized in a meaningful and easy way with other end metrics being better too.

You can say Itanium was way ahead of its time and was a more advanced technology and it truly was, but at the end it couldn't be utilized to its fullest extent and that was its down fall.

You can say Bulldozer with its more cores is more advanced and works better with heavily multithreaded applications and it does do well, with the cost of higher power consumption, and failing in older applications that don't need more cores but more IPC, and that was its downfall. Which doesn't make it "more" advanced.
 
Last edited:
AzixTGO said:
"Efficiency" is a loaded term. If we were to talk about efficiency in bitcoin mining, who is more efficient? Exactly. Its software dependent. Dx11 efficiency while having less dx12 efficiency is neither here nor there. People think "efficiency" just applies in <dx11 as if its that simple. IMO nvidia is not more efficient even after castrating their hardware. With kepler they managed to use similar power to hawaii yet arent even comparable now. With maxwell the 980ti uses similar to the fury x yet loses even while overclocked depending on which api you use. Which in perspective shows AMD might be ahead. What we saw with thermi is what we would see if nvidia tried to do what AMD is doing I think.

I meant advanced in terms of what was in the silicon.
Click to expand...


Most people talk about power efficiency while gaming. The GTX 1060 uses 120W while being roughly 14% faster than the reference RX 480 at 165W, and still 5-10% faster than aftermarket cards that use up to 220W. THAT is efficient. You can name products where AMD cards are faster - Doom (Vulkan), Hitman, Ethereum mining (more efficient here as well). For literally everything else, Nvidia is faster in current gen products. And even in those three mentioned programs, only one shows AMD as more efficient.

What you're doing is called "white knighting." It's ironic that you criticized Nvidia fanboys earlier.
 
AzixTGO said:
"Efficiency" is a loaded term. If we were to talk about efficiency in bitcoin mining, who is more efficient? Exactly. Its software dependent. Dx11 efficiency while having less dx12 efficiency is neither here nor there. People think "efficiency" just applies in <dx11 as if its that simple. IMO nvidia is not more efficient even after castrating their hardware. With kepler they managed to use similar power to hawaii yet arent even comparable now. With maxwell the 980ti uses similar to the fury x yet loses even while overclocked depending on which api you use. Which in perspective shows AMD might be ahead. What we saw with thermi is what we would see if nvidia tried to do what AMD is doing I think.

I meant advanced in terms of what was in the silicon.
Click to expand...

Just a ton of ALUs, and not much else. That's why AMD cards are good for mining, but not so much gaming, and are always under-utilized.

I wouldn't want to be in AMD's shoes. Having to have 20-30% more ALUs just to match the competitor's general gaming performance means they're fucked in terms of margins due to bigger die sizes, or having to make big trade-offs in terms of design, ending up with unbalanced designs and bottlenecks in the pipeline.

AMD really should put more thought into getting in some of that GPGPU business and compete against Tesla and Xeon Phi. Much, much higher margins. Their designs should be well suited for those kinds of tasks.
 
Last edited:
AzixTGO said:
if you want to talk about architectures you talk about everything. Not just dx11 gaming. That's the basic point. How you utilize it is a shifting thing. What is actually there to utilize is more relevant to the argument about R&D and what a company was able to pull off in hardware.



1060 is not faster than a 480 for less power. Again, this is the problem. This only happens in dx11 and only in some games. Its all very dependent on what you run on it. 1060 is not faster in all other cases like you imply. We'll see though. Still more games to come.

I put fanboy in quotes. And it refers to all camps really. I dont call people fanboy because its stupid to call people that just because they disagree with you strongly. Deal with the arguments first. If its stupid, ignore it.



I really meant the inclusion of hardware scheduling, greater dx12 support and ACEs. At the very least they aren't behind.

