AMD plays it cool with low-volt Istanbuls
No new server chips until 2010
AMD is today expected to complete its rollout of the six-core Istanbul series of Opteron processors with the final chip in the family: a low-voltage part for two-socket servers.
AMD has been dabbling with voltages and clock speeds to make Opterons of various performance levels and thermal envelopes since the 64-bit Opterons shook up the server racket in 2003.
The company was way ahead of the curve when it came to delivering low-voltage, low-frequency parts that had the same feature set as the standard x64 chips it sold, but for complicated reasons, server makers and customers were not quite ready to sacrifice raw performance for the kinds of performance per watt that the so-called Opteron Energy Efficient (EE) chips could deliver.
And so, AMD eventually kicked the EE parts over to an embedded business and didn't sell them through its normal channel to server OEMs. The low-powered parts were backburnered. But as the quad-core Shanghai Opterons lineup was expanded this past April on the sixth birthday of the Opteron chips, AMD recommitted itself to the EE parts as it saw its Highly Efficient (HE) parts start selling better.
That said, AMD is not crazy and is only trying to sell EE parts in high-end, four-socket and larger machines, but just in two-socket boxes.
The quad-core Shanghai chips had two EE parts, and both were rated at 40 watts like the new Istanbul Opteron 2419 EE part that comes out today.
Standard Opterons are rated at 75 watts, HE parts are rated at 75 watts, and Special Edition (SE) parts, which offer the peak performance in any Opteron generation, are rated at 105 watts.
Here's how AMD actually mixes and matches the voltages and clock speeds to get those numbers. The six-core Opteron 2419 EE runs at 1.8GHz and runs at 1.125 volts. That low-clock speed is what helps the part stay in that 40-watt thermal envelope. The two Shanghai parts, the Opteron 2373 EE and 2377 EE, ran at 2.1GHz and 2.3GHz, respectively, but because there was less stuff crammed into the chip, it could still meet the 40 watt envelope running at the same 1.125 volts.
The Istanbul HE parts run at 1.15 volts and top out at 2.1GHz, yielding that 55 watts, while the standard parts raise the voltage to 1.3 volts and crank the clocks as high as 2.6GHz. Because heat increases on a log scale as volts and clock speed rises linearly, the relatively small jump in volts (13 per cent) and clocks (23.8 per cent) yields a much larger increase in heat dissipation (36.4 per cent).
The SE parts are really ridiculous in this regard, with clock speeds only rising by 7.7 per cent to 2.8GHz and voltage rising by 1.9 per cent, but heat dissipation increases to 105 watts, a jump of 40 per cent over the standard parts.
The Opteron 2419 EE chip is available today, and costs $989 each when customers buy them in 1,000-unit trays. On integer workloads (SPECint_rate2006), the six-core Opteron 2419 EE running at 1.8GHz is rated at 158, which is 32.8 per cent higher than the 119 rating that the four-core Opteron 2377 running at 2.3GHz. However, the Opteron 2377 EE only cost $698. You have to pay 41.7 per cent more dough to get that extra 32.8 per cent oomph.
These numbers won't get the hyperscale data center customers that AMD is targeting with the EE chips very excited. In fact, they will probably be annoyed that the cost for a unit of performance has gone up.
That said, big data center operators and the server OEMs that are building custom boxes that attempt to squeeze all the power out of a box that they can will no doubt listen to the following numbers with some interest.
On the SPECpower_ssj2008 energy efficiency benchmark, a two-socket server using the Istanbul EE parts is delivering 31 per cent better performance per watt than a standard quad-core Shanghai chip running at 75 watts. The EE part is priced the same as the Shanghai chip running at 2.9GHz (the Opteron 2389), and it would have been interesting to see the SPECpower_ssj2008 comparison with a machine using this chip, which would answer the question how much more efficient can the machine be for the same money?
Even if the Istanbul EE chip only offers 20 per cent better bang for the kilowatt at the same cost, that is something hyperscale data center customers are looking for.
The other part of AMD's sales pitch for the Istanbul EE parts is that you can use less energy per server and get rid of some heat islands in the data center, or cram more physical boxes on the floor in a given power budget. This latter bit cuts against the whole grain of virtualization, of course, but so do big Web 2.0 operators, who don't tend to virtualize their workloads.
Using the 1.8GHz Instanbul HE chip, you can cram 42 1U, two-socket servers into a standard rack with a 7.5 kilowatt power budget, given the 178 watt platform power load on the SPECpower_ssj2008 test.
A fully populated rack of two-socket boxes using the 2.1 GHz six-core Opteron 2425 HE chips will consume 218 watts, or 9.2 kilowatts per rack of 1U machines, and moving to the top-end standard 2.6GHz Opteron 2425 HE chip boosts the power consumption of a rack of servers running the SPECpower_ssj2008 test to 12.9 kilowatts.
The difference per rack (at 10 cents per kilowatt hour plus scaling factors for data center cooling and power distribution inefficiencies) in terms of energy costs is this: a rack of the Istanbul EE chips costs $249 to run for a year, compared to $305 for a rack of servers using the Istanbul HEs and $430 for a rack of machines using standard Istanbuls.
The other way of using those numbers is this way: you can get 42 servers using the EE chips into a rack at that 7.5 kilowatt power budget, but only 34 servers with the HE parts and only 24 using the standard parts.
AMD is also bragging about the fact that the Istanbuls use less expensive and cooler DDR2 main memory instead of the faster and hotter DDR3 main memory used in Intel's Nehalem EP Xeon 5500 processors.
While AMD conceded that the faster DDR3 memory, which will be used in AMD's twelve-core Magny-Cours and six-core Lisbon Opteron chips early next year, does run faster and give some performance advantages, when you look at memory costs relative to the performance, DDR2 wins "hands down," according to AMD servers and workstations senior product manager Gina Longoria. "We have the right timing for our DDR3 crossover for next year," she says.
In memory-heavy configurations, such as for servers running the memcache-d web caching program, the difference in memory price on roughly equivalent configured Opteron 2400 and Xeon 5500 servers can be as much as $1,000, by Longoria's math, and that can work out to be $1.2m difference across a 25-rack server setup.
AMD's comparison put 64GB of 667 MHz DDR2 memory in each Istanbul box and 72 GB of 1GHz DDR3 memory into the Xeon 5500 boxes. All were 4GB modules. It costs $130 per DIMM for the DDR2 memory and $180 per DIMM for the DDR3 memory. AMD did not do the math on how much more heat the DDR3 memory would generate. ®