Why is AMD losing the next gen race to Nvidia?

[gongo]

So now Nvidia has release the 3G 1060...this is in addition to the full Pascal in a laptop form(gasps!). Nvidia has launched the full range of next gen nm gaming GPU...whereas AMD is struggling with Polaris.

What gives? I was expecting the move to 14/16nm is a form of reset. A new dawn, a new fight. Yet the noteworthy rumors i've heard about AMD is..."AIB are not happy that AMD has not let them know of a high performance part this year" ...

What is taking AMD so long?

On a high level, Polaris do not seem much changed to their 28nm family...almost similar to Pascal-Maxwell. Is AMD waiting for HBM2? Is AMD tied up with PS4Neo?

I am sad Nvidia is given free reign now, and prices are up. Normally, such 'delay-no news' turns out to be bad. AMD big GPU may be suffering from high power drain leakage i fear...RX 480 did not leave a good impression.

 

[ninelven]

Should be fairly obvious.... high bandwidth is the enemy of energy efficiency. The more energy you spend moving things around, the less you have in your budget to actually do stuff. Maximum performance is determined by perf/watt given a fixed max power consumption. Currently, Nvidia has better compression techniques, which allows them to save power potentially 3 different ways (clocks, bus width, lower data movement). For example, the better compression allows Nvidia to use a 192bit memory interface on the 1060, which saves power (vs the 256bit on RX480). And they are also moving less data due to the compression which saves more power. Then, they can either choose to spend the power they saved on better performance, lower TDP, or a mix of both.

If AMD is/was counting HBM2, then the problem is likely one of availability/expense. But compression is the gift that just keeps giving, as it will still save power even if one has an over-abundance of available bandwidth....

 

[Malo]

AMD is relying on DX12/Vulkan and ports using those as their new dawn. They're obviously strong with the new APIs but it's still going to be awhile before we see wider adoption considering game development timespans.

 

[lanek]

Its a question of balance ( i dont speak about gaming but we are not so far ), you need high bandwith, high capacity of storage and efficiency and high compute power, it is a balance between thoses .I personally dont care that my compute gpu's for raytracing eatt 600W if they can offer me the best possibilitty on thoses 3 aspects. ( with photogrametry,, i need 16k-32K textures, i need high resolution HDRi as lighting environnement sources ( let say 16-32K too ), i need extremely high poly counts ) with onlly that i can eat 32GB of Vram in a nutshell, for a singe frame and a single scene. ( animation are render anyway frame by frame, so then the problem is for VFX, there we move to CPU, because its impossible to have enough memory on GPU's )

AMD is relying on DX12/Vulkan and ports using those as their new dawn. They're obviously strong with the new APIs but it's still going to be awhile before we see wider adoption considering game development timespans.

I dont think, they rely on them, but it is in their interest to provide steam on them.. Since a good time, outside that defacto, some features of their architecture find their way with thoses "new API's" ( as it was finally a good idea ), they can provide easely to the developpers, new features, codes compatibles with their architectures.. without having to rely on drivers, for fix the games their gpus are running.. this is allready a good thing in his own.. It was absolutelly funny to see how much DX11 games, was just not made for run on AMD GPU's ... with absolutely no optimization and debug for run on thoses gpu's ... dont ask you where come the overhead of the driver in this cases. And when we look the benchmarks, AMD was not so bad with theirs drivers ..

I will say coulld change things, not willl.. I dont see Nvidia let it go so easely .

 

[silent_guy]

Should be fairly obvious.... high bandwidth is the enemy of energy efficiency. The more energy you spend moving things around, the less you have in your budget to actually do stuff. Maximum performance is determined by perf/watt given a fixed max power consumption. Currently, Nvidia has better compression techniques, which allows them to save power potentially 3 different ways (clocks, bus width, lower data movement).

I think you're putting a little to much emphasis on the compression part. I'm sure it helps, but the low level architecture of the SM itself is much more likely to be the biggest factor by far.

