AMD’s Hammer Less Than Four Weeks Away
For almost two years, AMD’s eighth-generation, 64-bit “Hammer” CPU architecture has garnered endless media exposure, while the actual processors have slowly made their way from paper announcements to real-world products. Now the wait can be counted in days, not quarters: While the desktop and notebook-oriented Athlon 64 chip has been scheduled for September 2003, the server and workstation-optimized Opteron will debut less than a month from now, on April 22.
While CPU Planet has already offered an introduction to the Hammer family, it’s time for an update of what’s known about the Opteron, the systems that’ll use it, and the software that’ll be available for it.
Opteron Changes Everything
The Opteron is radical — it represents a totally new CPU architecture, system platform, and even microcode support all rolled into one. The Opteron is conservative — AMD’s x86-64 architecture runs current 32-bit applications, and quickly, while giving forward-thinking buyers a transition path to 64-bit computing rather than the start-from-scratch approach of Intel’s 64-bit Itanium family.
The 0.13-micron-process silicon-on-insulator Opteron — a 90-nanometer-process successor will arrive in 2004 — features two more pipeline stages than AMD’s Athlon XP; instructions-per-clock-cycle boosters such as enhanced branch-prediction algorithms and larger translation look-aside buffers; support for the SSE2 streaming multimedia instructions that debuted in Intel’s Pentium 4; and up to 1MB of Level 2 cache, all in a new, plus-sized processor die or 940-pin ceramic package. (The Athlon 64 will use a different, 754-pin socket.)
Both the Opteron and Athlon 64 boast 64-bit data and address paths and break through current 32-bit CPUs’ 4GB memory addressing cap with 40-bit physical (up to 1 terabyte) and 48-bit virtual (up to 256 terabytes) memory addressing space. The Opteron also supports three HyperTransport links, providing up to 19.2GB/sec of bandwidth, versus the Athlon 64’s single HyperTransport link for 6.4GB/sec of data transfer.
In terms of architectural changes, the most noticeable is the Opteron’s integrated memory controller — a 128-bit, dual-channel design supporting DDR266 and DDR333 SDRAM. Both the Opteron’s memory controller and the Athlon 64’s — a single 72-bit channel — take that job away from its traditional place in the system chipset’s external Northbridge, greatly reducing the latency of read/write requests. This essentially controls the system at, or yields a front-side bus speed matching, the clock speed of the CPU.
New Model Numbering
PC processors have traditionally been identified by their clock speeds. AMD’s Athlon XP took a step away from literal megahertz to model numbers based on virtual speeds — the company says the Athlon XP 3000+ offers performance equivalent to an original Athlon design running at 3GHz, while most users compare it to Intel’s 3.06GHz Pentium 4.
The Opteron model numbering system moves away from clock speed altogether, rolling the CPU’s scalability and performance into one easy-to-digest, three-digit number. The first number represents the number of processors supported for simultaneous or multiprocessing installations; the basic Opteron lines are the 100 Series (single processor), 200 Series (dual processors), and 800 Series (up to eight-way multiprocessing).
The next two digits — for the first Opterons, numbers in the 40s — indicate relative performance within (not necessarily across) the series. For example, an Opteron 240 would be the base model for dual-processor server and workstation platforms, while an Opteron 242 would be a faster alternative for the same systems. (Actual model numbers haven’t been officially announced, but we think the above is a good guess: Before a hasty, likely-urged-by-AMD switch back to just “100 Series, 200 Series, 800 Series” yesterday, server vendor PSSC Labs’ site offered pre-order shoppers a menu of Opteron models 140, 142, 240, 242, 840, and 842 — Ed.)
This new numbering scheme is definitely a gutsy move, as the obvious questions in buyers’ minds will be how much faster the higher-numbered Opterons will be, and at exactly which clock speeds they will run. It seems obvious that Intel will continue to hold the clock-speed advantage (its Xeon currently peaks at 3.06GHz), while AMD might well win the performance race — it’s shown estimated 32-bit benchmark results that show a 2.0GHz Opteron comfortably ahead of the 2.8GHz Xeon.
AMD has already successfully shifted gears with the Athlon XP model numbers, and the server business is far more interested in performance, scalability, and architecture than outright core speed. If in doubt, head on down to your local IT department and ask what’s under the server hood; you might be surprised to find many systems with only 250MHz to 500MHz per processor.
By the same token, the new lingo might not be as easy a sale with the workstation crowd, where power and panache go hand in hand and owning the highest-clocked CPU can be a status symbol. Either way, AMD will need to spell out the Opterons’ clock speeds somewhere; the best predictions we have range from 1.4GHz to the 2.0GHz or higher goal that’s been mentioned in company speeches and briefings.
