It's not enough to make a power-efficient server or processor anymore; no, you also have to package it as being for "the Cloud." In this respect, AMD's announcement today is similar to the recent Tilera and SeaMicro news; the chipmaker is talking cloud, cloud cloud. By now, I've seen essentially the same pitch deck at least half a dozen times—only the company- and technology-specific slides are new in each case.
In the case of today's Opteron 4000 series ("Lisbon" platform) and Firestream coprocessor launch, AMD has a particularly good slide that summarizes most of the other slides in most of the other "here's our new cloud datacenter product" decks that I've seen in the past few weeks.
As a 4- and 6-core server part for one- and two-socket systems, Lisbon is definitely more traditional than the more exotic cloud-specific products that have been mooted recently. The latter have been aimed at Intel's Xeon as the epitome of the traditional, performance-oriented, chronically underutilized server processor, while AMD hopes to slot Lisbon in right beneath Xeon in the price/performance ladder.
What AMD offers with Lisbon is a lower performance, lower power, and lower cost alternative for x86 customers who still want the fairly wide range of dynamic performance/watt scaling that the emerging class of radically different "physicalized" designs can offer.
The one- and two-socket-capable 4000 series chips are priced to be competitive with Intel's single-socket-only Xeons, while AMD's two- and four-socket-capable Opteron 6000 series is priced to compete with Intel's two-socket Xeons. This pricing strategy will probably benefit AMD in the cost-sensitive datacenter market.
The 4000 series is AMD's first server processor family to target the 32W, 50W, and 75W per node power bands, and AMD claims that the 32W 4162EE and 4164EE have the lowest power per core of any server processor on the market as of launch. This low per-core power puts the 1.8GHz 4164EE at the top of the 4000 series price range, at $698. At the bottom of the series price range is the 75W, 2.2GHz, four-core 4122, for $99. Clearly, AMD is not pricing these parts on anything like raw performance, but on efficiency.
The new Opterons are based on pretty much the same microarchitecture that AMD has been using for some time now, so there's nothing new there. AMD has focused most of its efforts for Lisbon on getting the platform's overall power down and its scalability up.
For instance, Lisbon's BIOS offers users very fine-grained control over which parts of the computer are active. You can turn off the PCIe slots for a savings of 2.4W, or the USB ports for a savings of 1W, or a number of other components for a total of 62W per motherboard in potential savings. Across a large datacenter, saving a watt for each node adds up quickly, so this kind of feature makes plenty of sense for AMD's target market.
Lisbon also brings the Direct Connect Architecture 2.0 that debuted with Magny-Cours down into the lower end of the server space, which includes AMD-V (plus virtualized I/O), HyperTransport 3.0, and support for DDR3 1333.Firestream looks to the datacenter
The other big news from AMD today is the launch of the ATI Firestream 9350 and 9370 products, which are ATI's answer to NVIDIA's Tesla. The new Firestream parts are essentially the "Cypress" GPU from the Radeon 5870, but repackaged (and marked up) with more memory as an add-in board for high-performance computing.
The new parts have full hardware support for OpenCL 1.0 and DirectX 11, and they come with passive heatsinks for higher reliability in server environments. Both Firestream products will launch in the third quarter of this year.Real integration will wait for Fusion
One of the things that I had been expecting to see by now was some sort of deep, platform-level CPU + GPU integration—basically a product with a GPU in a coprocessor socket, instead of on a PCIe daughterboard. This never really happened, probably because the combination of PCIe and GDDR5 is good enough, and now it's clear that it won't happen. AMD has focused its CPU + GPU efforts entirely on wedding the two devices in silicon on the same die, and is skipping the deeper, non-commodity platform integration step entirely.
I asked AMD about this issue, and they confirmed that it's reasonable to expect a Fusion server part for high performance computing; they also reiterated that, until such a part arrives, they're focusing on getting the software side of the integration picture (OpenCL and DirectCompute) in shape.
Source: ars technica