AMD Phenom II X6 might pose a threat to Intel Core i7

AS CHIP DESIGNER AMD has announced its Phenom II X6 line this week, one wonders how AMD's latest line of x86 desktop processors will compete against the current performance leader, the Intel Core i7 980X, as well as its more reasonably priced mainstream LGA1156 socket siblings including the Core i7 880 and 875K.

You already know the specs of AMD's latest desktop chip. It's still fabbed on the 45nm process and six cores share 6MB of joint L3 on-die cache while their own individual 512KB L2 caches are still using the 'exclusive' approach with data in L2 not copied into L3.

That approach has its pros and cons. The plus point is that a total of 9MB of data can be stored in the combined six L2 plus one L3 cache for increased total cache capacity, and the minus point is that, whenever one CPU core needs some data from other CPU cores, it might have to search for it across all the other L2 caches plus the L3 cache, and then go to main memory if not found. Of course, improved search algorithms can cut the penalty quite a bit, but not totally.

In the case of Intel chips with their "inclusive" cache approach, the smaller 256KB L2 per-core cache contents for each CPU are all also copied to the shared L3 cache. For the 3.33GHz Core i7 980X, that's not a problem at all since it has a huge 12MB L3 cache. The benefit here is that, whatever data is required by any core beyond its own local cache, just one search in the shared L3 cache is required before going to main memory if needed.

In this case, the performance benefit for either side will depend on the application, but mostly Intel's approach errs on the side of larger total cache and fewer search hops, as latency benefits come into play.

Now, the Phenom II X6 die seemingly has more in common with AMD's latest dual die "Magny Cours" 12-core and single die "Lisbon" six core Opteron server chips than the earlier "Istanbul" six core DDR2-only Opteron server chips. The Phenom II X6 is expected to have the same core throughput enhancements as Magny Cours and Lisbon, as well as, of course, native DDR3-1600 dual channel memory support and a HyperTransport 3.1 link to the chipset at a full 6.4 Gigatransfers/sec speed, matching Intel's Quick Path Interconnect (QPI) link.

At launch, the fastest Phenom II X6 part is the 1090 model at 3.2GHz speed, even though a lot of preliminary benchmarks were done on the 3GHz 1075 and 2.8GHz 1060 parts. Ignoring the "Turbo" from both sides, as well as Intel multithreading, this now comes to within 5 per cent of the fastest Intel part's clock speed. However, Intel's parts do have a noticeable clock-for-clock performance advantage over AMD's previous Phenom chips, and Intel's six-core Westmere chips managed to widen that gap further.

Despite the $300 price range that the new Phenom II X6 is in, there's a chance for AMD to look one price bracket up assuming it can continue with performance advances. If the core and uncore improvements in the Phenom II X6 bring that per-clock difference to less than 10 per cent, and keep the clock speed differential low, AMD has a strong chance of getting back a bit of former glory in the higher margin $400 to $500 desktop CPU market, a segment where it has been absent for over two years now.

The Intel Core i7 980X is unlikely to be beaten in the performance race, and Intel can always announce a, say, 3.6GHz Core i7 990X if need be, as there is enough margin left. However, with the Phenom II X6 1090, AMD will still come close enough to, at very least, attract some attention from the buyers otherwise looking for an Intel Core i7 880 "Lynnfield" Nehalem update, for instance.

So, I do expect the Phenom II X6 1090 to benchmark at within 15 to 20 per cent of the Intel Core i7 980X performance in most benchmarks, assuming that HyperThreading is off and Turbo is not enabled. This means 15 to 20 per cent slower, just to be clear. However, it is not that great a difference, and remains to be seen in actual benchmark tests.

In overclocking, I am hearing that a lot of high end Phenom II X6 parts can easily run at 4GHz in every day production use if properly cooled on good boards, and AMD, unofficially, seems to agree. As you know, Intel's existing chips have had no problems in the 4+GHz segment for quite a while, so AMD having a six-core CPU that can do 4GHz without heroic efforts is a big plus.

However, whether that will lead to any market share improvement for AMD remains to be seen. First, it has to be able to produce as many of these high bin 1075 and 1090 chips as possible, and that volume is not known right now regarding the yields in the 45nm process. Secondly, it can't be too aggressive with pricing as the chips will still have to make acceptable profits, yet it will have to be more aggressive than Intel is in that particular market segment.

The third question revolves around the fact that - except for the Core i7 980X and of course Xeon 5600 workstation and server processors - Intel didn't bother to introduce the six core Westmere to the general PC market. Simply, the PC OS and application requirements and usage capability to maximise six cores are just not there yet. Speeding up the Lynnfield Nehalem CPUs in the LGA1156 socket by another notch or two - the Core i7 890 model slot is still unused - in Intel's mind could be a good enough answer to keep AMD at bay until Intel's Sandy Bridge and AMD's Bulldozer are launched early next year. At the same time, AMD needs to find more killer apps to show how it can optimally make use of six cores at the same price as Intel's four core chips, or alternatively offer four-cores-only versions of the Phenom II at higher clock speeds in the meantime.

Talking about the Bulldozer, the Phenom II X6 and its Opteron "Lisbon" twin are the last high end AMD CPU introductions before that next generation of chips appears. Since the AMD Bulldozer and Intel Sandy Bridge announcement dates are expected to be quite close, both vendors may want to drum up their respective cases well ahead of time. And what better way to do that than making more good sales of existing top bin CPUs before then? µ