AMD has ramped up its 45 nanometer chip making process, delivering lower-voltage and higher-clock-speed variants of its "Shanghai" Opteron processors. These Shanghai chips are coming to market just when AMD said they would, representing a tiny bit of good news in a twelve-month cycle that has been pretty hard on AMD, its partners, and its shareholders.
There are five so-called Highly Efficient variants of the Shanghai Opterons arriving today. These parts run at slightly lower clock speeds, but they have dramatically lower power consumption. Each is rated at 55 watts using AMD's own ACP scale. ACP is short for Average CPU Power, and it's distinct from the peak power consumption metrics that AMD used for many years. A standard Shanghai part has an ACP rating of 75 watts.
AMD has also released a Special Edition Shanghai chip, which comes in one flavor for two-socket servers and another for four-socket boxes. It revs at a slightly higher clock speed, but eats up a lot more power. It's rated at 105 watts. Considering the price premium and extra heat in the SE parts, you really have to need that extra oomph to pay the extra dough. Very few people do this, by the way.
For two-socket boxes, the Opteron 2372 HE part runs at 2.1 GHz and costs $316 in 1,000-unit quantities. The 2374 HE runs at 2.2 GHz and costs $450. And the 2376 HE runs at 2.3 GHz and costs $575. For bigger boxes, there's an 8374 HE running at 2.2 GHz that costs $1,165 and an 8376 HE running at 2.3 GHz that costs $1,514. The topper 2386 SE part for two-socket servers runs at 2.8 GHz and costs $1,165, while the 8386 SE part for four-socket boxes (running at the same speed) costs $2,649.
Whether they are standard, HE, or SE parts, all of the Shanghai chips have the same basic configuration. Each of the four cores on the chip has 512 KB of L2 cache memory and the cores share a 6 MB L3 cache (which is three times the cache on the earlier and buggy Barcelona quad-core Opterons that gave AMD grief a year ago). The chips also support 800 MHz DDR2 main memory. The original Shanghai parts were announced back in November.
While getting the HE and SE parts out on time is good news for AMD, many people had been expecting the 45 nanometer ramp to allow AMD to hit a high clock speed, perhaps even 3 GHz or higher. A standard Shanghai chip with a 75 watt thermal envelope runs at 2.7 GHz. The Opteron 2384 at 2.7 GHz costs $989, and the 8384 costs $2,149. So you pay 17.8 or 23.3 per cent premium for that extra 100 MHz of performance, but you have to put up for 40 per cent more power consumption and heat dissipation from the chip when you only get 3.7 per cent more oomph.
If you look at it based on the price scale AMD uses, a Shanghai SE part should run at something closer to 3.2 GHz - assuming a linear relationship between performance and price (which chip vendors do not, particularly at the high end of their lines). Based on the thermal penalties and a linear scale, you would expect a 105 watt chip to run at something closer to 3.8 GHz. And by the way, such a chip would crank out a lot more heat than this in the real world because heat dissipation is on an exponential curve, tied to clock speeds. As you can see from the SE chips above, a little more oomph can mean a lot more heat.
Given the heat and cost, it seems unlikely that the Shanghai SE parts will get much traction in the market. They just aren't worth it. Faster parts at the same prices and still within the 105-watt thermal envelope might, and hopefully, AMD will push its yields and get these out. But AMD has been pretty clear that it is focusing more on standard and HE parts these days.
John Fruehe, director of business development for server and workstation products at AMD, says that regardless of the adverse economic conditions, the Shanghai ramp is "going quite well" and that AMD can still get a piece of the IT budget. "A typical data center has two buckets of money - one for projects and one for upgrades," Fruehe explains. "We typically plan for upgrades to be about 5 per cent or so of the market, but with the opportunity to get more performance in the same server, and in a way, that is a lot less expensive than buying a whole new server."
Fruehe also hints that there is plenty of room in the 45 nanometer process ramp, so more speed bumps will be announced in the coming quarters. And that is good news for AMD, since the company believes that Intel will have to focus on selling "Harpertown" Xeon systems as it preps the market for its next-generation "Nehalem" architecture. "Across the board, Nehalem looks like it will draw down more power than Harpertown, which is exactly the wrong direction from the one that the market is demanding," he says.
The Nehalem server launch is expected around the end of March, according to the latest murmurings on the street. But it is possible that this gets pushed out a quarter given the state of the economy and Intel's own inventory and sales woes. As we previously reported, Intel is apparently delaying a variant of Nehalem called "Lynnfield" and its P55 chipset. Aimed at mainstream PCs, the platform is now expected some time in July. The delay seems to have less to do with any technical issues than with excess inventory of older kit and a PC channel that is in no mood to buy more parts. ®
Sponsored: Webcast: Simplify data protection on AWS