AMD's Opteron Comes Down Hard

The Workstation Conundrum

AMD freely admitted that the workstation market wasn't a primary target for the Opteron, which makes perfect sense for a number of reasons. AMD's strength seems to be in Systems Integrator channel, and in the digital content creation (DCC) market, the company has made inroads into production houses were price sensitivity exists. However, the big special effects houses, such as DreamWorks and Sony, are sill primarily Tier One brand name buyers.

However, if AMD manages to create a large enough installed base of 64-bit platforms with Opteron and Clawhammer, it should find it easier to get the support of workstation applications developers in key market areas such as DCC, simulation, and finance to port their software to AMD's 64-bit platform.

The one question mark that hangs over the Opteron in the workstation segment is the 8151 chipset and the way it works with AGP. AMD uses the term APG tunneling.

The 8151 with AGP Tunneling.

Into this mix you have to add an AGP Bridge: this connects the two contrasting buses of the HyperTransport interface and the AGP interface. The AGP tunnel from AMD allows AGP 8x to hook up to the Opteron HyperTransport FSB. Is this a good idea? It is logical since Opteron integrates the memory controller. This is the next best thing for an AGP attach. But, will it have issues? It could. It is more complex to convert from parallel data (AGP) to packetized data (HyperTransport). Packets are not pure data; they have overhead (commands, addresses, ECC, etc. that establish each packet of information and define its transport). This overhead can cause over-runs or under-runs that may impact performance and cause bugs.

While AMD readily admits that the workstation market is not going to be the major target for Opteron, it's worth noting that there is a slight question mark hanging over the approach taken with the 8151 chipset.

To get AMD's point of view we talked to Gabriele Sartori, Director of Technology Evangelism at AMD and President of the HyperTransport Consortium, who believes that HyperTransport is a very effective solution. He told us, "HyperTransport is very low latency, and very fast. The problem with AGP today is that it is plugged in directly to the memory controller. When you get a transaction from AGP the data may or may not be available. We don't have that problem with HyperTransport."

In addition, Mr. Sartori stressed that the greater bandwidth of HyperTransport can be calibrated depending on the application. With other bus solutions that require an embedded clock, the overhead can eat up as much as 20% of the bandwidth of the bus, further reducing the efficacy of the bus.

At the end of the day, this is all conjecture until we get our hands on real silicon and test it, yet, the issue of competing and contrasting buses is not going to go away. For instance, Intel, ServerWorks, AMD, VIA, et. al. all have their own favorite board-level interconnects; some proprietary, and some open. PCI Express and HyperTransport are both an open standard, so competition and customer demand will dictate what chip-level interconnect chipsets will choose. This issue mustn't be overlooked in assessing AMD's positioning statements.