- AMD's has advertised big performance per watt gains for the upcoming Puma chips.
- Examining the Jaguar design makes the performance gains highly plausible.
- Looking at information published following MWC (mobile world congress), combined with Geekbench 3 data and AMD's press data, we will explore Mullins performance potential.
Fellow Seeking Alpha contributor Alcaraz Research surmised that free Microsoft (NASDAQ:MSFT) 8.1 licensing for small form factor tablets could be a tail wind for Advanced Micro Devices (NASDAQ:AMD), spurring on growth of the gaming tablet.
His article focused mainly on AMD's Project Discovery Tablet, which was demonstrated at this year's CES (Consumer Electronics Show).
In this article I would like to focus on AMD's performance claims for the upcoming Mullins APU. I will also look at a potential entry for a Mullins reference platform in the Geekbench3 benchmark database to shed some light on possible specifications and performance of the Mullins chip. Finally, I will conclude with a discussion of Beema's chances to meaningfully impact AMD's financials.
As the first few sections are highly technical, if you are in a time crunch or don't want to suffer the details, I will provide a brief TL;DR (too long; didn't read) explanation at the end of each of the technical sections. But because milliwatts really matter in mobile, it's important to understand the minutiae involved.
As a point of reference, AMD names the company's smaller cores using cat names. Jaguar and Puma are the codenames for the underlying architecture. Kabini and Temash are the specific names given to APUs built around the Jaguar architecture. Mullins and Beema will replace Temash and Kabini and are APUs built using the Puma architecture.
Survival of the Fittest: The Puma and the Jaguar
AMD's Jaguar architecture was a major improvement over the Bobcat cores it replaced. This improvement manifested itself in the form of more performance at a lower power consumption.
But that was last year, and this year we have a new story on tap.
AMD will be releasing the new Mullins and Beema chips in short order based on the Puma architecture, and the company is claiming a substantial increase in performance per watt when comparing Puma and Jaguar chips.
Note that the Mullins and Beema chips above have substantially lower power consumption figures than the Jaguar chips they are compared against. Also notice the 4.5W Mullins chip achieves almost the same score as the 25W Kabini chip in PCMark 8, and is within throwing distance of the GPU score of the 25W Kabini.
Similar (but larger) performance claims were made when AMD transitioned from Bobcat to Jaguar. AMD's claims for the Jaguar chips stood up to independent benchmarking when the platform released.
Source: AMD's Slide Deck via TheVerge
The difference in these scenarios, however, is that there was a node shrink from Bobcat to Jaguar, while both Jaguar and Puma chips are manufactured at 28nm.
So while the performance increases from Jaguar to Puma may be a little harder to believe, they are completely within the realm of possibility if you understand all the low hanging fruit available to increase the performance of Jaguar.
TL;DR: Because the Beema and Mullins platforms are both 28nm products the performance increases may seem unrealistic. Therefore I will demonstrate why I feel they are believable.
Why the Performance Increases Are Believable
A quick stroll along the timeline of AMD's recent (within the past 3 years or so) of processor releases reveals a trend: AMD typically releases a second version of a very similar architecture that improves performance per watt. This was the case with both the transition from Bulldozer to Piledriver, as well as Trinity to Richland.
In the case of the transition from Jaguar to Puma, the performance gains are larger than for the releases above so I would like to do a quick dive into the Jaguar architecture to understand why these performance increases are plausible. I will build my argument based on:
- Clock speed vs. performance between Jaguar and Bay Trail on the CPU side
- GPU performance between Jaguar and Bay Trail
- Explore the differences between rated and actual power consumption
- It's possible that Puma will be built on a different 28nm node
- Information from a leaked GeekBench 3 score along with a video from a hardware review site reveal potential details about frequency.
Regarding Clock For Clock CPU Performance, Jaguar is Stronger than Bay Trail
Intel's (NASDAQ:INTC) top of the line Z3770 Bay Trail features a 1.4 GHz base speed, with a 2.4 GHz boost clock in a quad core configuration. Rated SDP (scenario design power) is 2W, and it's speculated that overall power consumption is rated for ~4W.
When Bay Trail debuted, the chip was often pitted against AMD's A4-5000. The A4-5000 features 4 Jaguar cores running at up to 1.5 GHz. Overall rated power consumption for the A4-5000 is 15W. I'll address the power consumption delta between these processors in a later section.
