From GTC 2012: AMD R7970 Preferred over NVIDIA Kepler in Real GPGPU Deployments?
During the Emerging Companies Summit at the NVIDIA GPU Technology Conference, we spoke with several startups and experienced companies, seeing an interesting emerging pattern.
Apparently, if a company requires a quick, raw compute speed where there isn’t a lot of need for complex coding, such as the image or file compression and decompression, that company will be advised for a following cadence – GTX 280 – GTX 480 – GTX 580 – HD 7970.
That’s right, while the Kepler GPU architecture offers a lot more capabilities, there are a lot of applications with simple code that requires a lot of raw horsepower. In such case scenario, 2048 cores inside a Radeon HD 7970 will mean more than 1536 cores on a GTX 690. This lead to clients ordering Radeon HD 7970 boards rather than going GeForce GTX 680 or a product which NVIDIA would prefer to be used in GPGPU – the Tesla C2075, or the new K10.
OpenCL is being pushed as the API of choice, which is a testament to the Khronos Group. For the future of the GPGPU computing, we’ve been told that there is a lot of interest in moving away from the conventional x86 CPUs and replacing them with ARM cores. However, ARM needs to resolve the bandwidth issue, since GPU requires much more data than ARM is able to provide. You can put this in “good and bad news for Intel” category.
To us, all of this sounds ideal for AMD, and our sources went on to say that the real limitation is in the way how AMD positions its products. “AMD went for cheap and affordable”, said our source. "They don’t have anything for the servers and that’s what guts them. Our clients go AMD 7970, but then we have to go with a 5U chassis and put up to eight Radeon HD 7970 cards, but we cannot compete on space optimized 1U and 2U, since AMD has no purpose-built (GPGPU) products for that."
And here lies the sad truth. GPGPU is turning to be a big business for NVIDIA, earning more money than AMD is able to pull from their consumer endeavours. BUT, since there are no contemporary server-grade parts from AMD and “if the company requires a 2U or 1U, AMD solution is SOL and we go NVIDIA”.
The quotes above come from several C-level executives, and you can see how similar they are.
Here’s an advice for AMD – if you guys want to make money using your existing technology, how about qualifying a HD 7970 with 6GB of ECC memory and call it FireStream 7970? Just that there aren’t any surprises when next round of financial results come in and NVIDIA once more demolishes AMD by effectively making a technology which is inferior in simple GPGPU apps.