AMD: Mobile Kaveri And Semi-Custom Better Explained

Jun.18.14 | About: Advanced Micro (AMD)


AMD recently launched the mobile Kaveri platform targeted at higher performance notebook PCs.

These new notebook chips, at initial glance, look like they will improve AMD's standing in the notebook space.

Recent tidbits revealed by AMD have given some additional clarity into the semi-custom business.

AMD has managed to hold onto a significant portion of the gains in share price during the second quarter. This has also been an eventful quarter, as AMD has disclosed more of the roadmap for future plans, as well as provided some additional insight into the semi-custom and growth strategies.

This article is best viewed as two stand-alone articles: one regarding mobile Kaveri and one regarding semi-custom.

Why Have Laptops Taken Such A Hit?

From a product perspective, AMD has essentially two separate APU lines destined for laptops: ultra-low power chips such as Beema, and low power versions of the desktop class APUs.

During early 2013, the company launched a new line of the ultra-low power chips, and refreshed the higher performing Trinity APUs in order to launch the Richland platform. The Richland platform was essentially the same as Trinity, but managed to clock slightly higher while consuming slightly less power. In essence, it wasn't anything new.

Looking back to the Q4 earnings call, AMD states that laptop sales took a hit as the company's partners were trying to thin out inventory.

During Q1, AMD started shipping versions of mobile Kaveri and Beema to partners, resulting in a slight uptick of notebook unit shipments. During Q1, CS revenues fell sequentially despite the uptick in notebook volumes.

2013 was a pretty brutal year for AMD regarding notebooks, as Intel (NASDAQ:INTC) launched the Haswell platform, while AMD was left to defend itself with a slightly tweaked version of the Trinity platform.

Click to enlarge

Source: ExtremeTech

How Can Kaveri Change the Trinity/Richland Situation?

Prior to the launch of Kaveri, I explained why the chip stood a pretty solid shot at showing some decent performance gains. However, once Kaveri was released, I pointed out that the 95W desktop chips weren't really that interesting, but at lower powers things could look differently. I spent quite a bit of time working via email with ExtremeTech's Joel Hruska to perform some frequency vs. power consumption testing in order to better understand how Kaveri behaved over a range of frequencies.

Source: The testing I performed with Joel

You can see above that the particular chip we tested behaves very linearly up to around 3.0 GHz, and we hit a knee around 3.3 Ghz where power consumption begins to rise drastically.

Once I noticed this behavior, it was evident to me why AMD released a chip that only slightly outperformed the chip it replaced. AMD could have clocked Kaveri higher, but that would have ballooned the TDP rapidly. Note the point in the slide below regarding CPU frequency at higher powers.

Click to enlarge

Source: AMD via ExtremeTech

This jives with most of the reviewer comments as well. Reviewers almost unanimously praised the 45W A8-7600 for offering most of the performance of the 95W Kaveri chip in less than half the power.

Does Kaveri Look Like It Improves The Situation?

I believe the answer to the headline of this section is a yes, and will explain why.

Source: AMD

As to not bogart my sources of information from the readers, GeekBench is an excellent tool for finding out some information regarding CPU products. Although there can be a lot of variance between some of the tests run, GeekBench collects a variety of data on the CPUs that run these tests.

Click to enlarge

Source: GeekBench

Above shows the comparison of an A10-5745M against an AMD Ballina reference platform. Using some nerd elbow grease decodes the "ZM207094H4468_32/20/12/05_130A" processor string to be a 2.0 GHz/3.2 GHz boost part with an ~500 MHz GPU frequency featuring the 130A variant of the Kaveri iGPU. Note that the decoded frequencies don't quite align with any of the Kaveri SKUs, but the reported frequency is actually 1.9 GHz. Also, the name Ballina suggests this is the reference platform for mobile Kaveri. Based on these reasons, I believe this chip to be the 19W A10-7300.

Source: AMD

Because there is a lot of variance between benchmark runs on GeekBench, I grabbed an A10-5745M sample that was somewhere in the middle of the performance pack. The A10-5745M is a 25W chip with a 2.1 GHz base/2.9 GHz boost frequency. You can see above what I believe to be the 19W A10-7300 achieving roughly the same multi-threaded score, but ~10% higher single threaded score, than the 25W Richland chip. While these results aren't earth shattering, they are encouraging as it shows the same, or higher, performance despite a lower-advertised TDP. Note there are no third-party power consumption benchmarks yet, so it's hard to make any final judgment calls.

