AMD cranks up Opteron clocks
Before Bulldozer, a few more megahertz
Advanced Micro Devices is firing off the first salvo in the ongoing x64 war between Intel and itself, revving up its high-end Opteron 6100 processors to try to get a bit more market share.
The crank on the eight-core and twelve-core "Magny-Cours" processors, which debuted last March, is not huge, so don't get too excited. AMD and its former foundry and now third party wafer baker, GlobalFoundries, have pushed the 45 nanometer processes used to make the Opteron 6100s about as far as they can while staying within the 65, 80, and 105 watt thermal bands that server makers engineer their machines to.
As it is, server makers were being asked to do a lot by AMD when the chip maker double-stuffed the G34 sockets with two six-core chips with the Opteron 6100s. (The Magny-Cours is not a single piece of silicon, but two tweaked "Istanbul" chips sitting side-by-side in the same package.) Not only was there a new and fatter socket to contend with in the belly of the Great Recession in late 2008 and early 2009, when AMD was laying out its Opteron 6100 plans.
The new chips burned more juice and threw off more heat than their Istanbul predecessors, which came in 55 watt, low-voltage versions (the Highly Efficient or HE parts), 75 watt standard parts, and 95 watt revved up Special Edition, or SE parts. Coping with the extra 20 or 40 watts is not trivial in a 1U or 2U server.
The Magny-Cours launch also required server makers to re-engineer their server platforms and adopt AMD's first server chipset, since Nvidia and Broadcom stopped making new Opteron chipsets with this generation. When you add it all up, it is no wonder that Hewlett-Packard and Dell were the only big players that really backed the Opteron 6100s, that IBM gave it tepid support, and Oracle ditched Opterons entirely from their product line. (I know Acer made some noise in Asia and Europe, but hasn't even entered the U.S. market with its Opteron-based systems yet.)
It was a bad time to have server makers and buyers alike cope with a new architecture, and it is no surprise that Intel did well because it did its major transition a year earlier, in March 2009, with the four-core "Nehalem-EP" Xeon 5500s, at the bottom of the recession. Server makers had adopted the new LGA 1366 sockets and the QuickPath Interconnect back in 2008 when they were still feeling rich, and when Intel launched the six-core "Nehalem-EP" Xeon 5600s in March 2010, these new chips had the same thermals and plugged into the same sockets as the Xeon 5500s. It only required them to recertify their configurations, which is the right amount of work for a lazy server maker.
Here in 2011, Intel has a socket shift it is making server makers do with its "Sandy Bridge" Xeons, and they are probably grousing about it a bit, but with the server racket and the major economies of the world on the mend, this transition is not as big of a deal as the one they were facing with the Opteron 6100s and 4100s last year. That said, AMD will have an easier time because the future "Valencia" and "Interlagos" Opterons, based on the "Bulldozer" core pair and due later this year, will snap into any existing C32 or G34 socket, respectively. (See here for details on the Bulldozer chips and there for AMD's Opteron roadmaps through 2012.)
In the meantime, AMD needs to hit Intel with something, and the only weapon it has is to sift through the Opteron 6100 bins and find machines that can run at higher clock speeds reliably.
After sifting through those bins, AMD came up with five new Opteron 6100 processors, which launch today. In the table below, they are shown in red italics:
The AMD Opteron 6100 server chip lineup
The Opteron 6100s are used in two-socket and four-socket servers, and given that each socket has two whole processors in it, these are really quasi four-socket and eight-socket machines as far as systems software is concerned. (AMD and its partners could do eight-socket machines, comprised of four two-socket boards or two four-socket boards, from the early days of the Opterons, but there were not enough HyperTransport links in the chip sockets to make all sockets able to link to all other sockets, so there was a lot more hopping around in the system. This had a predictable and bad effect on NUMA scaling for these early Opterons. But with the latest Opterons, their HT3 links, and more of them, the NUMA scaling is vastly improved.)
