At a press conference from its Round Rock headquarters this morning, AMD made good on one of its most important promised milestones: It's preparing to deliver its six-core Opteron server processors, which will beat Intel to market with the first drop-in six-cores for 4P and 8P configurations. Shipping, according to executives, begins now. Intel isn't expected to have Nehalem generation EP- and EX-class Xeon processors in configurations other than 2P until sometime next year.
As the company's server business director John Fruehe told reporters this afternoon, this new class of six-core Opteron will feature a new element of its HyperTransport bus, called HT Assist. As Fruehe explained, this feature will enable all the processors in a multi-way configuration to share portions of their L3 cache as a pooled lookup table. This way, calls to the table are directed to the appropriate processor, even across processors. The promise here is to dramatically reduce crosstalk and traffic, and cut stream memory bandwidth by as much as 30%.
*Update*
Over the past three years, since Intel's introduction of Core Microarchitecture, AMD has been in the habit of professing that it has been listening to its customer, and whatever it is that the customer needs happens to be whatever AMD presently produces.
Some days, the company might be right. The rest of the time, it's been something of a stretch, including the day two years ago, amid the worst days of Barcelona's lackluster performance, when AMD openly argued customers don't really want performance anyway. Certainly if there's an exception to that argument, it's in the server space, and AMD's July 2007 dance around that exception was now almost famously ineffective. This morning, AMD is in a position where it must produce a higher performance chip for every market segment, or else become relegated to the scrap heap of history along with Transmeta. So now that AMD is more comfortable with what it's seeing from its Istanbul product line, what is it that server customers want? Startlingly, something that looks a lot like Istanbul.
Source
As the company's server business director John Fruehe told reporters this afternoon, this new class of six-core Opteron will feature a new element of its HyperTransport bus, called HT Assist. As Fruehe explained, this feature will enable all the processors in a multi-way configuration to share portions of their L3 cache as a pooled lookup table. This way, calls to the table are directed to the appropriate processor, even across processors. The promise here is to dramatically reduce crosstalk and traffic, and cut stream memory bandwidth by as much as 30%.
*Update*
Over the past three years, since Intel's introduction of Core Microarchitecture, AMD has been in the habit of professing that it has been listening to its customer, and whatever it is that the customer needs happens to be whatever AMD presently produces.
Some days, the company might be right. The rest of the time, it's been something of a stretch, including the day two years ago, amid the worst days of Barcelona's lackluster performance, when AMD openly argued customers don't really want performance anyway. Certainly if there's an exception to that argument, it's in the server space, and AMD's July 2007 dance around that exception was now almost famously ineffective. This morning, AMD is in a position where it must produce a higher performance chip for every market segment, or else become relegated to the scrap heap of history along with Transmeta. So now that AMD is more comfortable with what it's seeing from its Istanbul product line, what is it that server customers want? Startlingly, something that looks a lot like Istanbul.
Source