 

[Infinisearch]

Does lack of an R&D budget fit somewhere into this?

 

[kalelovil]

Perhaps AMD didn't expect Nvidia to release so many Pascal GPUs in a short space of time. AMD hoped they could capture the midrange while Nvidia concentrated on the high-end market.

I think you are forgetting Hawaii and Iceland in there as well. Two release s in a year is pretty much the norm - i.e. have NV released more than GM200 and GM206 in 2015?

They might not have had much choice though, having to stretch a smaller R&D budget than Intel or Nvidia across CPUs. GPUs and SoCs.

Alternative optimistic theory: There was a bigger third Polaris GPU planned. Something went wrong and it was delayed. The delay would have meant it came out only months before Vega, and Vega is somehow a much better chip, so it was cancelled.

 

[xEx]

The answer is simple: Lack of resources.

 

[ninelven]

I think you're putting a little to much emphasis on the compression part. I'm sure it helps, but the low level architecture of the SM itself is much more likely to be the biggest factor by far.

Hmmm.... I'm not so sure. Certainly, there are many contributing factors, such at special function units, L0, TBR, z-rate, etc... But I would say the compression is not chump change. The 1060 has 192GB/s bandwidth. One could down-clock an RX480's memory to 6Gbps and measure the performance lost....

 

[DavidGraham]

AMD is relying on DX12/Vulkan and ports using those as their new dawn. They're obviously strong with the new APIs but it's still going to be awhile before we see wider adoption considering game development timespans.

I seriously hope not, one doesn't need to be a wizard to know such strategy is a failed one. First it relies on an uncertain outcome happening within an uncertain time in the future, secondly it assumes the competitor will stand still and do nothing about it while both history and present have shown the complete opposite.

Being unprepared and underqualified now, for the hope that things will change later is a bad business practice which gets companies nowhere but bankruptcy, For example. If we look at AMD's marketing push towards VR with the RX 480/Fury, we see them getting pummeled in almost every VR release compared to NV, so what gives? Well AMD is waiting for DX12 to chime in for VR games, a wait which shall be quite long indeed, so how the heck will AMD convince buyers to purchase their hardware for VR now, when they can't provide adequate performance "NOW"? with empty words?

 

[Razor1]

Does lack of an R&D budget fit somewhere into this?

Possibly but having both both main product lines not delivering starting over different periods of times and the ups and downs of one of those product line, speaks to more then just R&D.

ATI buyout came at a time when AMD thought they were in good position on the CPU side, but it was actually the start of the decline of their CPU side, and then the r600 hit, which didn't do anything for them, until the rv770 they were kinda in murky waters, then once they started working on GCN, they have lost ground ever generation not to mention their CPU side went to the toilet.

Which just speaks of taking wrong steps and not fixing those steps, mainly power consumption, which nV has been able to tout since (Fermi) and AMD took the lead shortly with the rv770.

Lower power consumption equals the ability to have greater performance at the end.

Big question is when ya got the money and ya don't do it, when can ya do it?

 

[silent_guy]

Hmmm.... I'm not so sure..

Let's look at it this way: when you need to save power, it pays to first optimize those units of which you have the most and/or those which run most often.

In a GPU, the calculation cores would qualify for both: you have many of them and they calculate something every clock cycle.

Nvidia promoted BW compression as a new feature for their second gen Maxwell, yet the GTX 750 Ti already had a major jump in perf/W. It's not much different from later Maxwells.

We have a pretty good idea what Nvidia did inside the Maxwell SMs compared to Kepler, they all improve perf/W, and GCN has none of that.

I think compression's primary reason is bandwidth amplification, and that the power savings are a nice second order effect.

 

[ninelven]

when you need to save power, it pays to first optimize those units of which you have the most and/or those which run most often

Well, yes and no.... The actual energy/op matters too. If you have 0.1% "bad" ops, but they consume orders of magnitude more energy than what you are normally doing, it pays to work on that too or obviate the need for them to exist in the first place.