The needing more cores is neither here nor there. Its more cores or higher clocks. They decide which route they want to take, tho I think more cores might be better if you can feed them. Ultimately you end up hitting a barrier where clocks have to go up too though. Regarding die size, it doesn't seem to always be that way. AMD seems to be able to put more cores in less space. Significantly less space. eg hawaii vs kepler even with the higher bus width and aces and hw scheduler, was smaller by far than kepler. The density might contribute to the lower clocks tho. Another comparison, tho not sure, is the 1060 and 480. !060 200mm2, 1280 cores. 480 232mm2 ... 2304 cores.

yes they need to get into GPGPU heavy. Thats another market they lagged behind in.
Click to expand...


And this is hogwash, cause end results dictate sales.

You want to go into the nuances of GPU's without understanding what is going on in what part, and then to come up with "more" advanced by looking at marketing slides, is a no no.

Tell me how the scheduling is done. tell me "greater DX12" support means to you in the case of AMD, nV has more support for DX 12 features (DX12.1 feature set), how are you quantifying them?

So let me get this straight it has more cores or higher clocks? Is either or? Why does Pascal Titan have more cores than the rx480 but can clock way above the rx480? How are you quantifying that?

"Ultimately you end up hitting a barrier" this is what you are saying, which is definitely correct, but the barrier is different for nV and AMD. So how do you quantify "more" advanced in that situation?

Do you know why they are able to get more cores into a smaller space (AMD)? There has been a lot of talk about this in past, and the benefits of that is actually diminishing returns on throughput. How are you quantifying more cores vs. more throughput?

There are many differences in both GPU architectures but to draw a base line because of A has this and B has this doesn't make one architecture "more" advanced then another, end results should be the only thing that matters.

And so far AMD seems to have an advantage with DX12 games that also do better for them in DX11, which out of the 5 games out thus far 3 of them are AMD sponsored they have a sizeable advantage in DX11 and 12, the others, one of them is nV sponsored is heavily nV favored, and the last one which has no sponsor works well on both and parity is similar to DX11 games.

PS the 1060 is faster w/ less power usage in DX11, and equal in DX12 with less power usage..... Either way ya look at it the 1060 is better, added to that it is more "efficient" in ALU throughput too, having less theoretical flops and still keeping up with the rx480 in DX12 and out matching it in DX11.
 
Last edited:
Yes! How do you evaluate 'dx12 support'?

Qualitatively? Quantitavely? Based on what criteria? It's a really weak argument.

You can always argue GCN has more flexibility with its ACE's and ability to effectively execute concurrent multi-engine at the CU level, but for this to translate into an inherent advantage vs competing products you need to establish the following correlation :

Two equally powerful (shaders) GPUs + DX12 = AMD victory, always.
 
Yes you can always argue those two points, but the first one is only part of the GPU, and we know ACE's aren't the end all of DX12, or LLAPI's. And definitely not the only part of a GPU that results in end performance and other metrics, which API's have very little to do with in that regard.

The second argument, is also moot cause right now AMD doesn't have anything the matches the upper tiers and the lower tiers are fairly equal in DX12 and loosing badly in DX11, I mean when ya got a brand new low end and mid range card matching the perf/watt of last generation low end, midrange and performance cards, that is nothing to write home about ;)
 
AzixTGO said:
1060 is not faster than a 480 for less power. Again, this is the problem. This only happens in dx11 and only in some games. Its all very dependent on what you run on it. 1060 is not faster in all other cases like you imply.
Click to expand...

The GTX 1060 is faster than the RX 480. Sure, they trade blows, and the RX 480 can come out on top in some cases, but on average the 1060 is noticeably faster. (I'll analyze this further below).

The GTX 1060 (most versions) tops out at roughly 120W under load. The reference 480s top out at roughly 165W under load. The aftermarket cards with raised power limits (MSI Gaming X, Sapphire Nitro+ OC, XFX Black) max at about 220W.

The GTX 1060 is faster. The GTX 1060 consumes less power. These are factual statements.

----------------------------------------

Now, as for the analysis. One of my favorite sources is TPU because they test a wide battery of games. Their latest results can be found HERE. After a full analysis of their results here's where the two cards stand. Note: I compared reference to reference and ignored the subject of that review (MSI 1060, green bar). Overall accounts for the overall in the performance summary, not ratio of wins.

