|The Leading Source for Global News and Information Covering the Ecosystem of High Productivity Computing / September 14, 2007|
After talking up its quad-core processors for more than a year, AMD finally released them into the wild this week. On Monday, the company kicked off a public relations extravaganza that celebrated the global launch of its new quad-core processors, which included events in Bangalore, Barcelona, Beijing, San Francisco, Seoul, Taipei and Tokyo.
Outmaneuvered by its larger rival over the past year, AMD has seen its worldwide server market share shrink from about 25 percent to 13 percent and has witnessed its bottom line bleed into the red. Since Intel's quad-core offerings have been available for almost a year, AMD has been under increasing pressure to deliver its quads into the market. AMD's six-month delay in getting its chips ready has given Intel the opportunity to romp freely in the x86 quad-core server space. AMD is hoping the new Opterons will not only stop the hemorrhaging, but will also start to reverse Intel's momentum.
"Frankly it's the biggest thing since the original Operton launch," said Steve Demski, AMD Opteron Product Manager, Server Workstation Division. "In fact, it might be even bigger than the original."
When AMD introduced the new processor back in 2003, there weren't a lot of opportunities for customers to purchase Opteron-equipped computers. Today, Opterons are found on 55 Tier 1 platforms as well as additional regional platforms worldwide. From that perspective, the new quad-cores represent a much larger opportunity for AMD than its offerings in years past.
The initial launch includes nine offerings: five two-socket processors (Opteron 23xx), which range from 1.7 GHz to 2.0 GHz in speed, and four four-socket processors (Opteron 83xx), which range from 1.8 to 2.0 GHz. They are being priced aggressively against their Intel counterparts. For example, the Opteron 8350 is listed at $1019 versus the comparable Xeon 7320 at $1177. Since HyperTransport and the integrated memory controller gives AMD a performance edge at the four-socket and above scale, the high-end Opterons should be very competitive. In the more mainstream two-socket market, the Opteron 2350 is listed at $389; its Xeon 5345 counterpart is listed at $455. Intel could decide to lower its prices now or wait for its 45nm Harpertown processors (due in November) to get a better price-performance position against the new Opterons. The quad-core competition is just getting started.
Besides just putting four cores on a die, AMD has made a number of other improvements in the design. The AMD engineers accelerated the floating point unit significantly. According to Demski, they've quadrupled raw FP performance compared to the previous generation. This was accomplished by doubling the FP pipeline from 64 bits to 128 bits and doubling the data and instruction fetch bandwidth. They've also added support for misaligned SSE operations, which should save a clock cycle on some instructions.
The on-die memory controller has been completely redesigned. The new controller has deeper buffers to play with, is able to prefetch into L1 cache instead of L2 cache, and has incorporated an optimized DRAM paging scheme. In addition, the 128-bit memory controller can be configured via a BIOS setting into two 64-bit channels that operate independently. This enables an application to read and write data in parallel. Overall, AMD says they've achieved 50 percent better memory bandwidth in the new quad-cores compared to their dual-core brethren. And they're going to need all that bandwidth, since there are now twice as many cores contending for memory.
AMD also put a lot of thought into minimizing energy consumption by taking advantage of the "native" quad-core implementation. For example, the clock frequencies on each of the four cores can be adjusted independently, based on an individual core's workload. The new chips also make much more aggressive use of clock gating, so that areas of the chip can be disabled when not in use. In addition, AMD has split the power supplied to the memory controller from the power supplied to the cores. This enables a 200 MHz increase in the memory bus clock, which translates into better memory performance. It also allows for better granularity of power management since it decouples memory needs from computational needs. For example, an application running in a two processor system, but which only requires the cores on one CPU, is able to access memory on the other CPU, with those processor's cores effectively turned off.
To go along with the new emphasis on energy conservation, AMD has decided to change the way it measures power consumption. In a nutshell, AMD is switching from a Thermal Design Power (TDP) metric to an Average CPU Power (ACP) metric. AMD's TDP will still be defined as the maximum theoretical power used by the processor, and is only intended to be used by system engineers for determining thermal design limits. ACP will represent the maximum power the chip will draw under real-world conditions. It's derived by running high utilization workloads such as TPC-C, STREAM, and various SPEC benchmarks.
Looking at the quad-core offerings AMD introduced this week: the standard processors have a TDP of 95 watts, with an ACP of 75 watts; the low-power processors -- what the company calls its high-efficiency line -- have a TDP of 68 watts and an ACP of 55 watts.
AMD claims that ACP is similar to Intel's TDP. And while Intel's TDP is almost certainly not equivalent to AMD's TDP, it's not clear if Intel's TDP is equivalent to ACP. Confused? For the end-user, the best metric will be plugging the system into the wall, running the applications, and measuring the power usage.
AMD provided some benchmark comparisons for a number of synthetic and real-world HPC-type workloads. Pitting its 2.0 GHz Opteron 2350 against the 2.3 GHz Xeon 5345, AMD reported that they beat the competition by anywhere from 7 to 189 percent, depending upon the workload. Independent testing from AnandTech (www.anandtech.com), using the same chip matchup, showed more mixed results. In these tests, the Xeon prevailed in the Linpack benchmark (using Intel's own Math Kernel Library), while the Opteron had a slight edge with a 3D rendering code. AnandTech's conclusion was that Intel and AMD are about equal in raw floating point power, but the Opterons have the edge when the application needs to access main memory a lot.
Neither set of tests was applied to the four-socket versions of the processors to be used in multiprocessor (MP) servers. Up until last week, Intel didn't have an MP solution on its latest Core architecture. But with the Tigerton Xeon introduction, it's now possible to match the 83xx Opterons against the 73xx Xeons. Because of the inherently better scalability of AMD's architecture, most analysts believe that the Opteron should still be the one to beat in multiprocessor server space. But for single-thread performance, the Xeons will continue to dominate.
Since Intel's next-generation Xeon Harpertown processors are expected to come online later this year, the longer-term success of AMD's quads will depend upon how quickly the company can ramp up clock speeds. "Special Edition" quad-core Opterons are already in the works and scheduled for release in Q4. Those CPUs will start at 2.3 GHz and go up from there. Even the standard and low-power processors are expected to get speed bumps before the end of 2007. If the quad-core Opterons can stay competitive on price and performance over the next six months, AMD should achieve the rebound it hoped for and then some.