To make a direct comparison between cores, I am focusing on single threaded benchmarks. According to the graphs above, the A4-5000 Kabini is evenly matched with the top of the line Bay Trail chip in Cinebench (stresses floating point performance), but Intel pulls ahead by 25% or so in 7-zip (an integer workload). While the type of work being performed isn't as important to understand, in his article Anand states that integer workloads are probably more important in a tablet, so I will focus on the integer performance specifically.
Keep in mind that the Z3770 is likely running at 2.4 GHz (the boost frequency of the chip). To normalize for clock frequency, the Intel chip would score a 689 (1655 / 2.4 GHz). AMD's A4-5000 would achieve a score of 882 (1323 / 1.5 GHz).
On a clock for clock basis, AMD's chip has enjoys about a 30% performance advantage over Bay Trail using 7-zip (representative of integer workloads) as a metric. A similar calculation using Cinebench would yield an ~50% performance advantage for Intel. However, keep in mind the Bay Trail chip is using significantly less power.
TL;DR: Intel's chips must run at higher CPU frequencies to achieve the same benchmark scores. However, despite higher clocks, Intel chips do consume less power.
Beema and Mullins Offer Up to a "250%" lead Over The Competition in Graphics
During her CES 2014 keynote, AMD's VP of Global Business Units Dr. Lisa Su made the claim that AMD's new low power APUs would offer "250%" better performance over Bay Trail regarding purely graphical workloads. That's an astronomical number, so let's see if it's believable.
As a standard of measure, I have found 3DMark 11 Performance scores for the platforms in question.
Source: AMD's CES 2014 presentation via SlideShare
For the sake of convenience, according to note 10, AMD's Mullins quad core APU scores 570 in 3DMark 11 Performance compared to 214 for Intel's Z3770. As points of reference, I have looked at a couple of tech sites (I, II) to verify AMD's claims for Bay Trail align with independent results, which they do. According to math, 570/214 is 2.66, so Dr. Su's claims are valid for pure graphical workloads.
3DMark 11 Performance setting benchmarks various systems using the DirextX11 API from Microsoft. The test is run at 720p resolution to provide continuity to ensure the metrics are comparable.
Source: 3DMark Database
In pure GPU workloads Mullins looks to completely outclass Bay Trail. Even if compared against the temash A6-1450, the 3DMark 11 database, backed up by reviews on NotebookCheck, shows scores of ~460 points for the A6-1450 chip.
Recall the above discussion of Turbo features. Purely graphical workloads may not necessarily be indicative of real world performance given the power consumption is shared by the CPU and GPU. If you're running a game that taxes the CPU, there is less power available to the GPU.
To show real world comparisons, according to data on NotebookCheck (I, II, III), the A6-1450 scores 20 fps on the Tomb Raider benchmark on low, whereas a platform based on the Z3770 only manages 13 fps. The A4-5000 manages a more respectable 27 FPS at the same settings, teetering on what most gamers consider playable.
There is a large performance difference in real world scenarios between Bay Trail and AMD's offerings. While the real world performance gap is more narrow than the theoretical gap, it is still quite substantial. And this is comparing Temash and Kabini against Bay Trail; not Puma. If AMD delivers with Puma, this gap will only widen.
In reality, I doubt the ability of either AMD or Intel to field a gaming powerhouse in a sub 5W chip. But for those that are okay with gaming at reduced settings and resolutions, the offerings from AMD's current line up look to be a better choice than Bay Trail. The issue now is that Bay Trail is in tablets, where AMD is essentially absent from the tablet space. In a later section I will demonstrate why graphics are arguably the more important part of the equation (provided the CPU is at least adequate) in tablet form factors.
TL;DR: AMD's chips offer substantially more powerful graphics. Neither AMD nor Intel chips will play AAA titles at decent settings. However, because of AMD's more powerful graphics capabilities it is still the better choice for the consumer interested in graphics.
Power Consumption: Max vs. Actual
Author's note: Thanks to reader "raghunathan78" for the discussion regarding power consumption.
The starting argument for AMD's viability in the tablet space begins with debunking the disingenuous claims surrounding Intel's 2W chip beating AMD's 15W chip.
Just because it's rated for 15W doesn't mean it uses nearly that much juice.
The first thing to point out is that, based on the graphic from AMD's slide deck (linked to earlier), SoC power consumption is a small fraction of total platform power consumption. When OEMs launch a tablet design, considerable effort goes into minimizing the power consumption of the overall device.
