3 ms·
AMD should put a good effort into GPUs too, and replace GCN with brand new and efficient architecture.
by shmerl 9y ago
AMD should put a good effort into GPUs too, and replace GCN with brand new and efficient architecture.
- IntelMiner 9y agoATi hasn't really competed with Nvidia at the top end of the market for a very long time They've basically thrown in the towel as far as the datacenter went with CUDA
- shmerl 9y agoSo they should make an effort to compete more.
- IntelMiner 9y agoThat's an incredibly ignorant comment Just because they can trade blows with Intel in the CPU space, doesn't automatically translate to them having equal engineering talent to take on Nvidia on GPU's
- shmerl 9y agoI didn't say they have talent (which they might have anyway). I said they should compete, which means they should find needed talent obviously if they don't have it.
- dragontamer 9y agoThey're behind in a number of areas. Their ROCm environment finally is catching up with NVidia's software / CUDA environment. OpenCL was simply a disaster and hampered AMD's role as a serious software solution for years. I think its more important to get ROCm fully functional with Tensorflow and other technologies (their current path). AMD can compete on price/performance, selling Vega HBM2 chips at ~$1000 to $2000 (compared to NVidia's chips at $8000 for their HBM2-based V100). With AMD's drivers properly in the Linux kernel and a nicer licensing model, AMD can achieve a position in the server world. But only if developers start coding in ROCm instead of CUDA. Honestly, I bet you that software is the bigger issue in the near term. No one wants to code in OpenCL / separated C-environment. ROCm achieves CUDA-parity to some degree with "single source" C++ programming and is beginning to become compatible with Tensorflow. Sure, they'll be slower than NVidia. But at least your python machine learning code will actually run at all. And paying 1/8th the price for ROCm acceleration will be fine as long as AMD is 1/4th the performance or better. That's how you can actually build a value argument.
- se6 9y agoOpenCL is not a disaster at all. It is just that NVidia were (and still are) too scared to have people move away from their proprietary solutions, so they tried to hide OpenCL has much as they could and only pushed Cuda. Even today OpenCL is a viable solution for GPU. It works fine on both AMD and NVidia GPUs. It is also pushed a lot by Intel for FPGA, which probably scare even more NVidia. OpenCL kernels are compiled at runtime, which is brilliant since you can change the kernels code at run time, use constants in the code at the last moment, unroll, etc.. which can gives better performances. (Nvidia only introduced the possibility of having runtime compilation as a preview in Cuda 7!) The "single source" argument is completely overrated. Furthermore, you can have single source in OpenCL putting the code in strings.
- aseipp 9y ago> It is also pushed a lot by Intel for FPGA, which probably scare even more NVidia. These tools aren't available for the wide majority of developers, and are still exceptionally difficult to use and maintain without hardware engineers. I'm going to assume you haven't used FPGAs at all? The ones that can compete at the same tasks for GPUs are not as easily available in terms of price, volume, or even over-the-counter availability (be prepared to ask for a lot of quotes), and the tools have only become more accessible very recently -- such as Intel slashing the FPGA OpenCL licensing costs, and Dell EMC shipping them in pre-configured rack units. > Nvidia only introduced the possibility of having runtime compilation as a preview in Cuda 7 In the mean time, Nvidia also completely dominated the market by actually producing many working middleware libraries and integrations, a solid and working programming model, and continuously refining and delivering on core technology and GPU upgrades. Maybe those things matter more than runtime compilation and speculative claims about peak performance... > The "single source" argument is completely overrated. Even new Khronos standards like SYCL (built on OpenCL, and which does look promising, and I'm hoping AMD delivers a toolchain after they get MIOpen more fleshed out) are moving to the single-source model. It's not even that much better, really, but development friction and cost of entry matters more than anything, and Nvidia understood this from day one. They understood it with GameWorks, as well. They plant specialist engineers "in the field" to accelerate the development and adoption of their tech, and they're very good at it. This is because their core focus is hardware and selling hardware; it's thus in their interest to release "free" tools that require low-effort to buy into, do as much dirty integration work as possible, and basically give people free engineering power -- because it drives their hardware sales. They basically subsidize their software stack in order to drive GPUs. > Furthermore, you can have single source in OpenCL putting the code in strings. This is a joke argument, right?
- hajile 9y agoMost of Nvidia's architectural changes have moved it closer to GCN (eg, adding async compute in Pascal). In other areas like multiple compute engines (A reason GCN generally has better VR latency numbers), Nvidia still lags behind (not sure if Volta changes that). AMD generally does quite well on compute tasks compared to Nvidia and is still generally much faster for anything using Vulkan. http://ext3h.makegames.de/DX12_Compute.html http://ext3h.makegames.de/DX12_Compute.html The big issue with AMD is software and money. Even with the most talented engineers it takes loads of man hours to build up an ecosystem. AMD's is currently valued at 8.3B. Nvidia currently has 7.1B cash on hand. With that kind of money, Nvidia will do things like loan AAA gamemakers some devs to re-write large parts of the game to optimize them for Nvidia graphics cards or build up the CUDA ecosystem.
- ahartmetz 9y agoIt seems like AMD has hardware bottlenecks (or missing optimizations relative to Nvidia) in the fixed-function geometry processing parts of GCN. The compute benchmarks are always better than the graphics benchmarks. If it was only optimizations in games, there would be outliers in less popular games (Doom is a bit of an outlier but popular). And in any case, Xbox and PS4 consoles use AMD GPUs so game devs have reason to optimize for AMD. Also, AMD has been improving its Windows drivers in the last 2-3 years according to reports. I use AMD GPUs on Linux where the driver codebase is different - the improvements there are huge. So much so that it may make sense for AMD to switch to that stack on Windows in the future.