-----

1600x900:
1060 wins = 12
480 wins = 3
Overall = 1060 by 12.8%

-----

1920x1080:
1060 wins = 11
480 wins = 4
Overall = 1060 by 11.4%

-----

2560x1440:
1060 wins = 11
480 wins = 4
Overall = 1060 by 10.1%

-----

3840x2160:
1060 wins = 12
480 wins = 3
Overall = 1060 by 5.4%

-----

Power Consumption (Typical Gaming):
1060 = 116W
480 = 163W

-----

Power Consumption (Peak Gaming):
1060 = 125W
480 = 167W

Conclusion - The GTX 1060 is shown to be >10% faster at the two most common gaming resolutions, while consuming 25-30% less power. If you're going to tell me that the GTX 1060 is not faster and more efficient, you'll need to provide some factual analysis to back up that claim. Otherwise, you may not like the word fanboy, but it would apply to the strategy shown in your posts.
 
trandoanhung1991 said:
Just a ton of ALUs, and not much else. That's why AMD cards are good for mining, but not so much gaming, and are always under-utilized.

I wouldn't want to be in AMD's shoes. Having to have 20-30% more ALUs just to match the competitor's general gaming performance means they're fucked in terms of margins due to bigger die sizes, or having to make big trade-offs in terms of design, ending up with unbalanced designs and bottlenecks in the pipeline.

AMD really should put more thought into getting in some of that GPGPU business and compete against Tesla and Xeon Phi. Much, much higher margins. Their designs should be well suited for those kinds of tasks.
Click to expand...

Also have to consider how the GPU is structured with SMs each containing double the ALU lanes of a CU and a geometry engine each.

Scalar units are also more expensive so that contributes to the difference.

Finally Polaris has a 10% size advantage thanks to glofo
razor1 said:
Yes you can always argue those two points, but the first one is only part of the GPU, and we know ACE's aren't the end all of DX12, or LLAPI's. And definitely not the only part of a GPU that results in end performance and other metrics, which API's have very little to do with in that regard.

The second argument, is also moot cause right now AMD doesn't have anything the matches the upper tiers and the lower tiers are fairly equal in DX12 and loosing badly in DX11, I mean when ya got a brand new low end and mid range card matching the perf/watt of last generation low end, midrange and performance cards, that is nothing to write home about ;)
Click to expand...

You can always unde4clock a 1080 or overclock a 980ti and compare to fiji
 
AzixTGO said:
Anyway, its not because of sub-par R&D that their products don't cost the moon. In fact the fury X did.

I think most of the arguments go towards GCN being more advanced than what nvidia offers. But I can't prove that. It just has more going for it on the hardware level beyond high clocks. You can't claim higher efficiency because that is dependent on the software run on it.



feature support? =/
Click to expand...

Yeah feature support and maxwell supports more features than Polaris, let alone Pascal.. Not that I would ever argue maxwell is better because it supports a higher FL. It just supports a higher FL.

I understand what you mean when you say GCN has more going for it than just high clocks, you could even say GCN is more complex and you would be right in a sense, but you just as easily argue that the maxwell and pascal are more 'elegant' solutions precisely because they don't need the complex auxiliary (control) logic.

At the end of the day how similarly priced cards perform relative to each other is entirely dependent on the balance of different loads and how that sits with the architectures bottlenecks.

What you can say pretty objectively is that it's harder to stall a maxwell gpu than a gcn gpu

Of course you claim higher efficiency.... I don't understand your point, yeah software optimization plays a role but at the end of the day what if I build a GPU that only achieves peak throughout under certain very specific and very restrictive conditions you can hardly argue it's more efficient just because of that one scenario
 
AzixTGO said:
Anyway, its not because of sub-par R&D that their products don't cost the moon. In fact the fury X did.

I think most of the arguments go towards GCN being more advanced than what nvidia offers. But I can't prove that. It just has more going for it on the hardware level beyond high clocks. You can't claim higher efficiency because that is dependent on the software run on it.



feature support? =/
Click to expand...