In the table above, each chip's clock speed and core count is shown. All of the Opteron 6100s have 512 KB of L2 cache per core and 6 MB of L3 cache per core, for a total of 12 MB for a Magny-Cours package. The Average CPU Power (ACP) thermal rating is not equivalent to Intel's Thermal Design Point (TDP) ratings. Pricing show is the per-unit price if you buy a 1,000-unit tray of chips from AMD.
8.7 per cent more oomph
The Opteron 6180 SE at the top of the table has about 8.7 per cent more oomph, based on raw clock speed, and at $1,514, the price is 8.5 per cent higher than the original Opteron 6176 SE part from last year. Sun Microsystems was the most enthusiastic supporter of the SE parts, believing that the extra oomph gave it an edge in the HPC space. But Oracle, after acquiring Sun last year, quietly killed off its Opteron products, so it is unclear who uses the SE parts. There's probably a financial services company monkeying around with your retirement funds on a few racks of these sitting near an exchange.
The 2.3 GHz twelve-core Opteron 6176 is the new top-end standard part, and should deliver a mere 4.5 per cent performance improvement over the 2.2 GHz Opteron 6174; it costs 8.6 per cent more money than the Opteron 6174, however, making it not a particularly good dead. But that is not the important comparison that AMD customers will be making. The important thing now is that the Opteron 6176 is a standard part burning at 80 watts instead of 105 watts and costs 9 per cent less than last year's Opteron 6176 SE part, which had the same performance.
The 2.6 GHz Opteron 6140 is now the fastest eight-core Magny-Cours processor, and it comes in an 80 watt thermal envelope and costs $989. This chip delivers 8.3 per cent more raw clock oomph than the 2.4 GHz Opteron 6136, but it costs 32.9 per cent more (and the same as the twelve-core Opteron 6172 running at 2.1 GHz). This is not a particularly good deal, but HPC customers using Cray's XT6 and XE6 supercomputers as well as those building or buying dense clusters to run HPC-style codes that need the fasted clock cycles they can get their hands on will look at these chips.
Then again, if they need the highest clock speeds and don't care as much about threads, virtual or physical, Intel has faster six-core Xeon 5600s that push up to 3.33 GHz and even four-core variants that can be pushed even as high as 3.46 GHz. These chips also cost $1,663 a pop when you buy in 1,000-unit trays.
There are two new low-voltage Opteron HE parts that are coming out today. The Opteron 6166 HE runs at 1.8 GHz, has twelve cores, and costs $873. That is 5.9 per cent more clocks than the Opteron 6164 HE for 17.3 per cent more dough. (Again, not a great deal.) The Opteron 6132 HE runs at 2.2 GHz in its 65 watt ACP envelope and costs $591. That's 10 per cent more performance for 13 per cent more money over the 2 GHz Opteron 6128 HE.
These chips are pretty much all that is coming out from AMD on the Opteron front this year until the Bulldozer-based chips - presumably to be called the Opteron 4200 (Valencia) and 6200 (Interlagos) for the C32 and G34 sockets, respectively - come out.
"We have no plans to do anything with the Opteron 4100 because we have the Bulldozer coming so close at hand," John Fruehe, director of product marketing for server/workstation products at AMD, tells El Reg.
The Opteron 4100s, which launched in June 2010, come with a single four-core or six-core chips in a package. The Opteron 4100s have slightly higher clock speeds and better thermals with ACP ratings of 75 watts for standard parts and 50 watts for HE parts. There are even 32 watt parts, the Extremely Efficient, or EE, variants. These C32-socket chips are used in servers with one or two sockets, and they are very inexpensive compared to the Opteron 6100s. (Which stands to reason with the core counts being half.) That said, AMD could probably sort through the bins and find 3 GHz or even 3.2 GHz Opteron 4100s if it wanted to.
Fruehe says that Interlagos will go into production in the second quarter using GlobalFoundries' 32 nanometer processes, followed in short order by the Valencia chips. (Interlagos will pack 12 or 16 cores, while Valencia will have six or eight cores.) Both chips will launch at the same time in the third quarter and be adopted shortly thereafter by HP, Dell, Acer, Cray, and IBM, but Interlagos will be "first through the pipe" to put more pressure on Intel and try to get some year-end server market share. ®