So although I may not have any power consumption benchmarks yet, press reviewers have performed limited testing of the upcoming FX-7600P (35W rated TDP) version of notebook Kaveri. And most of the testing performed seems to be very positive, with the caveat of not knowing how well the lower-rated wattage chips will stack up against Intel ULV chips. But to look at a quote from Hot Hardware, for example:

However, with what we've seen thus far, for full-sized notebooks and thin and light devices, the AMD Kaveri FX mobile chip we tested can hang with the best Intel has to offer on the CPU side with graphics performance that Intel's ULV processors can't currently match.

Regardless of the power consumption differences, the 35W Richland chip was often outperformed by 15W Intel chips.

Source: AnandTech

Note the A10-5750M is a 35W chip, and the i7-4500U one of the highest end 15W Haswell ULV parts. We see a sizable lead for Intel regarding both single threaded and multi-core performance in Cinebench R11.5.

Click to enlarge

Click to enlarge

Click to enlarge

Source: AnandTech, HotHardware

While there is some variance in the numbers obtained by HotHardware and AnandTech, rather than digging deep into the benchmarks and trying to extrapolate some fine-grained conclusions, rather notice that AMD is finally somewhat competitive on the CPU side with Intel; a place the company hasn't been in quite some time.

Intel still retains a fairly sizable single-threaded performance lead, and single-threaded performance is an important metric. But in multi-threaded performance, AMD's new FX-7600P parks itself firmly between the i5 and i7 ULV parts. Regarding the GPU performance, the FX-7600P nearly doubles (or sometimes triples) the performance of the Haswell ULV chips (source: AnandTech).

From a performance standpoint, Kaveri looks like a decent improvement in the notebook space over Richland. I suspect Intel will still have a power consumption advantage here, but I believe Kaveri has improved AMD's standing in the space overall.

The Budget Angle For Kaveri

The one interesting proposition for Kaveri is the knock-on effects from utilizing an AMD APU. OEMs that choose to use Kaveri have the option of adding in a discrete graphics chip that can be used in a dual graphics configuration. The discrete graphics chip can be used in tandem to increase overall performance.

Dual graphics setups utilizing desktop Kaveri was a mixed bag of results, showing anywhere from a decent performance boost to a slight performance regression. In the capacity of desktops, I would argue that dual graphics aren't a compelling reason to use Kaveri, as a DIYer could easily build a PC with a cheaper Athlon and higher powered GPU to get a better overall experience. Dual GPU setups aren't typically as stable or desirable as single card solutions, provided a consumer can achieve the desired level of performance from a single chip.

However, in the notebook space, the fact that you can't really build-to-order a platform from scratch severely limits consumer choice.

Click to enlarge

Amazon shows two different designs that offer interesting design choices. Above is a unit featuring an R7 M265 discrete GPU paired with an A10-7300 Kaveri for under $600. Lower on the product stack is a $400 version of the same model with an A8-7100 quad core Kaveri. These price points allow AMD to reach gamers on a budget at much lower price points than solutions from Intel.

Clarifying Potential Muddy Waters Surrounding Semi-Custom

Thanks to Mr. John Byrne and Mr. Drew Prairie from AMD clarifying and responding to inquiry.

Some of the biggest news, starting around December of last year, has been the potential for up to two semi-custom design win announcements from AMD. Seeking Alpha has a link to the Raymond James transcript, while the Credit Suisse and BMO transcripts are pay walled at various sites.

I listened to all three of these last year, and between them it was a little easier to paint a complete picture of semi-custom. Essentially, based solely on the transcripts, AMD was stating there was a strong potential for one to two additional semi-custom design wins in the works, but the deals had not yet been finalized (and still haven't). The deals would be in the neighborhood of $250M to $500M "over the lifetime" of the products.

I put "over the lifetime" in quotes, as this was close to the original wording during the BMO conference. However, during an interview on Barron's with AMD's CTO Mr. Mark Papermaster and Mr. Saeid Moshkelani, a VP in the semi-custom business, fleshed out more details regarding the semi-custom pipeline. Tiernan Ray, the author of the piece on Barron's, did an excellent write-up, so I highly recommend reading it. Some of the details brought out in the Barron's interview didn't completely jive with what was put out at the earlier conference, so I reached out to Mr. John Byrne and Mr. Drew Prairie for clarification.

After a few email exchanges, I came away with a much better understanding of what I was perceiving as discrepancies between statements made during the BMO conference and the Barron's interview.

First and foremost, Mr. Byrne confirmed to me his earlier statements. From our email exchanges:

...That said, as we talked about on our first-quarter financial results call in April we are working to close one to two more semi-custom wins this year in the $200 to $500M lifetime revenue range...

The following explanation is my understanding of the semi-custom business model based on email exchanges between myself and Mr. Prairie, supplemented with information from the Barron's interview.