Most of the arguments yet you can't prove it? What argument?

GCN being more advanced, in what metric? Also, outside of coin mining, where does GCN offer higher efficiency than Maxwell, let alone Pascal?

So Pascal supports 12_1, GCN only 12_0 and yet it has better DX12 support?
 
See AzixTGO you just said 480 is more efficient than a 970 in DX12, but is that true? Is it really just one specific game, or this a general rule in DX12?

Having said that, besting the perf/w of a 970 isn't very impressive, you generally expect that when comparing GPUs across nodes
 
AzixTGO said:
all dx11 games. You didn't read the comment you are replying to. Hardware vs hardware goes beyond any software. The best comparison is the lowest level software available that takes the most advantage of each hardware. Dx11 pascal vs dx12 gcn is fine if pascal down is better suited to dx11

saying the 1060 is faster than the 480 is to state an absolute that is not warranted. You should say the 1060 is faster than the 480 in dx11. I know we like to pretend this means everything because its more popular, but my thinking is a bit broader on that topic. Best bet is best api vs best api.



ACEs and async compute is added functionality, added complexity. +1 to advancement. Regardless of api.

perf/watt in dx11. Qualify statements. It does no good to think too simply about efficiency. a 480 obliterates the perf/watt of a 970 in dx12.
Click to expand...


API doesn't matter. I am a gamer. I can't tell developers what API to use. What matters is gaming. And in gaming today, the 1060 is faster than the 480 while consuming significantly less power. That's factual. Your stance that AMD will be better in the future with DX12 is theoretical. It also won't matter to me. I'll be upgrading to the 1160/580 on release. By then, the DX12/Vulkan performance of the 1160/480 will be meaningless to me.

In the present, the 1060 is faster and more efficient. The reason that you're not debating this with any analysis is because you can't.
 
AzixTGO said:
Anyway, its not because of sub-par R&D that their products don't cost the moon. In fact the fury X did.

I think most of the arguments go towards GCN being more advanced than what nvidia offers. But I can't prove that. It just has more going for it on the hardware level beyond high clocks. You can't claim higher efficiency because that is dependent on the software run on it.



feature support? =/

feature level support? =/
Click to expand...

nV cards are DX12.1, so they are more advanced in this regard, but are those 12.1 extra features being utilized to their fullest extent, I don't think they are. This is what makes it hard to say "more" advanced, yeah the features are there but if they aren't really used, what is the benefits of them?

all dx11 games. You didn't read the comment you are replying to. Hardware vs hardware goes beyond any software. The best comparison is the lowest level software available that takes the most advantage of each hardware. Dx11 pascal vs dx12 gcn is fine if pascal down is better suited to dx11

saying the 1060 is faster than the 480 is to state an absolute that is not warranted. You should say the 1060 is faster than the 480 in dx11. I know we like to pretend this means everything because its more popular, but my thinking is a bit broader on that topic. Best bet is best api vs best api.
Click to expand...

With a list of applications that are short and most of them know to be sponsored by AMD, coming up with that conclusion is ....... a bit short sighted

ACEs and async compute is added functionality, added complexity. +1 to advancement. Regardless of api.
Click to expand...

No they don't, that is what you don't understand, and this is why I asked you how scheduling is done, ACE's with additional hardware (two other command processors, that we know of because AMD told us this) on AMD work in unision to get async and concurrency to work. nV hardware, Pascal in this talk, we don't know how the they modified the gigathread engine to get dynamic load balancing to work, yeah AMD still has an advantage here when talking about possible latency since they are finer grained in some regards doing certain things that require preemption, but again, with async only talks, we don't know what the difference is in utilization vs async benefits of different architectures as scientific outlook unless we analyze the program, and even then its hard to do since we don't have access to the code to fully understand what is going on. And scheduling on nV's hardware is hardware based, software based scheduling is a myth, but because of queue lengths on Maxwell, it does some of the scheduling in software when its over tasked. GCN's advantage on this in Pascal, is not there.

perf/watt in dx11. Qualify statements. It does no good to think too simply about efficiency. a 480 obliterates the perf/watt of a 970 in dx12.
Click to expand...