Chipsets and Motherboards
AMD hasn’t actively produced a system chipset since the Athlon MP/MPX platforms, so its return to the chipset business with the 8000 series silicon for Opteron may raise a few eyebrows. On the other hand, this strategy allows new CPU architectures to have a viable, name-brand chipset immediately upon launch, which is a required element with server/workstation products seeking corporate approval. And as with past AMD releases, third parties like VIA, SiS, and Nvidia will pick up the slack with complementary offerings later — although the chipsets that have been most visible in trade-show previews, like Nvidia’s Crush K8 series, are aimed at the Athlon 64, so the 8000 series should dominate the initial Opteron market.
The integrated memory controller, multiprocessing support, and other features of the Opteron change the chipset parameters substantially, basically eliminating the need for a conventional, external Northbridge component. The closest thing to an old-fashioned Northbridge in the 8000 family is the AMD-8151 HyperTransport AGP 3.0 Graphics Tunnel, which provides AGP 4X/8X compatibility while using a 6.4GB/sec HyperTransport link to the host controller and 1.6GB/sec link to a downstream device.
The AMD-8111 HyperTransport I/O Hub and AMD-8131 HyperTransport I/O Bus Tunnel are more traditional Southbridge components, with the 8111 providing standard I/O and storage functions — including six USB 2.0 ports, AC97 six-channel audio, and dual ATA/133 channels — while the 8131 integrates dual high-speed PCI-X bus bridges with the HyperTransport tunnel.
The design of the Opteron allows greater commonality between the chipsets, whether running in two- or eight-way configurations. That translates into multiple uses of the same AMD 8000 series design, and also the option for third-party vendors to mix and match or use only those pieces of the chipset they need.
Companies that have shown Opteron motherboards — mostly two-way multiprocessing platforms — include FIC, Gigabyte, Rioworks, MSI, and Tyan. The latter’s Thunder K8 (S2880) is a good example, pairing two Opteron sockets with six DIMM sockets for up to 12GB of registered DDR333 memory; Serial ATA RAID and Ultra320 SCSI controllers; one PCI and four PCI-X slots; and AMD’s 8111 I/O and 8131 PCI-X components — since it’s a server platform, Tyan opted for a relatively low-performance ATI integrated graphics controller instead of AMD’s AGP 3.0 8151 component.
System Vendors and Software
AMD will not be releasing an official Opteron vendor list until at least April 22, but the company has made no bones about going after the largest server manufacturers: CEO Hector Ruiz said in a Reuters interview that AMD was “working with all of the top five [computer makers],” at least as far as the preliminary evaluation period that precludes any PC vendor’s decision to use a new processor. So far, that means nonstop gossip about Dell, and almost as much talk about — and perhaps a better chance of actual adoption by — IBM.
While we can look forward to at least one big name being dropped at the launch party, a growing number of smaller vendors are already trumpeting their Opteron support. Many have licensed designs from Newisys, an Austin, Texas, startup headed by former IBM and Dell execs that’s offering a complete dual-processor 1U rackmount server, configurable with 512MB to 16GB of DDR333, paired Ultra320 SCSI or IDE hard drives, dual embedded Gigabit Ethernet adapters, and two PCI-X expansion slots.
Angstrom Microsystems is promoting Titan64 dual-CPU servers, while RackSaver, whose Linux clusters have displaced Sun servers at companies like Pixar Animation, has promised to ship Opteron-based rack servers. And Appro International has been demonstrating what it calls HyperBlade server clusters that combine up to 80 dual-Opteron computers, at prices ranging from $160,000 to $320,000.
On the software side, the lack of 64-bit operating systems has been an oft-cited reason for the delay in bringing AMD’s Hammer processors to market. In the case of the desktop Athlon 64, this is a major concern, as going live without a compatible 64-bit version of Windows XP would curtail some of the advantages of the Athlon 64 architecture. The Opteron, however, is a bit different, in that it handles the server and workstation areas and is specifically positioned as a 32- to 64-bit transition engine.
So while Microsoft prepares to launch Windows Server 2003 a few days after the Opteron launch, it seems probable that it’ll run in 32-bit mode on the Opteron, with 64-bit support to follow in a later upgrade. SuSE, by contrast, has promised a full 64-bit version of Linux as soon as the Opteron ships, with Red Hat committed to follow a bit later. The 64-bit applications lined up so far aren’t very numerous, but range from the IBM DB2 and open-source MySQL databases to Valve’s CounterStrike online game server. Still, the real emphasis seems to have shifted from 64-bit out of the box toward a more flexible upgrade path from current 32-bit applications to 64-bit software when needed.
No one denies that AMD faces an uphill fight against Intel, but there has never been so much buzz surrounding an AMD server processor, nor such promise for truly competitive performance and scalability — both for entry-level and midrange servers, and for four- and eight-way systems in the $10,000-and-up segment that today’s Athlon MP has had to cede to Intel’s Xeon MP and Itanium 2.
AMD’s “32-bit today, 64-bit tomorrow” strategy should pay off with budget-conscious IT managers, and with one or two big-name server vendors on board and a fairly quick arrival of 64-bit Windows Server, the underdog’s gamble could pay off. Stay tuned for April 22.