Deep learning and artificial intelligence have been huge topics of interest in 2016, but so far most of the excitement has focused on either Nvidia GPUs or custom silicon hardware like Google’s TensorFlow. We know Intel is working on upcoming Xeon Phi-class silicon to throw at these problems, and AMD wants to enter the market too, courtesy of a new lineup of graphics cards based on three different product families. AMD will also offer its own software tools and customized libraries to accelerate these workloads. It’s still fairly early days for the AI and deep learning markets, and AMD could definitely use the cash — but what’s it going to bring to the table?
First up, let’s talk about the accelerators themselves. AMD is deploying three new cards under its new Radeon Instinct brand, from three different product families:
The MI6 is derived from Polaris, albeit Polaris running at a slightly lower clock than the boost frequencies we saw on consumer parts (total onboard RAM, however, is 16GB). The MI8 is a smaller GPU built around R9 Nano and clocked at the same frequencies, with the same 4GB RAM limitation. (It’s not clear how much AI and deep learning workloads depend on RAM, but AMD presumably wouldn’t sell the chip into this market if it didn’t have a viable use-case for it. Finally, the MI25 will be a Vega-derived chip that’s expected to be significantly faster than the other two cards, but AMD isn’t giving any details or information on that core yet. AMD hasn’t specified a ship date for any of these products beyond H1 2017, but we’d expect the company to bring its MI6 and MI8 cards out first, to test the waters and establish a foothold in the market.
It might seem crazy to think that AMD would seek to compete against Nvidia with older and midrange consumer hardware, but it’s probably a smart move. Nvidia still sells a range of HPC products based on Maxwell and Kepler hardware, and AMD’s GCN was actually a very strong competitor against Nvidia in a number of compute workloads. Toss in the fact that AMD continues to offer a CUDA compatibility layer, and Team Red has a plausible argument for its own hardware, at least if it brings pricing in appropriately (and in the HPC world, “appropriately” can still be plenty profitable). The question, however, is how many resources AMD will be able to dedicate to the software side of this particular equation, and whether it can overcome Nvidia’s near-decade lead in GPGPU computing.
Of all the reasons we’ve heard for why Nvidia took such a leadership position in HPC and scientific computing, one of the most consistent has nothing to do with hardware comparisons. AMD held a leadership position in multiple compute benchmarks and workloads during the Kepler and Maxwell eras, often by enormous margins (this is part of why AMD GPU prices spiked in 2013-2014). OpenCL, however, wasn’t really in a state to capitalize on the strength of AMD’s underlying hardware, and AMD didn’t have the resources to spend on a major bring-up or enterprise computing push. Since then, we’ve seen incremental progress on this front, with last years’ Boltzmann initiative, various server and virtualization product launches, and now the Radeon Instinct brand. Radeon Instinct products will use AMD’s MIOpen GPU accelerated library to “to provide GPU-tuned implementations for standard routines such as convolution, pooling, activation functions, normalization and tensor format” while its ROCm deep learning network “is also now optimized for acceleration of popular deep learning frameworks, including Caffe, Torch 7, and Tensorflow, allowing programmers to focus on training neural networks rather than low-level performance tuning through ROCm’s rich integrations. ROCm is intended to serve as the foundation of the next evolution of machine intelligence problem sets, with domain-specific compilers for linear algebra and tensors and an open compiler and language runtime.”
AMD is also partnering with some hardware customers to build custom Zen systems for server rack deployments with varying numbers of accelerator cards in them, but obviously this hardware won’t be available for quite some time, since Zen’s server launch isn’t expected until Q2 2017. We expect to see both Zen and Vega in consumer hardware first, before launching for server.
It’s good to see AMD pushing for markets where its graphics cards might be particularly well-suited, given GCN’s historic compute strengths, but it’s not clear if it’ll be able to muster the software expertise to win market share. Nvidia has been plugging away at this for nearly ten years and Intel has boatloads of cash to throw at the problem. Between those two companies, there may not be much room for AMD at the proverbial table. While AMD took pains to call out its expertise in heterogeneous computing and implied this could give it a leg up once Zen is shipping, that’s a very tenuous argument right now. Nearly three years after Kaveri launched, I’m not aware of any significant software with HSA support, and AMD’s presence in the GPGPU market is anemic at best. Easy-to-use tools and compatibility with both OpenCL and CUDA could change that going forward, but this is a long-term play. It’ll take a few more years before we can fairly gauge whether it’s a success.