How about the gtx 980, cause that has even better perf/watt than the 970.... Last generation cards, that are 2 years old now? That is something that is great? Is that "more" advanced?
 
trandoanhung1991 said:
Most of the arguments yet you can't prove it? What argument?

GCN being more advanced, in what metric? Also, outside of coin mining, where does GCN offer higher efficiency than Maxwell, let alone Pascal?

So Pascal supports 12_1, GCN only 12_0 and yet it has better DX12 support?
Click to expand...


I wouldn't even put mining in there

RX 480 hashrate - Crypto Mining Blog

Now reading this, I'm not a miner but to me it seems like the 1070 has twice the hash rate of the rx 480/470, Unless I'm reading this wrong.
 
I guess it depends on the hashing algorithm because circular shifts are not part of the ISA so they're emulate with more instructions, afaik this is what made gcn defacto mining king.

The context switch latency is super low on gcn, but to my understanding there is no pixel level preemption - haven't found details though.

I guess if you consider the total time T from moment you send the halt command to the moment the new shader begins execution it's essentially the preemption latency + a context switch - although I wouldn't consider it a context switch unless the shader isn't new (was swapped out earlier) then gcn reduces the delay in reading/writing context but Pascal should be faster in preempting
 
AzixTGO said:
When it comes to feature levels GCN is ahead on 12. You can claim nvidia has 12_1 but they also had async compute. The point I was making was that they are on par and better in a lot of places. There is no basis for the R&D argument against GCN. Sure they could do more with more money (maybe) but they did well vs what nvidia did.

nvidia said they removed the hardware scheduler and gained efficiency, going to chose to listen to them over you.

proper async > load balancing. imo load balancing should be default. Shame on them for not having it till pascal.

the 980? I'd assume its a complete fail vs the Fury in dx12 since the regular fury touches or exceeds the 980ti in the benchmarks I saw. A strix fury can use around 200W according to techpowerup.
Click to expand...

Do you know what the difference between dynamic load balancing and async is? LOL, man lets get back to the basics cause one needs the other. You are not talking about async compute, you are talking about async shaders. Now rewrite your stuff and replace async compute with async shaders.

Async compute has been an nV feature since Fermi or Keplar. Async shaders which devs call that async compute at times works on all nV hardware as well Maxwell 2 and up. But Maxwell 2 can't do inter SMM, kernel changes after the first partitioning of the SMM, dynamic load balancing solves this problem, a feature AMD does have with all GCN (probably done in a different way). The problem wasn't that nV's cards lacked a feature, the problem was the ability to do concurrency in Maxwell 2 was hampered by the inability to change the partitioning once initially partitioned, unless SMM finished all work and was re-partitioned again.

Now do you understand what the feature change in Pascal's SMM's to accommodate this feature?

Now do you understand why scheduling has nothing to do with this? And shouldn't even be talked about when talking about async performance? If you don't want to learn about something, don't talk about it and post about like there are some facts to back you up, cause outside of marketing slides from AMD, what you have been saying, anyone that knows what they are talking about can see its fluff. AMD marketing took too much of the meat out of what is going in the background, to get anything meaning full out of what the real problem was with Maxwell 2 (which wasn't talking about by nV either so we can blame them too) and Pascal solved that problem.

This has been the one time AMD marketing actually worked! They actually fooled people into thinking their marketing team actually knew what they were talking about, but the fact is, they didn't!

its not a very specific scenario. Unless you want to say the same thing about nvidia cards under dx11. The gap mainly exists under dx11, this one specific scenario, so why claim they are more efficient.


its faster than the 970 in dx12. If you say its similar in efficiency in dx11, then its obviously more efficient than it in dx12. I dont know whats impressive vs what isn't when it comes to power consumption. I only said that because someone said it was just catching up. Its not just catching up, it has exceeded it. It ALL depends on what software is run on the GPU. If you want to say the GPU itself is less efficient, then clearly you are mistaken. The GPU itself is not less efficient, it can be less efficient or more efficient depending on what you run on it.
Click to expand...