It's not much different from later Maxwells.

2% here, 3% there.... pretty soon you are up to 12%.... Just imagine if tomorrow AMD released at driver that increased graphics performance 12% across the board. Or better yet, if the RX480 had simply launched with +12% performance... This thread probably wouldn't exist.

 

[Erinyes]

Does lack of an R&D budget fit somewhere into this?

This. Nvidia spends in excess of $100M more than AMD per quarter on R&D ($350M v/s $243M from the latest results)

 

[liolio]

The answer is simple: Lack of resources.

Indeed, he best answer around.
There are many thing that set AMD and NVIDIA GPUs aside and it has been a while that AMD GPUs suffers from limitations that AMD did not address.
They lack money for the GPUs but also the CPU they gave up on their most successfully IPs (cat cores). They are also wasting money often pushing theirs chips way past the diminishing returns. I would think that the management sucks too.

 

[xEx]

Indeed, he best answer around.
There are many thing that set AMD and NVIDIA GPUs aside and it has been a while that AMD GPUs suffers from limitations that AMD did not address.
They lack money for the GPUs but also the CPU they gave up on their most successfully IPs (cat cores). They are also wasting money often pushing theirs chips way past the diminishing returns. I would think that the management sucks too.

I used "resources" instead of "money" for this reason. Its not just money AMD is lacking is also talent, the most crucial part. Raja said it already, no1 wants to work on AMD because of its "Dying image" so AMD only can hire what everyone else don't want. In my opinion AMD should have took its shot at ARM instead of trying to fight Intel which has like 20 times more money to fight that fight.

 

[liolio]

I used "resources" instead of "money" for this reason. Its not just money AMD is lacking is also talent, the most crucial part. Raja said it already, no1 wants to work on AMD because of its "Dying image" so AMD only can hire what everyone else don't want. In my opinion AMD should have took its shot at ARM instead of trying to fight Intel which has like 20 times more money to fight that fight.

They may indeed lose talents or fail to attract new ones though I think the management and marketing is not giving the wrong goals to the engineering teams. "More" seems to be the mantra ( in lot of companies) instead of better... Or even better do it right. The former is way easier to deliver and it is easy to hide responsibilities: "we were ask more, we delivered it, it failed but I held my side of the deal". I suspect those that would advocate to pass on some marketing checks boxes for the best or better are shunt fast... It happens in lots of place, it is like the corporate world decayed a lot since the " sellers" took over the "doers" but that is a more general issue.

 

[silent_guy]

2% here, 3% there.... pretty soon you are up to 12%...

Yes, it had to be a grab bag of tons of little improvements. But first gen Maxwell improved perf/W by around 80%, without any known improvements in compression over Kepler. That really all that I'm saying.

 

[Anarchist4000]

Alternative optimistic theory: There was a bigger third Polaris GPU planned. Something went wrong and it was delayed. The delay would have meant it came out only months before Vega, and Vega is somehow a much better chip, so it was cancelled.

It seems likely a larger one was planned, but considering the release date and rumors of Vega in October maybe it was just ditched because of that? It wouldn't really have been a delay as the FINFET designs came from both IHVs around the same time. If Vega does show up in October that still leaves it months prior to the release of it's successor. All the roadmaps indicate Vega is HBM2, so it's entirely possible they are geared towards high end and APUs where they may be on interposes. If they simply took Fiji and shrunk it down with FINFET they should have had a competitor for GP104.

 

[ninelven]

...compression over Kepler. That really all that I'm saying.

I wasn't (nor is this thread about) comparing Nvidia's compression (they started it with Fermi btw) to Nvidia, but w/e......

Anyway, I did a rough estimate based on the data available, and I would guesstimate that Nvidia's compression advantage is responsible for ~1/5 to 1/4 of their perf/watt lead. So a sizable chunk, but by no means all or even a majority of it. One could call it a necessary but not sufficient condition for AMD to regain competitiveness.