In Anand's review of the A4-5000, he finds that running Cinebench R11.5's single threaded benchmark causes platform power consumption to jump from 4.75W to 7.91W, or a change of 3.15W. He had Intel engineers perform similar tests for the Z3770 debut, with the difference being that the engineers were measuring energy consumption at the SoC, which doesn't encompass the increase from the rest of the platform. Anand stated the Z3770 was drawing ~1W during the Cinebench R11.5 single threaded benchmark. Using AMD's press deck slide as a guide, this leads me to believe a similar figure (if only SoC power consumption were measured) of around 2W would apply to the A4-5000.
To quote Anand directly concerning the Z3770 multi-threaded benchmark (Anandtech article linked to earlier):
Multithreaded performance puts Bay Trail and AMD's Kabini at similar performance levels. Once again, looking at SoC power however the Atom Z3770 pulls around 2.5W in this test. Looking at the increase in platform power for the A4-5000 here, I'm assuming that the equivalent data for AMD would put Kabini in the 6W range.
Intel currently enjoys a nice performance/watt advantage over AMD.
Looking behind the performance numbers you'll notice the A4-5000 scores .39 points in the single threaded benchmark, and platform power consumption increases by just over 3W. Looking at the multi-threaded benchmark, you'll see a score of 1.5 corresponds to a power consumption increase of 6.75W.
This information tells us that AMD's chip isn't operating a single core very efficiently. Taxing 1 CPU core causes an increase of just over 3W, whereas taxing 4 cores causes an increase of just under 7W. Based on the single core number, you would expect an increase of closer to 13W, but this is not what we see.
I'm going to steal tech writer Joel Hruska's phrase that "0W is the new 1 GHz" (sorry if I butchered your saying Joel). Around the turn of the century the cool thing to do was to race to 1 Ghz. Now with the focus on mobile, chip makers are trying to drive power consumption as close to 0 as is possible. The idea is that when a CPU is tasked the CPU will perform the task as efficiently as possible and then "go back to sleep" (known as sleep states). The more of the core that goes to sleep, the lower the idle power consumption. The more efficiently and quickly the task is completed, the lower the active power consumption.
All this is to say that Intel's 2W chip outperforms AMD's 15W chip is a specious argument. Sure, there is a performance per watt difference, but it isn't the 7 fold difference that looking at rated power consumption suggests. The 15W chip doesn't consume nearly 15W, and the workloads refer only to CPU tasks. There is very little power consumption data regarding Bay Trail's GPU so it is hard to draw meaningful conclusions with respect to graphics. You may think I'm nitpicking, but every milliwatt counts in mobile so these nuances are important to understand.
TL;DR: Although Intel does enjoy a perf/watt lead over AMD regarding CPU performance, the lead is often exaggerated. You cannot look at rated power consumption to make comparisons. But because we're discussing tablet form factors, small differences can still be very important, giving the advantage to Intel.
How Can AMD Improve Power Consumption?
The possibility of AMD producing Puma chips on a better 28nm node is speculation only at this point, and is based on the timing of the release of Jaguar, other chips using 28nm HPM, and AMD's claimed TDPs for Puma when compared against jaguar.
Based on AMD's press deck, we know the A4-5000 consumes roughly .77W when idle. Compare this to the Z3770 which Anand stated had about 1W power consumption on a single core at full load; AMD's idle power consumption with Jaguar is nearly equal to Intel's power consumption with one core loaded.
This is why advances such as Microsoft's Instant Go feature are important. It will allow the chip to enter a deeper sleep state to save power.
Active power consumption can be reduced by adjusting how the chip behaves (more tightly controlling clock frequencies for various workloads). The power consumption increase between idle and a single core being loaded was much higher than the commensurate increase between loading 1 core and what would be expected when loading four cores. If most of the chip gets turned on when the single core is loaded, this is going to drive power consumption much higher. Being able to keep more of the chip turned off would reduce this power consumption. Finally, controlling the frequencies of the CPU and GPU to ramp up or down for corresponding workloads can also help with further power reduction.
Finally, AMD's Jaguar platform began volume shipments in Q1 of 2013. This was before TSMC had the 28nm HPM up and running (I, II, III). TSMC has a few varieties of 28nm chip production, with 28nm HPM being the ideal choice for mobile.
