Following our look at AMD’s upcoming platform, processor and GPGPU plans, this final installment of a four-part series which digests AMD’s 2009 Financial Analyst Day takes a look the company’s efforts with enterprise hardware.
The story of AMD’s server segment throughout the last 18 months is one that mirrors the desktop: variable—if not declining—market share owed largely to financial struggles which have hamstrung the development of architectural and process advancements. Looking to the year ahead, however, we can see the trappings of a watershed moment. With ambitious plans for two new platforms, two new sockets, a migration to DDR3, a 12 core CPU, and early samples of a next-gen architecture, let’s see if 2010 marks the beginning of the turnaround AMD’s server segment needs.
Overall strategy
In a nutshell, AMD is targeting the volume server market that describes fully featured dual socket systems, and the so-called “volume 4P” market that consists of quad processor servers which run between $6,000 and $10,000 USD.
This is easily the best possible strategy for the company as volume server sales are expected to grow in 2010 by 4.3%, while the remaining segments will be flat, or even down by more than 12% at the top end of the scale. Additionally, by making the volume market even less expensive than the $25,000 ceiling billed by IDC, AMD will undoubtedly capture additional buyers who were reticent to upgrade in 2009.
From a technological perspective, it’s clear that AMD is looking to pursue virtualization. By dramatically increasing the number of physical cores from six to 12, or even 16 by the end of 2010, the parallel computing capacity of 4P AMD systems could, for example, push VMWare VMark scores higher than 54@39. Meanwhile, competing 4P Nehalem EX systems should top out at 44@31. In other words, the 16 threads offered by the Nehalem EX’s eight Hyper-Threaded cores just can’t cut it against the 12 real threads offered by AMD’s upcoming chips.
Efficiency will also be the name of AMD’s game in the year ahead. Though Average CPU Power (ACP) and Thermal Design Power (TDP) are not directly comparable, AMD’s 130W (ACP) dodeca-core chips should offer consistently lower power consumption than the 140W (TDP) octo-core Nehalem EX. This will translate directly to superior performance per watt per socket, and that’s a strong selling point for server vendors that are boarding the green bandwagon in increasing numbers.
Sockets and platforms
As AMD ramps up its core counts and transitions to DDR3 this year, new sockets with additional or reconfigured pins will be required to handle the increased I/O and the HyperTransport links to orchestrate it. AMD is using this transition as an opportunity to stratify its workstation and server markets into two separate platforms known as Maranello and San Marino, respectively.
San Marino
Taking up the ground floor, the San Marino platform is a 1-2 processor solution for workstations, web, cloud and infrastructure work. San Marino features the new Socket C32, which reconfigures the old Socket F’s 1207 pins into a format that’s compatible with Lisbon’s new on-die, dual channel DDR3 controller.
AMD’s Director of Business Development for Server/Workstation Products John Fruehe says that San Marino-based machines are designed with efficiency in mind.
“The platform choices around ‘San Marino’ are expected to help OEMs optimize their C32 systems for low power consumption and low cost. We believe these AMD Opteron 4000 Series processors will have the potential to help users achieve new levels of price/performance and performance/watt,” he says.
“When you walk through a data center and see rack after rack of servers, it’s clear that reducing the cost and power footprint of the ‘workhorse’ servers can have a huge impact on the bottom line.”
On the processor front, Lisbon will be the go-to chip for Socket C32 throughout 2010. It’s easiest to consider these chips the DDR3 version of AMD’s existing Shanghai or Istanbul products with a little optimization sprinkled on top. As the slides indicate, Lisbon will be sold under the AMD 4000 series, but it will be succeeded by the 32nm Valencia in 2011 (more on that later).
Maranello Platform
Targeting HPC, virtualization, and databasing, the Maranello platform comes in at the top of AMD’s heap. Featuring a whopping 1974 pins, Maranello’s Socket G34 not only has the pins to support quad-channel DDR3, but also to rock the veritable spider’s web of HyperTransport links required to keep all the Magny-Cours cores talking to one another.
On the processor front, Magny-Cours is being primed for a 1Q10 launch. As an MCM processor, the twelve core chip features a pair of hexa core Istanbul dies married at the middle by HyperTransport. We’ll touch more upon this chip in the next section.
Why stratify?
As has been indicated, processors designed for C32 and G34 are incompatible with one another. This is not a maneuver made out of greed or malice, rather it’s a decision made out of necessity and cost consciousness. Put simply, the additional cores and memory channels offered by Magny-Cours require more voltage and I/O pins in the socket–767 more, to be precise–than the simpler Lisbon. While AMD could easily make Lisbon socket-compatible with G34, that would be a cost passed onto the consumer for zero performance benefit.
