AMD Bets Big On CPU-GPU Fusion In 2009

ChannelWeb has offered its own series of predictions for 2009, covering everything from chips to networking to storage. But what do the real insiders think is going to happen in the coming year?

We asked three top executives at Advanced Micro Devices to break out the crystal ball and share their prognosticative insights with us. The upshot: Sunnyvale, Calif.-based AMD is betting its unique one-two punch as a leading maker of both x86-based central processors and cutting-edge graphics processors will really start to resonate in the market, that virtualization will go mainstream, and that general-purpose GPU (GPGPU) computing will move sharply away from proprietary interfaces.

Dirk Meyer, president and CEO:

The industry is embarking on the visual computing era, and computing trends in 2009 will amplify the role of graphics in delivering the enhanced experiences users crave. The next wave of industry innovation will come from the fusion of computing and graphics technologies. As the only company in the world shipping both x86-based CPUs and leading-edge graphics, AMD is in a leadership position to deliver the best end-user experiences at home, work or play.

Nigel Dessau, chief marketing officer:

Virtualization will jump the chasm from large enterprise projects to the mainstream. In addition, we should expect customers to demand multiarchitecture or heterogeneous solutions that enable live migrations between PC suppliers, chip and operating system types. This is the year where we will all finally agree: the speed of the processor is no longer the most significant factor in defining the experience of the user.

Rick Bergman, senior vice president and general manager, graphics products group

A transition from proprietary to standard interfaces for GPGPU is ready to occur in 2009, as the first generation of proprietary APIs is quickly replaced by industry standard interfaces such as Havok, OpenCL and the compute shader features of DirectX 11. This will be a natural and much-needed evolution, allowing GPGPU acceleration to fully penetrate mainstream programming environments, and further unlocking the massive floating point computation of GPUs for processing general purpose applications from the desktop to the data center.

History has shown that proprietary programming interfaces are almost always replaced by freely distributed industry standards. This is because industry standards typically encourage greater and more widespread programmer productivity and innovation, while also enabling programmers to address the largest possible hardware installed base.

And with a DirectX 10.1 hardware install base well into the tens of millions, we also expect to see an uptick in games using that API, as more visually complex games and an increasingly competitive game software landscape create the need to unlock advanced AA techniques and additional "free" game performance.

Server Wars: AMD`s Opteron VS. Intel

In the battle for 45 nanometer dual-process supremacy, AMD’s Opteron 2384 is pitted against Intel’s X5492 Xeon. The winner comes down to the classic factors of power consumption vs. performance.

With the launch of “Shanghai”, AMD has revamped their Opteron product line to do battle with Intel’s Xeon in the multi-cpu server space. Does Intel have anything to worry about?

Server CPUs tend to be unique beasts. Sure you can build a server with a desktop CPU, but once you move into the multi-processor space, things begin to change.

Intel has dominated that multi-processor server (and workstation) market with its 45 nanometer Xeon line, especially in the lucrative and large dual processor segment.

Recognizing Intel’s success, AMD recently launched its latest Opteron CPUs, code named Shanghai, to compete with Intel in dual-processor servers.

Intel would win hands down if winning in the dual processor race was based purely on speed. However, there are many more factors to consider when building dual-processor servers, including power usage, performance per watt, heat generated and, of course, price.

In the dual-processor server market, Intel’s 5400 series of Xeon (Harpertown) CPUs seem to be the most popular. Xeon 5400 series processors are available in several different models, ranging from the Quad Core E5405, which runs at 2Ghz and has a thermal design power (TDP) rating of 80 watts, to the X5492, which runs at 3.4Ghz and has a TDP of 150 watts.

AMD offers several models of the Opteron (45nm Shanghai) CPU for dual-processor systems, ranging from the 2376, which has a clock speed of 2.3GHz, to the 2384, which has a clock speed of 2.7GHz. All third generation Opteron (45nm Shanghai) offer a 4x512KB L2 cache, 6 MB L3 Cache and have an average CPU power of 75 watts.

Comparing the best of dual-processor servers meant pitting Intel’s 3.4Ghz X5492 Xeon against the AMD 2.7Ghz Opeteron 2384. For a direct comparison, we turned to the SPECjbb2005 benchmark, which scored the Xeon X5492 at 324,451 and the Opteron 2384 at 311,471–showing that the Intel CPU has a speed advantage, at least when it comes to synthetic testing of elements such as the JVM (Java Virtual Machine), JIT (Just-In-Time) compiler, garbage collection, threads.

AMD’s Opteron 2384 has a average retail price of about $1,050, while Intel’s X5492 Xeon retails for $1,700. That brings up an interesting point, how much Intel processor can $1,050 buy, and how would that compare with the Opteron 2384? The Intel Xeon E5450 sports a price of about a $1,000, and tones the specs down a little bit compared to the X5492. The Xeon E5450 is an 80 watt quad core 3.0Ghz CPU and uses significantly less power than the X5492 (150 watt), which translates to less heat generated and lower electric bills. The E5450 is priced close to the AMD Opteron 2384 and has comparable power ratings (80 watts for the Intel, 75 watts for the AMD).

That said, initial performance comparisons gives the edge to AMD with the Xeon E5450 offering a SPECjbb 2005 rating of
293,213, slightly behind AMD’s 311,471. Simply put the
Opteron 2384 offers about a 6 percent performance increase over the Xeon E5450.

Of course, Specjbb2005 is only one bench mark, when we switched over to PassMark to score the CPUs, we came additional ratings which mirrored the previous tests. PassMark’s Performance Test offers a “CPUMark” rating score, which focuses on testing the CPU performance– here a pair of Intel Xeon E5450 CPUs scored a CPUMark of 9025, while a pair of AMD Opteron 2384 CPUs scored a CPUmark of 9449, roughly a 5.5 percent performance advantage.

