Nvidia Corp. has confirmed in a brief media interview that it will be able to cut-down the Fermi-G300 graphics processors in order to address certain specific markets and price-points. The move is natural for all graphics chips designers, but this time Nvidia openly admits that many of implemented capabilities will hardly offer benefits for the consumer right away.
“We are paying a bit of a compute tax in that we launched a part where a lot of the consumer compute applications haven’t really taken hold yet. But over time as more consumer computer applications are developed that take advantage of our compute (consumer) features, I think it's going to give us a big leg up,” William Dally, chief scientist at Nvidia, told Cnet News.com.
The flagship Fermi graphics processor (NV60, G300, GT300) will feature 512 stream processing engines (which are organized as 16 streaming multi-processors with 32 cores in each) that support a type of multi-threading technology to maximize utilization of cores. Each stream processor has a fully pipelined integer arithmetic logic unit (ALU) and floating point unit (FPU). The top-of-the-range chip contains 3 billion of transistors, features 384-bit memory GDDR5 memory controller with ECC and features rather unprecedented 768KB unified level-two cache as well as rather complex cache hierarchy in general. Naturally, the Fermi family is compatible with DirectX 11, OpenGL 3.x and OpenCL 1.x application programming interfaces (APIs). The new chips will be made using 40nm process technology at TSMC.
At its GPU Technology Conference (GTC) last week Nvidia specifically noted that it paid a lot of attention on boosting double-precision floating point computational performance on its Fermi-G300 chip (about 750GFLOPs) and thus will be able to address new several new markets, e.g., high-performance computing. However, DP performance is rarely needed by average consumers on the desktop or laptop markets. Obviously, in order to create power efficient version of Fermi for notebooks or low-cost desktops, Nvidia will have to sacrifice some of its capabilities.
“We're not talking about other (chips) at this point in time but you can imagine that we can scale this part by having fewer than the 512 cores and by having these cores have fewer of the features, for example less double-precision,” said Mr. Dally, who did not explain how it is possible to reduce double-precision floating point performance without decreasing single-precision point speed, something which is needed by video games. In fact, Mr. Dally’s comment may imply that non-flagship Fermi derivatives will have not only be slower in terms of performance, but will be seriously different in terms of implementation.
Still, it is an open question when Nvidia releases its new family of graphics processing units for consumers. At the GTC, Nvidia demonstrated the A1 revision of the Fermi-G300 graphics chip made in late August, whereas usually the company uses only A2 or even A3 revisions on commercial products. It usually takes months to create a new revision of a chip. Besides, Fermi the graphics card that the company demonstrated at its GTC turned out to be a mock up, not a working sample. Officially Nvidia promises to start sales of Fermi products in 2009.