Magny-Cours also requires several additional HyperTransport lanes for the purpose of cache and logic coherency. Either every motherboard would ship with appropriate support and its corresponding costs, or vendors would be forced to maintain a complicated compatibility matrix of the Lisbon-oriented boards that don’t offer those lanes: talk about a support nightmare.
In short, stratifying 1-2P from 2-4p (with a selection of 2P G34 mainboards in the middle) is ideally suited to offering the right solution to customers at the right price.
Processors
As previously mentioned, AMD is arming itself in 2010 with the 4-6 core Lisbon, and the 8-12 core Magny-Cours. In this section we’ll not only take a closer look at their designs, but we’ll also take a peek at Valencia and Interlagos, two Opterons based on AMD’s next-gen Bulldozer architecture, which will begin sampling to OEMs in 2010.
Lisbon
Family: Stars (Phenom II)
Cores: 4 to 6
Process: 45nm
Socket: C32
Memory: Dual channel DDR3
Platform: San Marino
Role(s): Workstations, web serving, cloud computing and infrastructure
Launch date: 1Q10
If you’re looking to build a render box or do a little web serving with the green team in 2010, Lisbon is your go-to chip. Featuring dual channel DDR3, 4-6 cores, low TDPs and slotting into the most inexpensive SMP platform AMD offers, Lisbon is a rather modern and relatively inexpensive solution.
From an architectural perspective, Lisbon chips are exceptionally similar to Socket AM3 Phenom II processors.
Magny-Cours
Family: Stars (Phenom II)
Cores: 8-12
Process: 45nm
Socket: G34
Memory: Quad channel DDR3
Platform: Maranello
Role(s): HPC, clusters, virtualization
Launch date: 1Q10
The Magny-Cours is an exceptionally neat processor that combines two hexa core Istanbul Opterons with a HyperTransport link. Pull off the die cover, shown right, and you’d find two independent CPU dies–this architecture is known as an MCM, or multi-chip module.
While all signs point to a launch frequency of 1700-2100MHz, we know concretely that it will ship with 128K L1 per core, 512k L2 per core and 10MB shared L3, along with power management functions that can take the chip as low as 800MHz.
As we also suggested, Magny-Cours is a monster when it comes to parallel processing. The following image shows an 1700MHz Magny plowing through 32 million digits of Pi in less than 6.5 seconds–a 3600MHz Core i7 takes 24 seconds.
Valencia
Family: Bulldozer
Cores: 6 or 8
Process: 32nm
Socket: C32
Memory: (Dual channel DDR3?)
Platform: San Marino
Role: Workstations, web serving, cloud computing and infrastructure
Launch date: 4Q10-1Q11
Valencia is one of two processors based on AMD’s upcoming Bulldozer architecture that the firm will begin sampling in 2010, and launch in 2011. The subtleties of Bulldozer are legion, but major benefits include even better parallel performance, lower temperatures, lower power consumption, clock for clock performance enhancements, and improved floating point performance.
Valencia and its big brother, Interlagos, are particularly innovative chips explicitly designed for the kinds of demands servers make. And, as an added perk, AMD has already suggested that Valencia CPUs are drop-in upgrades for today’s Lisbon chips.
Interlagos
Family: Bulldozer
Cores: 12 or 16
Process: 32nm
Socket: G34
Memory: (Quad channel DDR3?)
Platform: Maranello
Role: HPC, clusters, virtualization
Launch date: 4Q10-1Q11
As implied, Interlagos adds additional Bulldozer modules to bump the core count to a maximum of 16. If you thought Magny-Cours had outstanding parallel performance, wait until Interlagos numbers start filtering into the public. VM nerds are going to bust a nut.
Final thoughts
Looking throughout 2010, it is clear that AMD is working very hard to regain the market share it has lost since the advent of Core 2-inspired Xeon designs. Through lower prices, lower wattage and superior parallel performance, AMD is attacking on the three fronts that are increasingly relevant to an enterprise market that doesn’t want to sacrifice performance to address its environmental and economic concerns.
More importantly, AMD appears to be on track with its schedule. This is incredibly important for a company which has slipped its deadlines on several key technologies, particularly 32nm and Fusion. Only through demonstrating to analysts, investors and bean counters that the company can deliver on time and in volume will AMD cultivate the image it needs to precipitate a recovery.
As we also come to the end of our series, we must step back and look at AMD’s many parts as a single machine. From this perspective, AMD has rarely been in better form. With six months of leadership in DirectX 11, a timely next-gen architecture in the wings, new chipsets for 2010, APUs prepared to clobber Intel’s offerings, best-in-class mobile GPUs and ambitious plans for the enterprise, AMD hasn’t looked this good since it sucker-punched Intel with the launch of the Athlon 64.
Given AMD’s recent history, the biggest hurdle for the firm will be its ability to deliver as intended, but we are very optimistic about those chances. All the signs are pointing in the right direction: AMD is back.
Articles RSS