Just hot air blowing out of their a$$
After being late to market with high-performance graphics offerings for a number of times, ATI, graphics product group of Advanced Micro Devices, is reportedly considering high-end graphics solutions that utilize more than two or, perhaps, even more physical dice. The method has been successfully utilized by Intel Corp., but will it be feasible for graphics processors too?
As Graphics Chips Become More Complex…
ATI Radeon HD 2900 (R600) graphics chip, which contains about 700 million of transistors had power consumption of 160W, or even more, but still did not manage to demonstrate performance on par with Nvidia GeForce 8800 GTX, a solution that also demands high amount of power and is rather expensive to manufacture. But ATI Radeon HD 3800 (RV670) graphics processing unit (GPU), which is made using 55nm process technology, has the same amount of horsepower as R600, but is cheaper to build and consumes less amount of energy.
While two of such ATI RV670 chips would still consume quite a lot of power, they will be able to offer performance and features that were not available before without necessity to develop a chip that would have about 1.3 billion of transistors, the amount of elements that would require very thin process technology – so that the GPU would stay cheap enough to manufacture – and quite a lot of time to design it and verify the lack of bugs.
It is projected that ATI Radeon HD 3800 X2 – graphics board running two ATI RV670 processors – will be announced at Consumer Electronics Show, or at another time early next year. While this graphics card will be the first multi-GPU consumer board developed by former ATI Technologies in years, it seems that multi-GPU is the future, at least when it comes to AMD’s graphics product group (GPG).
…Multi-Chip GPUs May Be the Future
It is little known about products code-named ATI R700 today, but, according to an article at PC Watch web-site, the next generation of graphics solutions from AMD may utilize multi-chip module (MCM) concept instead of multi-GPU concept, at least, in the high-end. Even though both approaches have drawbacks compared to single-chip solutions, in case of homogeneous MCM some issues are easier to solve.
Intel Corp., the world’s largest maker of chips, puts two physical dice – each containing two processing engines – onto a single piece of substrate to create quad-core central processing units. This allows Intel to boost its yields, as monolith quad-core microprocessor would have larger size of the die and would be more expensive to manufacture, according to Intel. It is rumored that AMD’s GPG may think the same way and cease to develop large GPUs, but concentrate on making smaller chips working efficiently together.
But ATI/AMD MCM graphics solutions will not be similar to Intel’s MCM CPUs. Instead of using an external bus to connect the two dice, a special chip-to-chip interface (or high-speed link) is expected to be used, which should improve their performance when they act together. Besides, it is reported that AMD’s GPG will attempt to use shared memory on its multi-dice ATI R700 graphics solutions and rely on the link between GPUs to organize their access to each other’s memory pools. The GPU dice are projected to be able to enter idle mode when their computing power is not needed, thus, preserving power consumption.
At present graphics processors in multi-GPU configuration, or on a multi-chip graphics board, communicate using special multi-GPU interfaces – dubbed ATI CrossFire or Nvidia SLI – or via PCI Express bus. While the bandwidth of CrossFire and SLI is believed to be relatively low, the bandwidth of PCI Express 2.0 x16 bus is 8GB/s in each direction, still well below of graphics cards’ memory bandwidth that can be over 100GB/s. However, chip-to-chip interfaces like Rambus’ Flex IO can provide speeds of 76.8GB/s (32GB/s read and 44.8GB/s write) and beyond, therefore, the problem of chip to-chip interface can be solved.
But an obvious problem with homogeneous multi-die GPUs is that a high-end GPU consisting of two dice will nearly always be two times faster compared to a performance-mainstream GPU with one die. Of course, AMD GPG will still be able to sell lower-clocked dual-die GPUs, but is not obvious that such solution will be viable from financial standpoint.
Therefore, with a large gap between the price and performance of single-chip and dual-chip products, it may be hard to form a comprehensive graphics card lineup that would cover all the price and performance segments. If AMD decides to install four homogeneous chips/dice on a high-end graphics card, three onto a performance-mainstream board, two on a mainstream solution and one will serve the low-end, then its driver team will have to spend a substantial amount of time tweaking each video game for single-, dual-, triple- and quad-GPU/die graphics sub-system; a task that neither ATI, nor Nvidia have so far been truly successful in.
Performance of all modern [homogeneous] multi-GPU solutions depends on drivers and in case the driver does not recognize an application, performance of a dual-, tripe- or quad-GPU graphics solution may be similar to a single-chip graphics card. Theoretically, ATI Catalyst driver developers may force so-called alternate frame rendering (AFR) multi-GPU rendering technology for all unknown applications for CrossFire configurations, but this may add lag effects in numerous games.
Or May Not
While homogeneous graphics solutions have been widely discussed in the recent years, the first successful consumer 3D graphics accelerators as well as professional 3D boards for workstations until the recent years utilized heterogeneous multi-chip architecture, where all (or nearly all) the chips onboard had different functionality. Maybe, this is the way to go?
Nowadays different units within a graphics processor have different memory bandwidth requirements, moreover, certain parts of the chips do not communicate with others. As a result, a heterogeneous multi-GPU, or heterogeneous multi-die GPUs, may become feasible solutions against ultra-complex single-chip GPUs and triple-/quad-core homogeneous multi-GPU/multi-die solutions.
Unfortunately, heterogeneous multi-GPU/multi-die GPU solutions will almost certainly not be viable for mainstream and entry-level graphics cards, where price matters quite a lot. Thus, ATI/AMD will have to develop single-chip solutions for segments of the market where price matters and create heterogeneous multi-chip/multi-dic graphics products for those, who demand ultimate performance. In both cases the gap between the price and performance of mainstream and high-end graphics sub-systems is likely to be fairly high.
At the end, both homogeneous multi-GPU/multi-die GPUa as well as heterogeneous multi-GPU/multi-die GPUs have their advantages and disadvantages. So, maybe a single-chip high-performance graphics card has still a reason to live?
Officials for ATI, graphics product group of AMD, did not comment on the news-story.