AMD Opteron 6100 Series review
For the first time in years, AMD has managed to trump Intel on the fundamental specification of a new processor. Launched in March, the “Magny-Cours” Opteron 6100 Series has eight or 12 cores, beating Intel’s Xeon 5600 series’ maximum of six cores. This high core count even stomps on the latest eight-core Xeon 6500 and 7500 “Nehalem-EX” chips that were announced shortly afterwards. We also have the first Opteron 6100 production rack server in the lab, courtesy of Boston Ltd. The GP 1120-T is equipped with a pair of 6174 modules, and we put it through a range of benchtests to see how it stacks up against the Xeon 6500, 7500 and 5600.
Compared with the Xeon 5600, the 6100 processor choices have been greatly simplified as they all have the same 12MB of L3 cache, the same HyperTransport speed and support for 1,333MHz DDR3 memory. All you need to decide is whether you want eight or 12 cores, whether you want one, two or four of them and the clock speed.
The Xeon 5600 offers 12 processor models, as opposed to AMD’s ten. However, along with three different QPI speeds, the 5600 offers a wide range of clock speeds, from 1.86GHz up to 3.33GHz, four or six cores, five different power ratings, varying HT and Turbo Boost support and different support for 800MHz, 1,066MHz and 1,333MHz memory speeds.
Along with a simplified product line, AMD’s product positioning puts the 6100 series directly into competition with the Xeon 5600 market and also takes a swipe at the 7500s. Intel has applied Itanium-like pricing to Xeon 6500 and 7500, so only those businesses with deep pockets will be able to afford them. The 6100 supports 1P, 2P and 4P systems and AMD’s pricing structure removes the cost barrier that the tiered 1000, 2000 and 8000 Opteron line-up imposed. Previously a 4P system required an expensive 8000 series processor to work, and a 2P required at least a 2000 series. The 6100 series removes that artificial barrier – they’ll work in 1P, 2P and 4P machines, for a clear upgrade path as long as you think about your requirements ahead of time and buy a suitable chassis.
AMD’s prices for its 12-core processors starts at $744 (around £490) for the 1.9GHz 6164 HE, going up to $1,386 (around £913) for the top-of-the-range 2.3GHz 6176 SE. This means that many businesses that traditionally go for 2P servers can now consider stepping up to 4P systems and some cost-reducing server consolidation. This leaves the Opteron 8000 hanging out to dry, as it supports 2P, 4P and 8P systems, and interest in the bigger boxes is showing a marked decline. AMD said it would continue to produce these processors but admitted that from a performance perspective there wasn’t much sense in buying them.
On the design front, the 6100 stays with the same 45nm fab that’s used for the Xeon 6500 and 7500. Only the Xeon 5600 has dropped down to a 32nm manufacturing process. Cramming in 12 cores has resulted in a larger rectangular processor, requiring AMD to abandon Socket F. The new Socket G34 means that moving to Opteron 6100 will require a complete upgrade, as they’re not backward-compatible.
The 12-core Opteron in action
The HyperTransport speed gets a bandwidth boost from 4.8GT/sec to 6.4GT/sec. This is a common feature across all 6100 models and makes for a less confusing choice compared with the Xeon 5600 range, which offers three different QPI speeds. AMD’s new Direct Connect Architecture (DCA) 2 brings it in line with Intel’s QPI, as it adds a fourth HyperTransport link to reduce the number of hops between processors. Previously, the three links in DCA 1 meant processors in 4P systems had to make more than one hop to reference memory controlled by another processor.
DCA 2 also doubles the number of memory controllers per processor to four, allowing each socket to support up to 12 DIMMs. This gives the 6100 the edge over the Xeon 5600, which can support only up to nine DIMMs per socket.
The 6100 series is more consistent with memory compatibility, as 800MHz, 1,066MHz and 1,333MHz speeds are supported by all models. They will also be able to handle low-voltage memory when it finally comes out, although this feature is present across the Xeon 5600 series.
Next up is CoolSpeed technology, which changes the way the management controller talks to the processors. Previously, this had to go over the north and south bridges and HyperTransport links, but it now communicates directly. It offers dynamic control over p-states and can switch off other components such as memory controllers when in idle.
AMD claims some impressive power figures, saying a 2P system can consume as little as 60W in idle. The C1E (enhanced hold-state) extends the sleep state to common parts of the processor such as the HyperTransport and memory controller with a view to making more savings when in idle. We didn’t see this with the Boston system, however; with the OS ticking over, we saw power consumption fall no lower than 163W.