NVIDIA Fermi Architecture
NVIDIA promises the GeForce GF100 GPU will be the most powerful graphics processor ever built, and on paper it appears the Fermi GPU architecture can deliver. Benchmark Reviews has tested AMD's latest graphics processor, the 40nm ATI Radeon Cypress GPU, which has held command of the discrete desktop graphics industry with its 1600 shader cores since first released on the Radeon HD 5870. But not all GPU cores are the same, and although ATI uses 1600 in their Juniper-XT, the end result is not a sum of its cores but rather a sum of the total package.
Running up to the NVIDIA GF100 GPU, the 40nm process has been nurtured to maturity after nine months of production at TSMC. After plenty of practice, patience, and some design tweaks, the NVIDIA GF100 GPU is the highest-performing graphics processor they've ever built. At the core of Fermi's architecture are improvements to visual quality and GPU-compute productivity, intended to deliver the best gaming experience available from a video card. In this section, we analyze NVIDIA's Fermi architecture and compare the incoming GF100 against the outgoing GT200 graphics processor.
GF100 is not another incremental GPU step-up like we had going from G80 to GT200. While processor cores have grown from 128 (G80) and 240 (GT200), they now reach 512 and earn the title of NVIDIA CUDA (Compute Unified Device Architecture) cores. The key here is not only the name, but that the name now implies an emphasis on something more than just graphics. Each Fermi CUDA processor core has a fully pipelined integer arithmetic logic unit (ALU) and floating point unit (FPU). GF100 implements the new IEEE 754-2008 floating-point standard, providing the fused multiply-add (FMA) instruction for both single and double precision arithmetic. FMA improves over a multiply-add (MAD) instruction by doing the multiplication and addition with a single final rounding step, with no loss of precision in the addition. FMA minimizes rendering errors in closely overlapping triangles.
GF100 Specifications
•512 CUDA Cores
•16 Geometry Units
•4 Raster Units
•64 Texture Units
•48 ROP Units
•384-bit GDDR5
•DirectX-11 API Support
Based on Fermi's third-generation Streaming Multiprocessor (SM) architecture, GF100 doubles the number of CUDA cores over the previous architecture. NVIDIA GeForce GF100 Fermi GPUs are based on a scalable array of Graphics Processing Clusters (GPCs), Streaming Multiprocessors (SMs), and memory controllers. The NVIDIA GF100 implements four GPCs, sixteen SMs, and six memory controllers. Expect NVIDIA to launch GF100 products with different configurations of GPCs, SMs, and memory controllers to address different price points.
CPU commands are read by the GPU via the Host Interface. The GigaThread Engine fetches the specified data from system memory and copies them to the framebuffer. GF100 implements six 64-bit GDDR5 memory controllers (384-bit total) to facilitate high bandwidth access to the framebuffer. The GigaThread Engine then creates and dispatches thread blocks to various SMs. Individual SMs in turn schedules warps (groups of 32 threads) to CUDA cores and other execution units. The GigaThread Engine also redistributes work to the SMs when work expansion occurs in the graphics pipeline, such as after the tessellation and rasterization stages.
GF100 implements 512 CUDA cores, organized as 16 SMs of 32 cores each. Each SM is a highly parallel multiprocessor supporting up to 48 warps at any given time. Each CUDA core is a unified processor core that executes vertex, pixel, geometry, and compute kernels. A unified L2 cache architecture services load, store, and texture operations. GF100 has 48 ROP units for pixel blending, antialiasing, and atomic memory operations. The ROP units are organized in six groups of eight. Each group is serviced by a 64-bit memory controller. The memory controller, L2 cache, and ROP group are closely coupled-scaling one unit automatically scales the others.
GF100 Streaming MultiProcessor
NVIDIA's third generation Streaming Multiprocessor (SM) introduces several architectural innovations that make it their most programmable and efficient yet. Inside each of the 16 streaming multiprocessors are 32 CUDA processors (4x increase over GT200's SM design). Fermi's GF100 CUDA cores are designed for maximum performance and efficiency across all shader workloads. By employing a scalar architecture, full performance is achieved irrespective of input vector size. Operations on the z-buffer (1D) or texture access (2D) attain full utilization of the GPU.
In GF100, the newly designed integer ALU supports full 32-bit precision for all instructions, consistent with standard programming language requirements. The integer ALU is also optimized to efficiently support 64-bit and extended precision operations. Various instructions are supported, including Boolean, shift, move, compare, convert, bit-field extract, bitreverse insert, and population count.
Each SM has 16 load/store units, allowing source and destination addresses to be calculated for sixteen threads per clock. Supporting units load and store the data at each address to cache or DRAM. Four separate Special Function Units (SFUs) execute transcendental instructions such as sin, cosine, reciprocal, and square root. Graphics interpolation instructions are also performed on the SFU. Each SFU executes one instruction per thread, per clock; a warp (32 threads) executes over eight clocks. The SFU pipeline is decoupled from the dispatch unit, allowing the dispatch unit to issue to other execution units while the SFU is occupied. Complex procedural shaders especially benefit from dedicated hardware for special functions.
•32 CUDA Cores (4x GT200)
•48KB or 16KB of shared memory (3x GT200)
•48KB or 16KB of dedicated L1 cache (none on GT200)
•768KB of L2 cache (256KB on GT200)
•ISA improvements: single/double precision with 32-bit integer operations
•4 Texture units
•New PolyMorph engine
The new Fermi architecture is designed to enable 16/48KB of on-die cache storage for improved performance with minimal trip penalty. While the dedicated L1 and texture L1 caches handle most operations, a 768KB L2 cache delivers greater texture coverage and improved computer performance over previous L2 cache (GT200).
