AMD's Polaris Architecture Is A Double-Edged Weapon

| About: Advanced Micro (AMD)
This article is now exclusive for PRO subscribers.


AMD has presented Polaris 10 XT as a RX 480 video card at a very competitive price tag of $199/$229.

But in view of its power consumption, there are issues for the desktop and mobile sectors to be considered.

RX 480 has the same TDP of the GTX 1070, while it will perform similarly to the GTX 1060 that will probably need only the PCI-E power (only 75W).

Nvidia is going to release GTX 1080 and 1080M for the mobile market (150/125W) with desktop-like performances. AMD, on the other hand, won't be able to go further beyond the RX 480 performance (150W already).

Considering how important the laptop market is, this is a concerning factor, and the upcoming mobile Polaris 11 chip is a budget solution only.

I have stated in a previously article that AMD, Inc.'s (NASDAQ:AMD) Polaris could have to face some power consumption issues, mostly related to the desktop high-end and the mobile sector.

This kind of statement is based upon what the Polaris 10 XT (RX 480) is able to do from the performance/power ratio point of view, and similar reasoning can be done considering the Polaris 10 Pro (RX 470) and Polaris 11 (RX 460).

The issue is that AMD will have problems facing Nvidia (NASDAQ:NVDA) in the desktop high-end and mobile sectors, because its architecture inherent power consumption is a weakness that will be hardly counterbalanced in the mobile sector, where power consumption must be restricted.

In addition, it looks like AMD has no real hardware tweak like Nvidia VR features (Simultaneous Multi-Projection, single stereo pass, Lens Matched Shading), and this lack may be an additional and consistent weakness for the company's solutions in the following months.

RX 480, Benchmarks And Power Consumption

Starting from the RX 480 specs, the Polaris 10 XT GPU exhibits 2304 cores that run at 1,050 MHz for the base clock and 1,266 MHz for the boost clock, achieving an average 5.5 TFLOPS of computational power (from 5.1 to 5.9 TFLOPS). The video card comes with a memory bus width of 256 bit and a bandwidth of 256 GB/s, which may be a little low given the well-known GCN architecture bandwidth hunger.

RX 480

Credits: Ars Technica

Considering its TFLOPS performance, this video card is probably going to perform like a R9 390 or R9 390X, and leaked benchmarks show this kind of performance range:


Source: 3DMark database

In the above table, you can see a series of 3DMark Fire Strike (Ultra and Extreme) results found out on the 3DMark website. These results comprises similar memory configurations and frequencies, similar CPUs, clocks and stock frequencies for the GPUs.

The result is that the RX 480 provides little lower performance than the R9 390X (-2.1%), but it is +5.3% faster than the R9 390. It is also -6.7% slower in comparison to the GTX 980.

If we consider 3DMark 11 Performance (Graphics) instead, we see a similar picture, where the RX 480 is 2.9% slower than the R9 390X, 6.3% faster than the R9 390, and 1.1% faster than the GTX 980. By the way, this benchmark is 5 years old, and it is less relevant in comparison to Fire Strike.

If we consider that this video card is going to be sold starting from $199 (4GB) or from $229 (8GB), and that it meets the $500 performance of the previous generation, it is something really astonishing. Nvidia has still not announced anything to challenge these solutions at these price levels.

But there are some issues to be considered, and I think they are relevant.

Still High Power Consumption

First of all, its TDP is specified at 150W. The video card will probably consume 120-130W in average gaming, a value that is comparable to the effective power consumption of the Nvidia GTX 1070 (it consumes an average of 135W while gaming). The fact is that the 1070 is a lot faster, and thus, it is more efficient, so it looks like Nvidia has maintained the efficiency advantage it gained with Maxwell architecture.

Obviously, this performance/price achievement from AMD is remarkable, and it still has no serious competition in that price range, but we must consider that this Polaris' behavior has a series of consequences that may pose serious problems for AMD.

Mobile Sector More Important

The first issue shows itself in the mobile sector: it requires restricted TDPs in order to maintain a decent battery life and sensible heat generation that must be managed by the dissipation system. Let's try to consider what is the actual top-end solution for the notebook sector: we have the GTX 980 notebook version with a TDP of 145W, which is very high even for a gaming notebook.

The point is that the RX 480 already meets that TDP with 150W. Okay, it doesn't mean that it consumes that energy, but that AMD won't be able to provide anything better than a RX 480 performance class - that is what a GTX 980 notebook version is capable of with the same TDP (at least with Fire Strike Ultra).

And this looks not good at all for AMD, because Nvidia is going to hit the mobile market with the notebook versions of the GTX 1080 (150W), the GTX 1080M (125W) and the GTX 1070M (100W-80W). In practice, Nvidia is going to provide a consistently better performance with the same TDP or a consistently lower TDP with the same performance, providing a very attracting performance/watt ratio that is fundamental in this kind of market.

GTX 1080M

Credits: VideoCardz

Note: In my previous article, I have written about rumors concerning the GTX 1080 Notebook Version (150W), GTX 1070 Notebook Version (125W), GTX 1080M (100W) and GTX 1070M (75W). But given the latest leaked information, it is more likely that Nvidia won't release a GTX 1070 Notebook version, since its CUDA core count (1920) is very similar to the GTX 1080M (2048). At the same time, given that the GTX 1080M starts with a consistently lower clock (1,442 MHz), and given that the GTX 1070 consumes an average of 135W, the GTX 1080M is more likely to be set with a 125W TDP, in my opinion.

