The 45 nm Shanghai/Deneb die from AMD. Now that’s a bit-o-cache.
Shanghai is the server level product based on the Phenom architecture. Just as Barcelona was to Agena, Shaghai is to Deneb. Deneb is the 45 nm Phenom aimed at the desktop. The 45 nm Phenoms have some significant differences between themselves and the older 65 nm Phenoms, other than just the smaller geometries that the 45 nm process enables. The first major difference is the amount of L3 cache that AMD is able to squeeze into the design. Current Phenoms have 2 MB of L3, but the upcoming Phenoms will have that expanded to a full 6 MB of L3. Considering how much the current Phenom design likes its fully functional L3 cache, this extra space will allow a greater amount of efficiency from the architecture. AMD has also done a lot of fine tuning and re-architecting to enable a higher IPC from the previous 65 nm parts. Pat goes on to say that they are seeing a 20% gain in performance per clock with the new design as compared to the old. This brings per core performance to the same type of levels as the mighty Core 2 Duo/Quad chips. Add onto that the ability to clock above 2.6 GHz without breaking any power and thermal envelopes. We see that AMD finally has a part that can compete against the Core 2 architecture in most instances. This is assuming that AMD is not feeding us a line about what to expect, and then not be able to deliver.
AMD seemingly is turning their ship around. Their graphics and chipset divisions are putting out compelling parts, and the price/performance of the current 65 nm Phenoms is pretty impressive. While AMD is not making bucketloads of money from their CPU division, they are at least putting out products that people can afford, and do give a lot of performance compared to the money spent. They have also worked their way into other markets with their triple core products as well as their low power Athlon 64 and Phenom products. The latest Radeon 4000 series of parts have achieved widespread acclaim for their performance, features, and highly competitive price. The Radeon 4870 is in the range of GTX 260 performance, and typically costs slightly less, the Radeon 4870 X2 is the fastest “single” video card out there, and the newly introduced Radeon 4670 takes gaming performance to new levels at a $79 price point. The chipset division is also fighting strong with their introductory level 740G, the higher end 780G, and the latest 790GX.
Put against that background, AMD’s potential resurgance in the server field is a welcome one for the company. The server and workstation markets are very high margin, and obviously a good money maker for AMD when they are hitting on all cylinders. The delay of Barcelona hurt that division quite a bit, but with the introduction of the 45 nm server chips, their fortunes may well change. AMD certainly has a leg up on Intel when it comes to multi-socket systems with their mature and widely adopted HyperTransport system, and with the release of the new 45 nm parts AMD will allow their partners to start utilizing the full HT 3.0 featureset and bandwidth.
As I have stated in the past, AMD typically starts shipping new products into the server market before they start releasing their desktop counterparts. This has been happening since the Athlon MP/XP days, and has been further refined to what we see now. The majority of production for the first several months of 45 nm parts will all go to the server and workstation guys, mainly because of three reasons. The first is that it is a smaller market and AMD can fulfill demand while still ramping production on these new parts. The second reason is the obviously high margins on these products that will help AMD regain as much money as possible from what was spent ramping initial production. The final reason for this mode of operation is that the server and workstation markets do not demand the highest MHz/GHz, and so AMD can still get away with shipping lower speed products but still cashing in on the nice margins. This allows AMD to further refine their 45 nm process in full production conditions, and allow them to tweak it so they can start achieving higher speed bins throughtout the next few months. When they reach a certain point where they are content with the speeds they are seeing, then they will start stockpiling parts to allow a desktop release with 45 nm parts and higher speeds.
AMD has seemingly announced this right at the end of Q3, which is when they were expecting the first production parts to be out. It seems they are slightly behind that schedule, but still targeting a mass release of 45 nm products in the upcoming Q4. We also expect to see the first 45 nm desktop parts to hit in early Q1 2009, but the official stance is that initial products will be seen later this coming quarter (but I think that may be a bit too optimistic for AMD at this time).
The long and short of this is that it is good to see AMD in a fairly competitive position across most of their lines. The big “if” is going to be the Nehalem release from Intel. This is certainly an impressive looking part, but the motherboard infrastructure is going to be immature for several months. Especially on the server side. So, AMD does have a window of opportunity here to regain some marketshare. Even if the 45 nm chips are not equal to Nehalem in per clock performance, AMD can still carve a niche out for themselves and make some money. OEMs, even though they may not see AMD as the most performant part, will gladly take a solid second supplier of parts to keep the pressure on #1. Now we only have to wait and see what AMD will do with their design and manufacturing groups with their “Asset Smart/Lite” program. Hopefully, we will get some solid answers on that as well.