So what exactly does all of this mean for the system builder? The quick take is that despite the launch of new processors and changes in technology, little has changed in the market. It all comes down to two elements: speed vs. cost.

For system builders looking for the most bang for the buck, a dual-processor server using AMD’s 45nm Opterons is the way to go. Those looking for max performance, the Intel Xeon proves to be the top dog.

There are some further complications coming down the pike. Intel will be expanding upon the technology that makes it latest “Nehalem” CPUs impressive performers, while AMD is sure to learn a thing or two when their new Phenom II CPUs hit the streets early next year.

With the new Opterons, AMD has proved that it can play “catch-up” with Intel, and still keep the market competitive. Simply put, AMD has put the ball is in Intel’s court for the dual-processor server market. Don’t count Intel out, though, since it’s sure to mount a strong reply to the third-generation Opteron.

Multicore doesn't mean equal core

As anyone who has worked on a group project knows all too well, not all team members contribute equally to the success of a project. And now Virginia Tech researchers have found the same holds true for the cores in multicore processors.

Depending on how your code is distributed across seemingly identical cores, the speed at which that code is executed on a multicore processor can vary by as much as 10 percent.

If you've ever had a program perform slower than expected or perform quickly on one day and not as spritely the next, you might want to examine how that CPU is executing the job.

"The solution to this is to dynamically map processes to the right cores," said Thomas Scogland, a Virginia Tech graduate student who summarized this quirk at the SC08 conference in Austin, Texas, last month. Scogland and fellow researchers, with help from the Energy Department's Argonne National Laboratory, developed prototype software that could one day help balance performance more equally across all cores. DOE also helped fund the work.

In the past few years, Advanced Micro Devices and Intel have developed multicore chips as a way of boosting performance over previous generations of commodity microprocessors. They have moved from two cores to four or even eight cores per chip.

Developers and systems engineers have mostly expected each core in a multicore processor to have the same effective capability. However, that is not necessarily the case for a variety of reasons, Scogland said.

In all fairness, it's not the cores' fault, technically speaking. Although the cores are identical, how a program is distributed among the cores can affect how quickly it runs. And in most cases, the operating system and hardware spread a program across multiple cores rather arbitrarily, which leads to varying performance.

A number of factors contribute to that variance, the researchers said. One factor is how the CPU hardware handles interrupts. In many cases, they could be directed to a single core, which could slow other applications on that core. However, if the interrupts are distributed across all the cores dynamically, there is no guarantee that the core handling the interrupt will be the same one that is running the program for which that interrupt was intended. Therefore, additional communication time is needed between the two cores.

AMD Phenom II 940 review from ocw

Crysis Warhead with 9950 and 4870x2 test

AMD 9950 unpacking

I found it on youtube for someone who wonder is there a heat sink in 9950 black edition.

How ATI stole Christmas

Yesterday we talked about Nvidia's Christmas lineup, bargains, or rather lack of them, here. Today it's ATI's turn, and things are looking good for the Canadian outfit, although considering what it's up against, this is hardly a surprise.

Let's start cheap and work our way up the ladder. At €33+ you can get the HD4350, while HD4550 prices start at €46. Don't let the numbers fool you, both are based on the RV710 core and differ in memory size and type (256 MB DDR2 vs. 512MB DDR3). While we can recommend the HD4350 for undemanding users, or HTPC builders, the HD4550 doesn't look like a good deal. Here's the catch, for just seven euros more you can get the HD4650, an RV730 card with 512MB of DDR2. The DDR3 version of the RV730, branded HD4670, costs €69.

In the same price range you'll find ATI's top value card of yesteryear, the HD3850. The cheapest one comes from MSI, and sells for €56. It's overclocked, but has just 256MB of memory. If you need more, you can get Sapphire's HD3850 with 512MB card for €69. HD3870 cards sell for €87+, and non-reference versions still cost over €100. Nvidia's G92 based 8800/9800GT cards have slipped into this price range, not to mention the HD4830, and both offer better performance. An overclocked dual GPU HD3870 X2 from MSI costs €174. It still packs quite a punch, and the price is reasonable.

So much for the RV670, let's move on to Nvidia's nemesis, ATI's RV770 based cards. The lineup starts with the HD4830, probably the best bang for buck card on the market at the moment. The cheapest ones come from HIS and Sapphire, and sell for €99. In our tests the HD4830 managed to outperform an overclocked 9800GT, not to mention the old HD3870, so even this crippled RV770 card packs quite a punch. As for the HD4850, it currently sells for €122+, a fair price, while the cheapest HD4870 costs €189. The latter seems a bit overpriced, especially following Nvidia's revamp of the GTX260 216 and the launch of new 180 Forceware drivers.

The R700 was never meant to be a budget card, but in recent weeks it's gone up in price, not down. Thank the weak euro and the bankers who led us straight into a recession for that. The cheapest one now costs €418, whereas three months ago it was selling for €30 less. The good news is that someone in ATI finally realized that the HD4850 X2 has potential, and what's more, they also realized that it can't be sold for €350. This, along with the two month delay, was the trouble with this rare card from the start. It was supposed to be a more sensible version of the R700, but it ended up costing almost as much as the regular one. Not anymore. As of today you can get the HD4850 X2 for €271 and it's finally starting to make sense.

ATI truly managed to steal Christmas this year, from the low-end, through the mid-range, all the way to the top, it dominates the market with affordable, yet powerful products. A year ago, Nvidia couldn't even meet demand for its G92 based cards, forcing ATI to fight for scraps, but the tables have turned and we can only hope Nvidia makes a comeback soon