TSMC also provides high performance for mobile applications (HPM) technology to address the need for applications requiring high speed as well as low leakage power. Such technology can provide better speed than 28HP and similar leakage power as 28LP. With such wide performance/leakage coverage, 28HPM is also ideal for many applications from networking, tablet, to mobile consumer products.
As it turns out, Microsoft's Xbox One uses TSMC 28nm HPM.
This is significant as the Xbox One APU uses Jaguar cores and the GCN GPU architecture. AMD does not typically disclose which products are made at which foundries, or which processes are used at those foundries, but based on timing I believe Kabini and Temash chips were launched on 28nm HP. We can see that the same Jaguar CPU architecture and GCN GPU architecture is up and running on 28nm HPM, so there is a chance that Mullins could be built on a better 28nm process than it's predecessor. This would help with power consumption all around.
TL;DR: There are a few changes AMD could make between Jaguar and Puma, namely deeper sleep states and active power management, to achieve some decent power consumption gains over Jaguar. There is also the possibility that the chip is made on a 28nm process better tuned for mobile, but at this point it's only speculation.
How Much More Performance Could Mullins Bring? A Look At Clock Speeds
This section is also speculative. Using two separate data points, I believe the CPU frequency of the 4.5W Mullins chip will be 1.2 GHz base and 2.2 GHz boost.
AMD's press materials point to between 20% and 30% on both the CPU and GPU. Very little information about the GPU portion of Mullins is available, so I am assuming that it uses the same 2 GCN cores as Jaguar, and implements some of the changes described above to increase performance.
CPU performance is more interesting. This is the area where AMD bears like to point fingers.
Geekbench 3 is a popular benchmarking tool used to quickly analyze performance metrics of various chips across different platforms. There is a data base that collects information that readers can browse.
Source: Geekbench Database
The above represents what I believe to be an entry for the 4.5W Mullins chip. A little elbow grease and some nerdery decodes the processor information string to reveal that the frequencies for the chip should be 1.2 GHz base with a boost frequency of 2.2 GHz boost. I set the comparison point as the 8W A6-1450 Temash processor. Geekbench has numerous entries for the A6-1450 at varying performances, so I randomly picked an entry that looked to be middle of the road for the A6-1450.
As a second point of reference, Semiaccurate forum poster "kef" pointed out this video (courtesy of NordicHardware) and this article (courtesy of YourNewsTicker) to me. Although the video is in swedish, I believe the video states the boost frequency is between 2.0 and 2.5 GHz, with the article stating the base frequency is 1.2 GHz and power consumption is 4.5W.
Note that I cannot verify the Geekbench entry is for the 4.5W chip, or that these clock frequencies will match final silicon, but based on the material provided above I believe this chip is likely the same SKU AMD is using to generate the press information. Also the If so, it points to a healthy performance lead over the A6-1450. And if this is indeed a 4.5W chip, a huge increase in performance per watt, despite being at the same node.
A higher base frequency, combined with an aggressive boost frequency, should make this product feel much snappier when compared against the Temash platforms. Also, the stronger CPU could easily benefit the GPU, improving real world performance.
TL;DR: A combination of press material from AMD, press reports from the recent MWC, and a Geekbench 3 entry point to a possible clock speed of 1.2 GHz base/2.2 GHz boost frequency for Mullins. If this is indeed turns out to be the case (which it may not as there is no official release from AMD), Mullins should offer a nice performance bump over Temash. This will benefit both the feel of the device (snappiness), as well as allow stronger GPU performance.
Tablets Aren't My Main Concern or Hope
As the rest of the world focuses on mobile, that's a fairly bold statement to make.
First I would like to look at AMD's change in market share regarding notebooks. Second, I will look at a couple of the hurdles for AMD to be successful in the tablet form factor.
Market Share Trends
ZDNet's report shows the fluctuations of market share between Intel and AMD. You'll notice that AMD actually gained market share between Q4 2012 and Q3 2013, only to see these gains erased and then some in just one quarter.
IDC data shows that the split between notebooks and desktop PCs is roughly 60/40. So losing basis points of market share in the notebook space is more substantial than a loss in desktop. Breaking this down further, PC sales between Q3 2013 and Q4 2013 were roughly flat at around 82M units for each quarter (source: IDC I, II).
Using the 60/40 split and an average of 82M units per quarter, AMD's shipments likely fell from around 7M units in Q3 to 5M units in Q4.
Everyone is focused on tablets, forgetting that PC sales are still important. I would like to see nice design wins come to market quicker on the Beema and Mullins platforms, and I'm not particularly concerned whether the design wins come in the form of a gaming tablet or just some nice notebook offerings.