The number of ROP (Render Output) units per ROP partition is doubled to 48 and fillrate is greatly improved as a result. 8x MSAA graphics performance is also improved through enhanced ROP compression, and allows 32x CSAA antialiasing mode. The additional ROP units better balance overall GPU throughput, even for portions of the scene that cannot be compressed.
GF100 implements a new 32x CSAA antialiasing mode based on eight multisamples and 24 coverage samples. CSAA has also been extended to support alpha-to-coverage on all samples, enabling smoother rendering of foliage and transparent textures. GF100 produces the highest quality antialiasing for both polygon edges and alpha textures with minimal performance penalty. Shadow mapping performance is greatly increased with hardware accelerated DirectX 11 four-offset Gather4.
Fermi GF100 Gaming Features
During the 2010 International Consumer Electronics Show (aka CES), NVIDIA was host to a day-long deep technical session that covered architecture, technology, and implementation. While the bulk of this article has been spent discussing the former, this section will detail the latter with some examples of Fermi in real-world gaming action. NVIDIA's Tony Tamasai (a familiar personality from the NVIDIA Editor's Day 2008 event) introduced GF100 gaming, and with the help of development staff Benchmark Reviews sat back and enjoyed the show.
NVIDIA hopes to push a state-of-the-art experience on video games with the GeForce GF100 GPU. While gamers are familiar with NVIDIA GeForce products and drivers, the company partners with developers on a routine basis to ensure that all games developed (regardless of video card support) will operate flawlessly and perform optimally with GeForce products. NVIDIA estimates that 90% of their resources are dedicated towards video game engineering, although numbers are considered an industry secret. Tony Tamasai mentions that NVIDIA has often worked with game developers up to 2-4 years before production, and help introduce new tools into their SDK to enable modern engineering paired to fresh game titles.
There were numerous in-house test results quoted in the NVIDIA press deck during the seminar, but only one DirectX-11 video game actually ran in real-time before our eyes. Dark Void is a video game that NVIDIA first began working on with Airtight around September 2008, and assisted in adding GPU PhysX support to enhance jet-pack turbulence and particle weapon effects. Of course, a game that allows the player to blast into the atmosphere isn't complete without GeForce 3D Vision support.
Even some older game titles benefit by the NVIDA GF100 GPU, beyond just an increase in frame rates. For example, Far Cry 2 will receive 32x CSAA functionality native to the game, but drive updates could also further add new features into existing co-developed video games.
Supersonic Sled
At the end of our deep tech-dive day, NVIDIA's Mark Daly regaled our attention with a demonstration of the Supersonic Sled mini-game. The fundamental purpose of the soon-to-be public demo is to illustrate new GF100 feature sets such as thrust turbulence and particle behavior. The premise is rather simple: control your supersonic rocket sled down the railway while keeping the pilot (and thrusters) in-tact. It's harder than it sounds, especially since you can pelt the sled with chickens and drop boxes at will. I suggested that NVIDIA use corn-on-the-cob instead of chickens, and include popcorn particles if they near the thrust. They said it was a great idea that could be easility implemented. We'll see.
NVIDIA 3D-Vision Surround
The original release of 3D-Vision was already impressive, as it truly earned our respected Benchmark Reviews Editors Choice Award for 2009. So how can 3D-Vision be outdone? The answer: allow it to work with up to three monitors. The newly dubbed NVIDIA 3D-Vision Surround (stereo) requires three 3D-Vision capable LCD, projector, or DLP devices and offers bezel correction support. Alternatively, NVIDIA Surround (non-stereo) supports mixed displays with common resolution/timing. 3D-Vision Surround is supported on GF100, as well as GTX 200 GPUs, but requires at least two video cards to enable (due to device connection restrictions).
NVIDIA GF100-based video cards share the same two-device pipeline limit for video output that previous GeForce GT200-based products had. When I asked Jason Paul as to why NVIDIA has not overcome this hurdle with Fermi despite three video outputs (2x DVI + HDMI) being available on the video card, he stated that this has been a GeForce GPU limitation that has existed for years. Ultimately, you'll be required to use at least two NVIDIA GeForce products in SLI to output three video streams and enjoy NVIDIA 3D-Vision Surround.
Final Thoughts
NVIDIA has a tough decision on their hands, and Fermi is only the start of things to come. If GPGPU is to continue as it has in the GF100, then NVIDIA will one day need to hedge their bets and develop separate processors: one for retail consumers and another for high-performance computing (HPC). Making things even more challenging to chip designers is the push towards APUs, or Accelerated Processing Units. You'll hear this term used more and more once (if and when) products like Intel's Larrabee or AMD's Fusion finally make it to market.
At the beginning of this article I compared the way GPUs have overtaken CPUs in terms of technical development progress with each new architecture. It's easy to make an assertion like this when neither the CPU nor GPU are every going to be directly comparable. To begin with, CPUs are strong at handling a wide variety of operations or controlling applications requiring limited threads, while GPUs are still best suited for multi-threaded applications with limited range of operations. It's true that one day the APU will have both side of the house under control, but until then NVIDIA is setting the pace for innovation while CPUs merely get bigger and faster.
As for NVIDIA's Fermi architecture and its real-world performance results in GeForce GF100-based video cards: wait and see. We could always take NVIDIA on their word that GF100 is a remarkable improvement, but Benchmark Reviews really enjoys living up to its name. Samples are expected sometime before March 2010, and you'll see our results posted in our Featured Reviews: Video Cards section as soon as they're available. If the in-house test results are correct, NVIDIA will have easily made-up for the lost ground they've suffered to ATI's Radeon HD 5000 series.