Athlon 64 and Opteron dual-core explained

ONE OF THE MORE intriguing things about future processors is the move towards multiple cores. It's a feature that some specialist high-end processors have had for a while and it can mean a significant improvement in performance. When AMD launched the Athlon 64, it came as a pleasant surprise that the firm had designed for the possibility of dual-core AMD64 chips from the outset.

Some might consider the inclusion of dual-core capability into an already late design to be over-engineering but the fact that it is there is likely to reap significant rewards for AMD. Modern compilers are aware that cache misses cause a huge performance hit so they try to avoid fetching data from the main memory as much as possible.

One of the main reasons that Intel introduced HyperThreading was to alleviate the effects of a cache miss but, so far, that technology has proved to be far less effective than many chip designers had hoped. The essential idea of HyperThreading is to build multiple execution units into a single processor core, if one execution unit has a cache miss, at least the other is still doing some useful work. In practice it seems that the two execution units are constantly fighting for data, often slowing the system down marginally.

Having dual processor cores is a bit different. Where the execution units in a HyperThreading processor have to share the cache, in dual-core chips, each core has a separate cache, the only contention is when going to main memory.

Dual-core SRQ
So far, AMD has decided to avoid using anything similar to HyperThreading. It took the decision to, if necessary, head down the dual-core path instead. Given the current AMD64 architecture, it would be quite difficult for the firm to produce anything similar to HyperThreading, the firm having chosen to concentrate on using HyperTransport for inter-processor communications.

The dual-core concept fits in perfectly with HyperTransport. The diagram below shows how AMD designed dual-core capability in from the beginning.

The main feature to note is that the System Request Queue can handle the two cores, 0 and 1. In the firm's current processors, only core 0 is present but, at the launch of the Athlon 64, AMD made a big deal of the fact that adding core 1 would require minimal work. Indeed, when Intel announced it would be introducing dual-core processors, AMD was ready to respond.

Strengths
The great advantage of dual-core over HyperThreading is that both cores have their own cache. It means that the only contention is over access to main memory and HyperTransport. Provided the level 2 cache is a decent size for each core, that contention should be kept to a fairly low level. But should is the operative word. Intel's HyperThreading should have had a significant impact on performance but that impact turned out to be in the negative rather than the positive that Intel had hoped for. That's not to say that HyperThreading won't start making improvements in future applications that are aware of the technology.

However, AMD will gain from HyperThreading aware applications too as Intel designed the technology to appear as though there were dual processors. A dual-core processor will do exactly the same so the both Intel and AMD are likely to reap any benefits from the move.

Weak Points
It's all very well saying that dual-core is better than dual execution units (HyperThreading) but there are some problems with following that route. The biggest spanner in the works is memory contention. As long as the applications that are running fit comfortably into the cache, everything will be fine but these are the days of large datasets. Manipulating images, video, databases and pretty much anything else these days means that the main memory gets used a lot. It's why having the memory controller built in to the Athlon 64 and Opteron caused such a significant performance boost.

Two cores fighting for data from the main memory is nothing new, you can see exactly the same thing in older dual processor designs like the Xeon and the Athlon MP series. Both of those processors have to request memory access from a northbridge chip and it means that you rarely get anything more than a 50% gain in performance over using a single processor. The Athlon 64 and Opteron dual-core design is little more than that old dual processor architecture compressed onto a single chip.

There are definite advantages to everything being on a single chip, most noticeably a reduction in latency, but it's unlikely that a system with one dual-core Opteron processor would ever manage to outpace a system with two single-core Opterons. The latter has two lots of main memory so it simply doesn't have the same contention issues.

There are some dual Opteron motherboards on the market that only have one lot of main memory, the second processor gaining access to that memory via its HyperTransport link. Needless to say, systems built using those motherboards are significantly slower than systems where both Opterons have their own DRAM. Performance wise, a single dual-core Opteron is likely to sit between those two.

The Competition
The biggest question still waiting to be answered is what Intel is up to memory architecture wise with its dual-core processor plans. If the firm intends to stick with the older northbridge-based memory architecture, it could find itself left even further behind. Perhaps there is some significance in the fact that Intel's next socket is 775 pins. Note how close that number is to AMD's 754 pin Athlon 64 socket, it could be that Intel has decided to go the same route as AMD and build the memory controller into the chip.

One thing is for certain, AMD has shown that it was at least one step ahead of Intel when it designed its current AMD64 chips. Although the dual-core processors will have some weaknesses, those weaknesses will be present in the processors from both firms. In mid to late 2004, AMD will introduce dual-core chips that have taken it next to no effort to produce; its step ahead will be used up. It remains to be seen if AMD can keep up this pressure on Intel after it has used up this advantage. Intel's huge research and development budget is almost certainly being poured into finding ways to cut the upstart down to a size that Intel shareholders would be more comfortable with. µ