GTX 1060 Is Coming

An additional issue arises from the fact that Nvidia is rumored to anticipate its mid- to low-end solutions release in the following weeks, in order to counteroffer the AMD RX 480 and 470 solutions. In particular, it is expected to release the GTX 1060 Ti with the GP104-150 GPU, capable of a performance between that of the GTX 980 and the GTX 980 Ti, while the GTX 1060 is expected to be equipped with a GP106 GPU that would provide a performance near that of the GTX 980 and the RX 480.

Considering that Zauba reported the import of GP 104 GPUs in the first two weeks of May, and that the GTX 1080 has been released at the end of May, Nvidia will probably release the GTX 1060 in the following days or couple of weeks.

This means AMD will have very restricted time to exploit the Nvidia Pascal market absence in the mid- to low-end segment.

Nvidia Has Really Improved SLI

Another thing that must be considered, even if it is niche, it is the SLI/Crossfire comparison: leaked benchmarks of 2 x RX 480 (Crossfire) show that a dual RX 480 scales up by 75% in 3DMark Fire Strike Ultra. This is remarkable, since it brings a $400/460 Crossfire to perform like the GTX 1080 ($599).

But there are two issues here: first, a dual RX 480 has a TDP of 300W against a TDP of 180W of the GTX 1080; and second, a dual SLI of GTX 1080 scales up by 97% at least (3DMark database). AMD has usually been able to scale better in multi-GPU configurations, but with Pascal and its beefed-up SLI connection, the situation looks to be reverted. This is an aspect that could be also replicated by the lower video cards like GTX 1070 or GTX 1060, meaning that the convenience of buying two RX 480s may not last for so long.

Vega Is Late

Vega will be the high-end GPU chip from AMD, and I am sure its performance will be very good. The problem is that this kind of GPU, considering what Polaris is capable of, will have a TDP in the range of 250-300W (air cooling). This is a very high power requirement, and it remains to be seen how effective AMD's work on this new architecture will be. Such power consumption obviously means we won't see anything with Vega 10 in mobile workstations and mobile gaming laptops, because the company may be able to improve the power efficiency, but it won't be able to reach the same efficiency leap that 14nm adoption has provided. This means it will lose (or not gain) high-margin segments in a market that is essentially growing and partially adsorbing the high-end desktop market.

But most importantly, it looks like AMD has postponed Vega to year 2017, revealing that either Global Foundries still has no capability to achieve decent samples and yield to hit the market in late 2016, or Vega GPU is still behind schedule from the architectural point of view, or maybe both. It must be remembered that Vega was initially forecasted to be released in 1H 2017, but the accelerated release of Pascal video cards caused AMD to try (without success) and anticipate Vega.

Vega Will Face Volta

The fact that Vega will be released in 2017 means that AMD will have to face Nvidia Volta after few months. Given the first announcements from AMD and Nvidia, we can expect a similar boost from Polaris to Vega and from Pascal to Volta, which means the actual balance will not change so much.

In addition, Nvidia is going to complete its NV link strategy with Volta, and that would be an additional attracting enterprise feature. AMD has to work hard on the power efficiency aspect in order to really gain relevant market share from Nvidia.

No VR Nvidia-like Features

Another point that will be relevant with time is that Nvidia has provided a very important feature set for the VR market: Simultaneous Multi-Projection and single stereo pass enable Pascal GPUs to generate and project one to sixteen different geometry projections in a single pass, with an irrelevant performance loss and cancellation of screen distortion. The Lens Matched Shading, instead, is able to greatly improve performance by eliminating the rendering of those pixels that wouldn't be reproduced due to virtual reality lenses' distortions.

AMD still does not have anything similar at the moment, and this may be troublesome in the following months of software/gaming development.

RX 470

As regards the RX 470, this video card still uses the Polaris 10 chip, but with less cores (2048) and lower GPU/memory frequency, providing a performance that is somewhere around that of the GTX 970. This is essentially the same RX 480 GPU chip, and therefore, the previous considerations do not change.

Polaris' Low Margin

The lack of a Polaris for the high-end market, and the fact that Vega is dated to year 2017, make me think AMD has serious problems with the architecture and the miniaturization process in Global Foundries, implying that the company is not able to achieve high enough frequencies with bigger GPUs. At the same time, even the relatively low frequency of the RX 480 sounds like an alarm alert.

The decision to target only the mid- to low-end of the GPU market will drive low profitability and margins.

What To Expect

Anyway, AMD will surely sell well. However, the company will sell relatively low-margin video cards, and it will have to face an Nvidia counteroffer in the same market range. AMD still has no breaking feature for the VR market, and that could be problematic in the following months.

Not to mention that the company still has a long way to go before releasing its high-end solutions, which means Nvidia will have free rein in that market range for nearly a year.

Personally, I think AMD was constrained to release Polaris as a mid- to low-end GPU due to fundamental power consumption and/or architectural limitations that may be partially corrected with future Vega architecture.

The company has been able to raise its stock price beyond $5 per share, which is a very good and remarkable result. As I stated before, it is possible that AMD will reach $10 just before the ZEN release: the hype generating around the ZEN architecture is very high, but such a target price may be possible only if Nvidia releases the GTX 1060 in late 2016. If Nvidia releases it in the following weeks, AMD's life would be more tough until the ZEN release. In any case, we must wait for the first RX 480 reviews in order to understand if AMD will meet all the recent hype beyond the RX 480. If everything goes well, the run for the $10 target will still be sensible.

Disclosure: I/we have no positions in any stocks mentioned, and no plans to initiate any positions within the next 72 hours.

I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.

Additional disclosure: The author does not guarantee the performance of any investments and potential investors should always do their own due diligence before making any investment decisions. Although the author believes that the information presented here is correct to the best of his knowledge, no warranties are made and potential investors should always conduct their own independent research before making any investment decisions. Investing carries risk of loss and is not suitable for all individuals.