Magny-Cours is comprised of two 4 or 6 core dies, and as such has some interesting characteristics. First off, it requires a special socket as compared to current Opterons. Each CPU supports 4 DDR-3 memory channels, which should give plenty of memory bandwidth for the density of the compute power. Very little has been talked about with the inner workings of how they are fusing these two cores on each substrate together, but obviously they are connected by multiple HyperTransport connections. This is something that many around the industry were confused about in the past… namely, why didn’t AMD produce dual core and quad core products before they did with previous generation parts (eg. two single core Athlon’s to make a dual core, or two dual core Athlons to make a quad core)? One would have thought that HyperTransport would have been robust enough to encourage such a product
Heat and power is a concern for AMD here, and obviously we will not see and 8 or 12 core parts hitting 3 GHz or more. I would hazard a guess and say that we would expect to see the highest clocked, highest TDP part to be at 2.2 GHz. TDPs for these parts are expressed as ACP by AMD, and we should see 55 W, 75 W, and 105 W processors. Even though the clockspeeds will not be nearly as high as what we would hope, the actual computing density of these parts is pretty impressive. A 4S system with 48 cores and 16 DDR-3 memory channels makes for a mouthwatering VM server.
Another positive offshoot of this release is that we can expect to see the newly redesigned quad core and six core desktop Phenom IIs based on the slightly revised Istanbul chip. AMD is being closemouthed about what all internal changes they have made, but apparently they are big enough to differentiate these parts from the older 45 nm Phenom IIs. I have heard it whispered that these upcoming Phenom II X4s are in fact standalone designs, and not just a six core die with two cores fused off (though there will likely be some of those to help recover yields).
Though AMD is sampling the Fusion design to partners in the first half of this year, GLOBALFOUNDRIES is still not ready to mass produce 32 nm parts. Fusion is again a slight reworking of the Phenom II architecture with better performance per clock than what we have here. Of course, the big change is the addition of a fully functional DX11/Compute portion to the product. Unlike Intel’s Clarksdale chip, which features a 32 nm Nehalem based core attached to a 45 nm graphics/memory controller chip, this is a fully integrated 32 nm part. I believe that AMD is really hoping to create a software environment where a lot of the heavy floating point work will be directed to the streaming portion of the graphics side.
Once GLOBALFOUNDRIES starts churning out Fusion parts, you can bet that AMD will take a good look at shrinking the Istanbul core, especially if there are any further delays for the Bulldozer architecture. Hopefully for consumers and AMD itself, we will not see any delays in Bulldozer and can expect to see initial samples coming out in late 2H of this year.
