NVIDIA GF100 Architecture and Feature Preview

Back in late September of last year, NVIDIA disclosed some information regarding its next generation GPU architecture, codenamed "Fermi". At the time, actual product names and detailed specifications were not disclosed, nor was performance in 3D games, but high-level information about the architecture, its strong focus on compute performance, and broader compatibility with computational applications were discussed.

We covered much of the early information regarding Fermi in this article. Just to recap some of the more pertinent details found there, the GPU codenamed Fermi will feature over 3 billion transistors and be produced using TSMC's 40nm processes. If you remember, AMD's RV870, which is used in the ATI Radeon HD 5870, is comprised of roughly 2.15 billion transistors and is also manufactured at 40nm. Fermi will be outfitted with more than double the number of cores as the current GT200, 512 in total. It will also offer 8x the peak double-precision compute performance as its predecessor, and Fermi will be the first GPU architecture to support ECC. ECC support will allow Fermi to compensate for soft error rate (SER) issues and also potentially allow it to scale to higher densities, mitigating the issue in larger designs.  The GPU will also be execute C++ code.

NVIDIA's Jen-Hsun Huang hold's GF100's closest sibling, Fermi-based Tesla card

During the GPU Technology conference that took place in San Jose, NVIDIA's CEO Jen-Hsun Huang showed off the first Fermi-based Tesla-branded prototype boards, and talked much of the compute performance of the architecture. Game performance wasn't a focus of Huang's speech, however, which led some to speculate that NVIDIA was forgetting about gamers with this generation of GPUs. That obviously is not the case, however. Fermi is going to be a powerful GPU after all. The simple fact of the matter is, NVIDIA is late with their next-gen GPU architecture and the company chose a different venue--the Consumer Electronic Show--to discuss Fermi's gaming oriented features.

GF100 High-Level Block Diagram

With desktop oriented parts, Fermi-based GPUs will here on in be referred to as GF100. As we've mentioned in previous articles, GF100 is a significant architectural change from previous GPU architectures. Initial information focused mostly on the compute side, but today we can finally discuss some of the more consumer-centric details that gamers will be most interested in.

At the Consumer Electronics Show, NVIDIA showed of a number of GF100 configurations, including single-card, and 2-way and 3-way SLI setups in demo systems. Those demos, however, used pre-production boards that were not indicative of retail product. Due to this fact, and also because the company is obviously still working on feverishly on the product, NVIDIA chose NOT to disclose many specific features or speeds and feeds of GF100. Instead, we have more architectural details and information regarding some new IQ modes and geometry related enhancements.

In the block diagram above, the first major changes made to GF100 become evident. In each GPC cluster--there are four in the diagram--newly designed Raster and Polymorph Engines are present. We'll give some more detail on these GPU segments a little later, but having these engines present in each GPC segment essentially allows each one to function as a full GPU. The design was implemented to allow for better geometry performance scalability, through a parallel implementation of geometry processing units. According to NVIDIA, the end result in an 8X improvement in geometry performance over the GT200. Segmenting the GPU in this way also allows for multiple levels of scalability, either at the GPC or individual SM unit level, etc.

Each GF100 GPU features 512 CUDA cores, 16 geometry units, 4 raster units, 64 texture units, 48 ROPs, and a 384-bit GDDR5 memory interface. If you're keeping count, the GT200 features 240 CUDA cores, 42 ROPs, and 60 texture units. The geometry and raster units, as they are implemented in GF100, are not in the GT200 GPU. The GT200 also features a wider 512-bit memory interface, but the need for such a wide interface is somewhat negated in GF100 in that the GPU uses GDDR5 memory which effectively offers double the bandwidth of GDDR3, clock for clock.

If we drill down a little deeper, each SM core in each GPC is comprised of 32 CUDA cores, with 48/16KB of shared memory (3 x that of GT200), 16/48KB of L1 (there is no L1 cache on GT200), 4 texture units, and 1 PolyMorph Engine. In addition to the actual units, we should point out that improvements have also been made over the previous generation for 32-bit integer operations performance and for full IEEE-754 2008 FMA support. The increase in cache size and the addition of L1 cache were designed to keep as much data on the GPU die as possible, without having to access memory.

The L1 cache is used for register spilling, stack ops, and global loads and stores, while the L2 cache is for vertex, SM, texture, and ROP data. According to NVIDIA, the GF100's cache structure offers many benefits over GT200 in gaming applications, including faster texture filtering and more efficient processing of physics and ray tracing, in addition to greater texture coverage and generally better overall compute performance.

The PolyMorph and Raster Engines in the GPU perform very different tasks, but in the end result in greater parallelism in the GPU. The PolyMorph Engines are used for world space processing, while the Raster Engines are for screen space processing. There are a total of 16 polymorph engines placed before each SM. They allow work to be distributed across the chip, but there is also intelligent logic in place designed to keep the data in order. Communications happen between the units to ensure the data arrives in DRAM in the correct order and all of the data is kept on die, thanks to the chip's cache structure. Synchronization is handled at the thread scheduling level. The four independent Raster Engines serve the geometry shaders running in each GPC and the cache architecture is used to pass data from stage to stage in the pipeline. We're also told that the GF100 offers 10x faster context switching over the GT200, which further enhances performance when compute and graphics modes are both being utilized.