Istanbul: Server Processor with Desktop Genealogy
Having successfully introduced 45 nm production process, AMD started manufacturing several CPU cores for mainstream desktop processors. We are very well familiar with them already, they are: Deneb that is used in Pehnom II X4, Propus used in Athlon II X4 and Regor used in Athlon II X2. AMD usually builds server processors with the cores similar to desktop ones that differing a little by external interfaces support: they have more HyperTransport busses and the memory controller is designed to support Registered memory DIMMs. For example, Shanghai core used in contemporary Opteron 2300 and 8300 processors is built exactly like that. It can be regarded as the closest relative of the desktop Deneb core.
However, Istanbul core stands out of this rule: so far it has no desktop analogue. Although we can’t deny that Istanbul, just like Shanghai, is close to Deneb in microarchitecture. However, in this case the differences go beyond the memory controller and the number of HyperTransport busses. Istanbul has six computational cores on a single processor die, which makes these processors the most expensive solutions in the Opteron lineup: they are priced starting at $450. There is nothing surprising about it: Istanbul die size is 346 mm2, which is about 1/3 larger than Shanghai die. The six-core processor has 904 million transistors. All this fits onto a monolithic semiconductor die, which holds not only six computational cores each with 512 KB L2 cache, but also a shared 6 MB L3 cache.
Istanbul processors have three HyperTransport busses, which allow using them in dual-, quad- and eight-processor systems. As for the memory controller, Istanbul works with Registered dual-channel DDR2-800/667/533 SDRAM with or without ECC support in order to maintain compatibility with existing Socket F platforms. By the way, compatibility is another specifically stressed advantage of AMD server processors: Socket F platform was first introduced three years ago, but even the latest Istanbul CPUs work perfectly fine with any Socket F mainboards after reflashing the BIOS.
Compared with their predecessors, six-core processors boast a significant improvement – HT Assist technology that reduces the CPU idling time in multi-processor systems during intensive memory operations. In systems like that cache-memory of different processors may contain copies of the same data from the system memory of different currency. That is why if the system needs to access any sort of data, the CPUs first of all send requests to other processors checking if they have more current version of the needed data in the cache. It creates big parasitic traffic along the HyperTransport bus and causes the performance during work with the memory subsystem to drop dramatically. In order to eliminate this negative effect, Istanbul processors can isolate a special area within L3 cache, that stores a constantly updated directory of data cached by all processors in the system. As a result, when the request for information from the memory is sent, each processor knows in advance, the cache-memory of which CPU in the multi-processor system it should address. This is the essence of HT Assist technology, which proves the most efficient in quad- and eight-processor systems.
However, if we regard Istanbul as a possible solution for a single-CPU platform, then there is no need for HT Assist and three HyperTransport busses. However, even in this case Istanbul may offer special advantages over Deneb and Shanghai, even if we do not take into account more computational cores. First, Istanbul supports higher HyperTransport bus frequency that has been increased to 2.4 GHz (in contemporary desktop AMD processors this frequency doesn’t exceed 2.0 GHz). Second, the built-in L3 cache inside Istanbul processor is overclocked to 2.2 GHz instead of 2.0 GHz.
As for the six-core processor clock frequencies, they are obviously lower than the frequencies of quad-core CPUs. The top AMD Opteron processor models with six cores work at 2.8 GHz. Moreover, these solutions belong to HE class that includes CPUs with higher power consumption and heat dissipation. The general-purpose six-core Opteron processors with average CPU power of 75 W work at 2.6 GHz maximum frequency, which is 30% less than the frequency of top quad-core Phenom II X4 CPUs. This must be another reason why AMD is not in a hurry to move their six-core processors into the desktop segment: at this point their clock frequencies won’t let them outperform the top quad-core CPUs even in well-paralleled tasks.
For our today’s test session we picked a six-core Opteron 2435 processor working at 2.6 GHz clock frequency. The table below lists the complete specifications of the new CPU:
Opteron 2435 is the top Istanbul model of those widely available in retail these days. It costs around $1000, but when these CPUs will be available in the desktop segment, they will most likely be considerably cheaper. Opteron 2435 processor is a server solution that is why it has a 1207-pin LGA packaging and differs substantially from the usual Socket AM2/AM3 CPUs even in its exterior looks.
Taking into account the specifications mentioned above, we would like to draw your attention to one more peculiarity of six-core Opteron CPUs, besides Registered DDR2 memory and higher operational frequencies of the North Bridge built into the processor and HyperTransport bus. The table above shows a relatively low heat dissipation of only 75 W, which seems a little surprising as the fastest desktop processor currently boast 125-140 W TDP. However, this inconsistency can be explained easily. Firstly, server CPU uses lower core voltage – this is AMD’s deliberate strategy aimed at lowering the electrical and thermal specifications of solutions targeted for rackmount servers, where it is challenging to use high-performance cooling solutions. Secondly, the physical meaning of the parameter that AMD uses to describe the heat dissipation of their server processors is slightly different from the TDP (Thermal Design Power) used for desktop CPUs. The thing is that AMD introduced a special parameter for their server processors called ACP (Average CPU Power) that is calculated not as maximum, but as average processor heat dissipation. In other words, Istanbul heat dissipation is lower than that of top Phenom II X4, but we can’t say how much lower, because it is impossible to directly compare ACP and TDP values.