AMD Needs a Multi-Core Killer App

After it released its first dual core Opteron for servers on April 21, 2005 followed by the Athlon 64 X2 for desktops a month later, Intel released its first dual core, the Pentium Extreme Edition. Around the same time, AMD splashed full-page ads in newspapers calling upon the bigger rival to join a duel to decide whose dual core was better. Intel declined, but several agencies did the testing and ultimately gave AMD the thumbs up.

In the case of quad-core processors, Intel is shipping them since November 2006, but AMD’s first quad-core Barcelona, unveiled only last month, is said to have a couple of features that Intel cannot match.

For one, Barcelona is a native quad-core. This means that it has 4 identical cores on a single piece of silicon, while Intel puts two dual-core chips side by side on a board, and connects them to form a quad core.

AMD says its design performs better. Intel plans to bring out its own native quad-core design in the H2/08, a good one year away.

The other unique feature of AMD’s Barcelona is that it has an on-chip memory controller. Experts say that this makes the processing faster than having the controller on a separate chip set, as Intel does it. Here too, Intel plans to put an on-chip memory controller on its devices starting in late 2008 or early 2009.

Superior technology is helping AMD slowly build on its market share. A study by Mercury Research shows that AMD’s market share stood at 22.9% for Q2/07 (76.3% for INTC) compared to 18.7% (80.5%) in Q1/07 and 21.6% (72.9%) in Q2/06.

For all that is apparently going well for it technically, AMD’s stock price doesn’t evoke cheer. It is in the midst of a long-term downtrend, and presently changes hands at the $12.80 level (52-week low $11.27) giving it a negative forward P/E of 5.73.

AMD’s Q2/07 results proved to be mixed. Revenue was $1.378 billion, up 12% sequentially and 13% y-o-y. However a net loss of $600 million translated to per-share net loss of $1.09 even as ROAE is markedly negative 49.74% against just –3.63% in the whole of 2006. Q2/07 gross margin (33.4%) improved from that of Q1/07 (28%), but was much lower than in Q2/06 (57%). AMD’s strategy of competing aggressively on price is certainly not one without consequences.

As for its future plans, at the CES 2007 AMD announced 8-core processing in H2/07, while a 16-core super-CPU codenamed Bulldozer is slated for H1/09 release.

AMD merged with ATI Technologies on October 25, 2006 after acquiring it for $5.4 billion in July. The fruit of the AMD-ATI combination is the announcement of AMD Fusion, a next-generation microprocessor design that aims to combine general processor execution as well as 3D geometry processing and other functions of today’s GPUs (graphics processors) into a single package. Fusion is expected in 2009.

The likely tremor that Fusion may generate is already apparent from Intel’s speedy snapping up of Dublin-based Havoc, a gaming middleware company, for an all-cash $110 million deal yesterday.

One wonders why isn’t AMD zeroing on Tilera Corporation. Tilera’s 64-core TILE64, that has started shipping only last month, is a futuristic multi-core design combining less consumption of power and higher performance. Skeptics question whether the level of processing power that Tilera promises to deliver is necessary in networking (embedded processors).

It is, however, evident that AMD and Intel have decided that the high-end microprocessor battle will be fought on the multi-core battlefield, so having an advanced R&D capacity such as Tilera would presumably give AMD a huge leg up against Intel in the mainstream microprocessor world.

At any rate, while addressing the multi-core R&D issue may be feasible, and even beating Intel at this game may well be possible for AMD, how it advances shareholder value remains a looming question.

One possibility that I see is for AMD to work with some key OS and application vendors to get to 16-, 32-, 64-core processors faster than Intel. In doing so, the key would be in porting their software to the advanced multi-core chips, optimizing them to be able to take true advantage of the processing power, and thus, creating compelling reason for consumers to want machines with high-end multi-core chips. As it stands, most of the software in the market cannot take advantage of multi-core chips, so why buy machines that boast to have them and pay the extra price?

In summary, AMD needs a killer app vendor partnership. Intel seems to have identified graphics middleware as one of their core software strategies. What is AMD’s killer-app alignment plan? Without understanding this, I just can’t get excited about AMD’s prospects as a company.