When I asked about the "over the lifetime" vs. the seemingly implied annual revenue streams mentioned during the Barron's interview, Mr. Prairie pointed out to me that the 3 to 5 years mentioned is more of a generality than a hard and fast guideline.

The following examples are purely hypothetical. Imagine a semi-custom data center design win compared to the next generation consoles.

I had penned an article several months back regarding HEVC (high efficiency video coding). HEVC was designed so that it doesn't need a ton of processing power to work with, but in the data center efficiency is king. So let's assume that Netflix (NASDAQ:NFLX) decided to build out some hosting centers using a semi-custom APU from AMD using HSA-enabled ARM Holdings (NASDAQ:ARMH) A57 cores and some sort of hardware accelerator for HEVC encoding. Netflix may decide to build out several data centers in 2015, but none in the following years. This would mean the "$200M to $500M over the lifetime" would all apply in a single year, rather than spread out over several years.

Now consider Sony's (NYSE:SNE) Playstation 4. The PS4 is expected to have around a 5 to 7 year (give or take) life cycle, so this semi-custom design win will represent a varying amount of revenue spread over the 6 years or so we're all buying PS4s.

The biggest take away I had regarding semi-custom is the dynamic nature of the design wins. Review the Barron's article I linked to, and you'll find numerous examples of where semi-custom could make sense: handheld gaming, data centers running specific workloads, wearables, etc. And if you think about each of these various potential markets, you can understand why there are going to be a lot of moving pieces and parts.

Between the recent conferences (I, II) and the Barron's interview, the picture is much clearer for semi-custom than it has been. CFO Mr. Devinder Kumar explained that pushing out AMD's debt gives the company more flexibility. To quote Mr. Kumar:

And if you pull off the second part of it, of extinguishing the 2017 debt tower to the tune of $452 million, I could stand here and say to you that for five years we are virtually debt free and that allows us to go pursue the very transformation in about as aggressive a manner that we could.

The entire focus of AMD's management team has been to create business units that can profitably stand on their own. Since this article is specifically centered around semi-custom, consider the 500 person team that Mr. Moshkelani has at his disposal. He mentioned that $100M would be an annual revenue target could be considered a "floor" for when AMD would consider attacking a semi-custom design win. Alleviating near-term debt concerns could allow AMD the flexibility to expand operations, but it seems management is focused on keeping OPEX tight.

So rather than trying to pick apart wordings from interviews and conferences, I am taking a more aggregate view. These semi-custom design wins will be confidential, so it is my belief that AMD is being intentionally vague regarding specific numbers, time frames, and revenue amounts from potential deals. Semi-custom also only makes sense when there will be a larger volume of sales involved, and will have to either leverage operations that are graphically taxing, or lean on AMD's ability to integrate third-party IP on die.


Regarding mobile Kaveri, I want to ensure my words are not mistaken. If you read the above and come to the conclusion that I am overly bullish on Kaveri, I think you are missing my point. If you look at my previous articles, I am bullish more in the frame of reference that the bear thesis seems to conclude AMD's PC sales will nearly implode. From an earlier article, I stated the following in a conclusion:

Rather, I'm taking a more metered approach. If bearish analysts are expecting CS revenues to contract, what happens if CS revenues don't fall? No growth CS represents a positive if the bar is set for falling revenues.

Am I expecting AMD to come in guns blazing and turn the notebook market on its head? Not a chance, in my opinion. I expect Intel will retain its lead in efficiency and single-threaded performance. I expect Kaveri will extend AMD's current lead in gaming and graphical workloads. But Kaveri is moving the ball down the field in terms of overall performance and graphics. If AMD's power consumption claims stand up to third-party testing, it will also mean AMD is making strides in efficiency as well.

Kaveri offers OEMs some pretty interesting design choices, as the on chip graphics isn't necessarily wasted if it is paired with a discrete graphics chip. Again, dual GPU solutions aren't really ideal, but the restrictive nature of the notebook market, in my opinion, makes these solutions more attractive than in the desktop market. This provides OEMs with incentive to use an AMD APU, and also allows AMD to sell two chips to a consumer rather than just one.

Moving on to semi-custom, the biggest takeaway is that the nature of the deals is going to be highly dependent on the type of design win. The lifecycle of the design win is a moving target, but given a limited budget, AMD is trying to go after design wins that fiscally make sense. Hopefully we get some more color on the potential deals in the pipeline during the next earnings report.

Disclosure: The author is long AMD. The author wrote this article themselves, and it expresses their own opinions. The author is not receiving compensation for it (other than from Seeking Alpha). The author has no business relationship with any company whose stock is mentioned in this article.

Additional disclosure: I own both shares and options in AMD and may add to or liquidate shares or options at anytime. I also own a small number of near term puts that I may add to or liquidate at anytime.