AMD unmasks Opterons of servers future
Faces for 'Magny Cours' and 'San Marino'
It's financial analyst day at chip designer and seller (but no longer wafer baker) Advanced Micro Devices, and that's reason enough for the company to divulge a few more details about its future Opteron processors and related chipsets and platforms, due early next year.
First of all, the official names of the processors. The "Magny-Cours" chips, which AMD has been chatting up as a game changer for its server business in 2010, will be known as the Opteron 6100 line, while the "Lisbon" processors, which will follow them to market a little bit later, will be called the Opteron 4100s.
The Opteron 6100s will be available in two-socket and four-socket configurations (known as the "Maranello" platform) using AMD's own "Fiorano" chipsets, which El Reg told you all about in September. The Opteron 6100s will are essentially tweaked versions of the current six-core "Istanbul" Opteron 8400 series chips that are equipped with DDR3 memory controllers on the die, plunked side-by-side in a single chip package, and plugged into a new G34 socket.
AMD plans to have versions of the Opteron 6100s with eight cores activated (with four of them being duds, presumably two per physical chip) as well as the full twelve-core complement, something it has been saying for nearly a year now.
Aside from the official name of the Magny-Cours chip, the big news - and the one that will allow AMD to get some traction in the server racket - is the big leap in memory bandwidth that AMD will show in the jump from Istanbul to Magny-Cours systems. According to John Fruehe, director of server product marketing at AMD, there is enough cores and memory bandwidth in the Opteron 6100 series that the company doesn't believe it has to create an eight-socket server. Indeed, the Fiorino chipsets only support eight-way configurations with the current Istanbul processors, but max out at four sockets with the Magny-Cours.
Here's why Fruehe is not worried about scalability. Using the Stream memory bandwidth benchmark as a guide, a four-socket Opteron 6100 box using the SR5690 I/O hub and SP5100 south bridge (which together have support for hot plug PCI-Express 2.0 peripherals and which offer 42 PCI lanes) is able to hit 100 GB/sec of performance.
That is not just a lot of memory bandwidth for a four-socket box, but compares favorably to the eight-socket servers on the market today using Istanbuls, which delivered a little north of 80 GB/sec on the Stream test; a four-socket Istanbul machine delivers just over 40 GB/sec today, and a two-socket server using Intel's "Nehalem EP" Xeon 5500s comes in at around 35 GB/sec.
Next page: QuickPath v HyperTransport
If economies of scale and price are the chicken and egg here, how about bringing back the CPU and memory daughterboard, using HTX3? Instead of making 1P 2P 4P and 8P boards, They could go for 1P 4P and 8P boards, or 1P and 2P for budget, and 8P for the rest.
The customer just populates with the number of CPUs they want.
Hell, AMD, if no one buys the 8000s because they're too expensive (and surely they cost near enough the same to make) - why not give all the Opterons the full number of links, reduce the number of different CPUs, and watch them fly of the shelves? Seed the market.
And while they're at it bring on the GPU HTX cards.
If you build it, they will come.
The price premium on 4P & 8P systems goes some way toward explaining their small market share?
Surely these systems would consume less power than an equivalent number of 2P or 1P servers?