Also, are we about to see the transformation of hardcore chip companies into companies that also do software?

Disclosure: Author has no positions in stocks named.

Comments

[DEcho:]

I think the article had a too strong bias towards AMD. The general consensus in the industry is that AMD is struggeling real bad, see f.ex. theinquirer.net/?artic... - the current situation is clearly not sustainable. Intel is executing better than ever, while the opposite can be said about AMD.. It's partly reflected in stock price, but I think the divergence between AMD and INTC stock price to widen. For one, the article does not mention AMDs execution problems, with delays and poor performance of their Barcelona line, while Intel is actually one generation ahead in terms of process technology (better technology, better margins)

2007 Sep 22 12:15 PM

[Scully:]

I agree with the first poster.... you give way to much credit for technical leadership.

When I read an article as such, it is typically with a measured lack of detailed knowledge behind the computational inner workings of a CPU. Before going into detail on this topic, just a recount of history. In 2003, AMD released the K8 core which possessed two features -- an integrated memory controller to connect direct to memory and hypertransport, a serial connection technology to interconnect periphrial devices. Innovative as they were, they form only one part of the entire computational puzzle. Nonetheless, Intel's decision to continue pushing a clock-scalable architecture, Netburst, resulted in AMD holding a decisive performance lead for desktop and server (Intel pioneered a mobile architecutre, that at it's base was actually superior the K8 -- that data is, in scant quanity, available around the Net). It was the performance lead that AMD began gaining share, winning design wins, and become more successful. One may argue that AMD's success was more a poor decision of Intel, nonetheless... the interconnect technology was only on part of the equation.

In July 2006, Intel released the Merom series processor -- spanning desktop, server, and mobile. This processor, even on the aging frontside bus and lacking an integrated memory controller proved far superior in performance to any AMD offering, where AMD only could win in the 4 socket server space (the area where their interconnect technology really, truly makes a substantial difference). As a result, of an agressive pricing of the new products, AMD had to adjust to be price/performance competitive with a computational inadequate performance curve. We now the result, almost 1.8 billion lost in the last three quarters.

Why, then is it that an Intel architecture can out perform similarly clocked AMD processors 20-30%, and top bins are holding a commanding lead? AMD would have you believe that the IMC and HT inteconnect make the processor, something many, including the author of the article, have bought. Intel contains superior core architecture, a superior caching system that renders BW in most situations irrelevant.

Alas, Intel produced a quad core part > 10 months ago, it prooved scaling was achieved in applications that could multithread, commensurate with the increased core count... why? Again, the bandwidth/IMC argument has been rendered irrelevant with a well designed caching technology and incremental increases in the performance of their own bussing technology. The sales pitch that 'native' is better is odd... intuitively it should be better, what AMD does not tell you is by how much. Say at a single core level Intel performance is 30% better at given clock speed, but in quad core performance that scaling achieves roughly 0.8x gain for Intel and 0.82x gain for AMD, that 'native' scaling gain will not be enough to overcome the raw performance over 4 threads no matter how native that product is....

In short, there is no such thing as a needed killer app for AMD.... their killer apps are clear... HPC scientifically intensive programs that actually push the BW argument and, as the data, shows... is where they show a very competitive position. Unfortunately, this is not the volume market and, though the margins for this niche product is good... the volume will not be high enough to pay the bills.

I would strongly, strongly encourage the author of the blog to rethink his position... and rather than take at truth the marketing hype from either company, take a moment to educate yourself on a) how a CPU actually works, the impacts of large vs small caching systems and b) simply study the emperical performance data.... it is quite telling and your faith in one company will most certainly be shaken.

2007 Sep 24 12:58 PM

[User 108694:]

A few short points.
1) The Intel Core architecture is far superior to the much newer Barcelona architecture (see benchmarks on any site, such as ananadtech.com). AMD do not hold the price/performance crown and won't for the forseeable future (1+ years).

2) One of the reasons for AMD's poor stock performance is that in the desktop scene, they are being trashed by higher performing better value chips (see point 1).

