AMD says new 'Shanghai' chip is ready to go
AMD said Monday it is set to roll out its next-generation "Shanghai" chip--minus the mistakes of the last generation.
The No. 2 processor maker wants to make one thing crystal clear: Shanghai is not Barcelona. The latter chip was rolled out in September 2007 to great fanfare only to be delayed a whopping eight months (or more, depending how the delay is calculated) due to production glitches and bugs. The chip was also hampered by speed (core clock frequency) limitations. This gave Intel an opportunity to regain ground it had lost to AMD in the server chip market.
"We had some mis-starts in getting Barcelona to market and wanted to bring as much velocity to Shanghai as possible. Learn from our mistakes and, as a company, never do that again," said Pat Patla, general manager of AMD's server and workstation chip business.
Shanghai--a quad-core product targeted at servers--will be AMD's first 45-nanometer processor. (Barcelona is 65-nanometer.) Typically, the smaller the geometries, the faster and more power efficient the chip. Intel has been shipping 45-nanometer processors since last year and these processors now make up most of Intel's offerings.
AMD needs Shanghai to succeed. It is reeling from a string of losses and is on the verge of announcing a major restructuring. "To bring it back to profitability the execution of the server product line is absolutely critical," said Ashok Kumar, an analyst at investment bank Collins Stewart. "That is really their only profit pool."
AMD's server chip roadmap includes a six-core Istanbul processor in Q4 2009
(Credit: AMD)To ensure this, AMD designated a lead engineer to take over the entire Shanghai project to establish goals for the "health of the silicon, the schedule for the silicon, and the confidence level of the silicon," Patla said. AMD had to make sure that "the product that we put in the hands of our partners is going to be of substantial stability so they can do lots of early validation," he said.
As a result, the schedule for Shanghai has been pulled in. "Originally the plan was that Shanghai would launch in Q1 of '09 and we were able to pull that into Q4," according to Patla, adding that the product will not only be announced in the fourth quarter but vendors will be shipping servers in the fourth quarter.
"We're in full production right now in the factory," he said. "People will start getting first silicon from the final production very shortly."
Patla asserted that Shanghai is a "very power efficient product" and will perform much better than Barcelona because the smaller 45-nanometer process yields "a lot more (clock) frequency."
At the same frequency (speed), Shanghai will outperform Barcelona by about 20 percent, Patla said.
AMD is also boosting the size of the cache memory, which typically speeds performance, from 2 megabytes to 6 megabytes. Another speed improvement will come from increasing "instructions per clock," Patla said.
"We're also turning on HT3 (HyperTransport 3) and you'll see partners start to validate that in the Q1 time frame," Patla said. HyperTransport is a high-speed communication link technology between silicon.
Shanghai will be followed by a 45-nanometer desktop processor code-named Deneb, which is due to launch in the fourth quarter of this year or first quarter of 2009, AMD said.
In the fourth quarter of 2009, AMD will add a six-core processor. "We'll take what we've learned from our 45-nanometer process and Shanghai core and bring out an Istanbul six-core product," Patla said. Like Shanghai, this will be targeted at servers with up to eight processor sockets.
Brooke Crothers is a former editor at large at CNET News.com, and has been an editor for the Asian weekly version of the Wall Street Journal. He writes for the CNET Blog Network, and is not a current employee of CNET. Contact him at mbcrothers@gmail.com. Disclosure.
Granted you can manually go in and change the affinity of an app to work on an individual core, but isn't it about time that software catches up with the hardware?
Why not work at increasing speed, lowering power and heat on the dual/quad core chips before moving into the 6-8 cores?
Several reasons.
1. This is for servers ... not desktop applications. Server software does consume lots of processes and/or threads. If you dozens, hundreds, or thousands of users interacting with your sever there certainly not just one application running on one thread on that machine.
2. Virtualization. Just because these CPUs may be in one physical machine that doesn't mean that there will be one OS and mostly one application running on top. Again if you aggregated several users onto one machine can find a steady stream of workload.
In short. one of the limiting factor on desktop machines is that most users primarily engage in only one activity at a time. If you are talking the "average joe" desktop market then perhaps can get by with just duals.
"Why not work at increasing speed, lowering power and heat"
Again servers. There is huge momentum in virtualization. For virtualization servers you need a "bigger" machine in a smaller and more cost effective box. By the way the article does seem to indicate that get a increase in speed and probably get a lowering of power since .45. ( heat and power are pragmatically the same thing . ignoring how 'leaky' the circuits are.) If you can increase speed and increase cores why not?
It won't take as long as you think. You don't have rewrite the software every time you double the cores. If you can write good multicore code it'll scale as you add more cores. Going from one core to two cores may need a rewrite. However, one you realize cores will keep increasing it won't take as many software changes to go from 4 to 6 or 8 as it did to go from 1 to 2. They just have to get over the first hump of teaching their software to work with many cores instead of one.
While this may be true in the desktop market, it is not at all true in the server market. Shanghai will first released as a server part under the Opteron name because the server market is their bread and butter. Server workloads are definitely optimized for multi-core processors, way more than desktop applications. A server could potentially sit above 90% utilization all day long depending on the application, while your desktop probably idles 90% of the time. Vastly different usage patterns.
On desktop you may use many applications paralel, so multicore will work fine there also.
-
by sang u kim
October 5, 2008 6:02 PM PDT
- AMD says its first 45nm chip dubbed Shanghai is not Barcelona, and will outperform Barcelona about 20 percent. However, it doesn?t mention what process technology is used in Shanghai chip, and its performance and power, particularly compared to Intel?s 45nm process technology that utilizes advanced high K HfO2 /metal gate stacks as detailed in IEDM 2007.
-
Reply to this comment
-
(14 Comments)Barcelona is PD (partially depleted) 65nm SOI based chip. There is no mention whether Shanghai is also based on PD SOI or FD (fully depleted) SOI or conventional bulk based oxy-nitride SiO2 gate or advanced bulk based like Intel 45nm. For Barcelona to be successful, it has to be equivalent or superior to Intel?s 45nm technology in device performance, power, reliability as well as profitability.
The Shanghai chip is one year behind Intel?s 45nm which is in high volume manufacturing since 2007, and AMD's 32nm might be one or more year behind Intel?s 32nm process technology yet to be announced, possibly late 2009 or early 2010. Intel is the only company today manufacturing high volume 45nm products. Doing so, Intel has gained critical learning and experiences in high volume manufacturing, circuit design and reliability required for easy transition to 32nm and beyond. AMD will face enormous challenges in getting back to profitability in high volume manufacturability race with Intel for the next generation technology development.