You are comparing to a 2 year old card on a larger node that consumes more voltage per transistor! Polaris should be more efficient, but it barely edges out 2 year old cards lol.

If the discussion is about which architecture is more advanced, it really does matter what API you use. when you make statements like the 1060 is faster than the 480 without any qualifications, you do end up getting into "in what". Because that statement is factually incorrect.

its not the future. Its faster in almost every dx12 game to date and significantly faster in the one proper vulkan game.
Click to expand...

Please look at reviews, the GTX 1060 is within 0-5% of the rx 480 in all dx12 games yet it uses 20% less power, hmm.
 
Last edited:
AzixTGO said:
When it comes to feature levels GCN is ahead on 12. You can claim nvidia has 12_1 but they also had async compute. The point I was making was that they are on par and better in a lot of places. There is no basis for the R&D argument against GCN. Sure they could do more with more money (maybe) but they did well vs what nvidia did.

nvidia said they removed the hardware scheduler and gained efficiency, going to chose to listen to them over you.

proper async > load balancing. imo load balancing should be default. Shame on them for not having it till pascal.
Click to expand...

There's no such thing as proper async. It makes no sense to say load balancing < async; that's like saying engines are worse than motor vehicles, or steering wheels are worse than driving...

We need to clear up some misonceptions here.

DX12 basically says; here are three QUEUES exposed to the API. Graphics, Compute and Copy - they are independent. Have fun.

That's it. Asynchronous ~= independent.

The Graphics queue encompasses commands of all types, Compute is limited to compute and copy, Copy is just ... Copy.

DX12 does not impose a performance benefit from this. Maxwell has the hardware to do this, but to extract additional performance from it you have to program it explicitly to do so.

Load balancing is simply "dynamic assignment", resources can be partitioned among running shaders by the on-die logic, Maxwell had to do this at drawcall boundaries.

If GCN wasn't able to switch context quickly, thus achieve concurrent execution on a single CU, it would be in the exact same position as Pascal lol.
 
I always post these cause they specify clocks in overclock testing

r9a5pyI.png


Sadly there's no 480 overclock review yet, or 1060 for that matter
djKOhsb.png

Look at 980 at 1488 vs 390 at 1100.

It's 5.6tflops vs 6tflops at those clocks, in favor of the 980; perform equally at 1080p and 1440p. Factor in approximate ~8% performance decrease because of DX essentially emulating the compute queue (it's disabled for maxwell, in drivers) and scheduling the tasks and you have basically equal "architectural efficiency"; the performance extracted from both cards scales similarly with available resources.

the 970 has 20% fewer SMs and less cache than the 980.

I don't imagine they reran all the maxwell cards benchmarks with the Pascal release driver, but that brought a performance improvement of about 9% for me, so it negated the async hit when that was enabled.

There's really nothing to suggest DX12 is inherently problematic.

Edit;

Ext3h on B3D discovered that NV didn't lie; Async was disabled on Maxwell and enabling it in game led to DX essentially emulating the compute queue then serializing by slotting those tasks between others on the graphics queue, so were essentially at the mercy of this heuristic, if it improved performance then your code sucked :p
 
Last edited:
AzixTGO said:
when you make statements like the 1060 is faster than the 480 without any qualifications,
Click to expand...

You're making things up. I never said that the 1060 was faster than the 480 without qualifications. I qualified my statements VERY well. in GAMING, TODAY, the 1060 is PROVEN to be >10% faster than the 480 while consuming 25-30% less power. It is faster and more efficient.

Based on your short yet current posting history, it is blatantly obvious what you are doing. So like all of the other Nvidia and AMD fanboys, I'll also add you to ignore. You've proven a complete inability to grasp a simple concept.
 
You must log in or register to reply here.
Back
Top