3) 90% of computer users will see benefits to moving to a dual core processor. Assuming that 0% of applications are actually able to use both cores simultaenously, you'll still get a massive benefit by being able to run 1 intensive application *without* slowing down your system to a crawl.

4) 16/32/64 core CPU's are nigh on useless for the average joe on the desktop. Even 8 is well pushing the bounds of reason. There isn't much out there that can effectively utilise 4 cores, and there's even less out there that actually needs that kind of processing. 8 Cores is overkill and probably will be for at least the next 5 years. The only thing that would be likely to push 8 cores would be games, and games are only just getting into the whole multithreading scene.

2007 Sep 24 01:35 PM

[Peter Kazanjy:]

There already is a killer app for heavily multi-core physical machines: virtualization.

Just because you don't readily have an app that needs 100% of a quad core processor, let alone the four or eight quad core procs that can be stuffed in a box, doesn't mean there isn't demand for those boxes.

I work at VMware, and we had a demo box running about 120 virtual machines with a four-way quad-core box.

There are plenty of customers using quad core-based boxes to host all kinds of virtualized workloads, including hundreds of virtualized desktop XP sessions for thin-client-based end users.

You can slice it any way you want. Run burly database VMs each with two cores and a boatload of RAM assigned to them, or dozens of low impact file servers and domain controllers, or even more lightly utilized desktop VMs. It's all just granular compute once you've virtualized the box.

With massively multi-core boxes and bare metal virtualization like VMware Infrastructure 3, the future of x86 servers looks like high capacity mainframes of the past: massive chunks of fully utilized compute that can be purchased (capex) and run (management and power opex) at a lower cost per compute unit than ever before.

As Intel's investment in VMware, and AMD's strong partnership with VMware both show, the future of multi-core servers is tightly integrated with and enabled by virtualization in much the same way that desktop x86 computing was advanced through a Wintel combination wherein ever more demanding software fully utilized ever stronger desktop chips.

2007 Sep 24 01:38 PM

[Saturn V:]

sramana,
You seem to be a victim of the AMD propaganda machine. Specifically you state 'but several agencies did the testing and ultimately gave AMD the thumbs up.". the reality is that the relative performance was a mixed bag, depending on the applications used, AMD lost on performance in most cases. AMD only won in high bandwith,multi-sockete... applications,which is a small subset of the market.The loss in other segments was a major reason in AMD major financial losses for the last 3 quarters.

The reality is that AMD has gained a tiny market share by selling its products at a loss. Its Average Selling price in the High Performance Markets has collapsed, and it is reduced to a "value vendor".

Unfortunately the latest Barcelona chip is not the promised savior. It is losing even more performance benchmarks to Intel's Penryn chip family.

AMD has stubbornly refused to downsize, and the red ink continues to flow with no end in sight. A restructuring is needed, or AMD desperately needs some one to bail them out, someone who can stomach losing $500 million every quarter.

2007 Sep 24 07:13 PM

[CPU watchtower:]

AMD quad core does have some advantages: better performance on multi-threaded memory-intensive code and much less power dissipation when idling. One place where AMD will certainly help save on electricity bills is in data centers where many servers sit at idle throughout the night. In other areas, AMD lacks a performance edge. There are a couple main reasons for this. First, the L2 cache is too small. Barcelona's L2 cache at 512 KB per core is even less than that for the current dual core Opteron and far less than what is available on Intel CPU's. The other reason is that the Intel Core 2 architecture executes more instructions per cycle: up to 5 per clock cycle per core if you include the macro ops fusion feature, while Barcelona comes in at a maximum of 3. For AMD to win, it cannot stay one step behind. One has to wonder why AMD makes design decisions that leave it one step behind. It is even behind on clock speed. Its fewer pipeline stages makes it more difficult to introduce the product at higher clock speeds. Compare that to the Intel Tigerton or the IBM POWER6 which are available at high clock speeds immediately upon introduction. AMD will have to tweak more to get the frequency up (another delay).

2007 Sep 26 10:12 PM