Regaining lost notebook share likely represents an opportunity for AMD far larger than anything I can reasonably fathom for tablets.
AMD's Obstacles In Tablets
First, in my opinion, Windows 8.1 is not a mobile OS. Minimum system requirements for Windows 8.1 require at least 16 GB of storage for the 32 bit version of the OS, with the 64 bit version requiring 20 GB. That's a lot of space on a tablet.
Looking at an entry level tablet on Amazon using Android, you can see that it only has 4GB of storage. Just based on the sheer size of the program, Windows tablets require more storage which drives up price. This doesn't take into account an increase in the required amount of RAM to run Windows either.
Windows is a much more complicated program that requires a lot of horsepower to run effectively. Take for example the 3rd generation iPad. This particular model launched in March of 2012, about a year before Temash. The A5X processor offered about 1/3 of the processing power of AMD's A6-1450.
The A5X would've struggled to run Windows, yet it was extremely successful in the iPad, thanks to a streamlined OS designed from the ground up for mobile; not a desktop OS masquerading as something good for tablets. Windows requires more processing power, memory, and storage space to run effectively. This puts Microsoft, Intel, and AMD all at a disadvantage concerning Windows tablets.
When Temash first launched it was competing against Clover Trail, not Bay Trail. This video shows the performance discrepancy between the two. Yet despite the stronger performance from Temash, we never really saw enough tablet design wins materialize to move the needle.
I believe AMD really pushed to get Temash out the door well before Bay Trail, and with Temash based devices starting to show up in May 2013, the company succeeded. The chip was a more potent solution than Clover Trail, yet it didn't manage to get any higher volume tablet design wins.
AMD had attractive reference designs last year, yet none of these designs materialized. The Project Discovery tablet is a phenomenal looking device, but this doesn't mean we will see the unit available for retail.
When the tablet was first announced in January, during an interview with TechRadar AMD's Dr. Lisa Su "immediately warned that this device will not see a commercial release."
In the interest of keeping the cart behind the horse, I am going to wait with bated breath to see some product announcements before I get my hopes up for tablet sales to prove accretive to AMD's financials.
Notebooks are a different story entirely. AMD has had a reputation of offering very strong mobile solutions at entry level price points. Brazos (an earlier low power chip from AMD) was extremely successful. Jaguar was a leap above Brazos, and it looks like Pumas can outrun Jaguars.
To make sure my meaning isn't lost I believe the Puma platform will be extremely solid. CPU performance increases are highly plausible and AMD already had a commanding lead in graphics performance.
Source: AMD's 2013 Mobility Press Deck via SlideShare
AMD's slide deck from last year regarding PC projections shows the traditional form factors in a down trend, but it also shows thin and light notebooks growing. You'll also notice how small the company is projecting the Windows tablet market to be.
IDC projected about 220M units in 2013, with Windows commanding a 3.5% market share. This translates to about 8M Windows tablets for 2013. Fudging the numbers north to 10M units for 2014 to account for potential growth still represents a market much smaller than traditional PCs. AMD has also shown very little inclination in competing in non-Windows devices.
Don't get me wrong; I absolutely love the Project Discovery concept and would be one of the first ones to buy one.
But the PC market is much larger. Beema and Mullins, based on current information, look to be solid performers. If it turns out the Puma chips fall short of the performance/watt necessary to make them ideal tablet solutions they should still be extremely potent notebook chips. These chips are ideal for the thin and light and small form factor notebooks that AMD is projecting to grow over the next couple of years.
Hopefully the launch of Beema and Mullins, a new desktop chip line, and Kaveri for notebooks will be enough to stem the loss of market share in notebooks for AMD at a minimum, or ideally providing an avenue for growth. Some nice tablet design wins would be welcome, but I feel the obstacles I presented above present some genuine challenges. Windows 8.1 is less than ideal for tablets, and Mullins won't sell if OEMs don't adopt the chip and consumers don't adopt Windows in tablets.
But with an extremely strong portfolio of chips perfect to increase notebook marketshare, my attention will be more focused toward AMD getting some design wins in the notebook space.
Additional disclosure: I own both shares and options in AMD and actively trade my position. I may add or liquidate part or all of my position at anytime, or initiate a hedge via puts prior to the upcoming earnings call. I may also liquidate my position in INTC prior to the upcoming earnings call.