ARM adds neural accelerators to GPUs

(newsroom.arm.com)

170 points | by dagmx 4 days ago

41 comments

  • armchairhacker 1 day ago
    I think ML has lots of potential in this area specifically.

    Imagine a game with bare-bones graphics and lighting, and a NN that converts it into something pretty. Indie developers can make AA-looking games and all game developers can devote more effort into design and logic. Artists will still be needed for art direction and possibly fine-tuning, although there will be less needed for each game (also less developers needed with AI agents and better tools).

    Related, ML also has potential for AI enemies (and allies). Lots of players still prefer multiplayer, in part because humans are more realistic enemies (but also because they want to beat real humans); but multiplayer games struggle because good netcode is nontrivial, servers are expensive, some players are obnoxious, and most games don’t have a consistent enough playerbase.

    • mdp2021 23 hours ago
      > Imagine a game with bare-bones graphics and lighting

      https://eu-images.contentstack.com/v3/assets/blt740a130ae3c5...

      # The Art Of Braid: Creating A Visual Identity For An Unusual Game

      https://www.gamedeveloper.com/design/the-art-of-braid-creati...

      • 15 hours ago
        • cyanydeez 21 hours ago
          I don't think you have a realistic view of how this will be used.

          First, porn.

          Second, artificial botting to make your game look active.

          Third, hire a art developer in india, VPN them to your AI tool, fire them when the game is done.

          You really should check your prescription rose colored glasses.

          • serf 4 hours ago
            >Second, artificial botting to make your game look active.

            MMOs have been using artificial players produced by the developers since at least the early EverQuest days.

            The choice space in an mmo isn't that great, it's trivial to make a realistic acting NPC that mimics player behaviors and hand-wave the poor language capability as the other player being a unable to understand your chosen language.

            NCSoft was involved in things like this in the early Lineage days and were fined for it. I would have a real hard time thinking this behavior is now uncommon given how low the fruit is.

            Deck an NPC-Player in the most expensive cash-shop goods and have it stand around in a social area doing emotes but otherwise silent just to make the other players jealous and apt to purchase goods -- self-generated whale-bait.

            • viraptor 15 hours ago
              > artificial botting to make your game look active.

              There's no reason to involve an NN in this one. We had convincing bots with varied behaviours for ages.

              • cyanydeez 3 hours ago
                Not ones that say gg and call you a f*g.
              • BobbyJo 20 hours ago
                So? Why would that stop us from doing cool things for games?
                • cyanydeez 16 hours ago
                  [flagged]
                  • mathiaspoint 16 hours ago
                    You actually sound insane. I think you need to take a break from politics.
                    • cyanydeez 3 hours ago
                      If thats insane, gotta assume youre an ostrich for the last decade.
            • N_Lens 1 day ago
              This article 2 links deep had better technical details -

              https://community.arm.com/arm-community-blogs/b/mobile-graph...

              Upscaling solution mainly targeted at mobile gaming, with an 'AI pipeline' for upscaling graphics (They claim 540p upscaled to 1080p at 4ms per frame). I'm a bit skeptical because this is a press release for chips that are in the works and claim to be releasing in DEC-26, and then on actual devices after that. So sounds more like a strategic/political move (Perhaps stock price related manoeuvring).

              Unreal Engine 5 plugin will allow previewing the upscaled effects using the though, which will be nice for game developers.

              • wmf 1 day ago
                It's a copy of DLSS/FSR4 which are pretty well understood by now. As for the schedule, Arm always announces IP ahead of time.
                • ksec 13 hours ago
                  Is DLSS really that mature by now? I thought only DLSS 4 was good enough and we should still have ways to improve on it.

                  And there seems to be a lot of hate towards DLSS from Gaming community.

                  • wmf 12 hours ago
                    DLSS 2.x is pretty good; I'd expect Arm NSS 1.0 to be similar to that.
              • bobajeff 4 days ago
                It sounds like this a geared towards games. However, I like the idea of exposing all of the ML features through Vulkan extensions rather than some proprietary API. Though I think exposing them through OpenCL extensions would work for me as well.
                • pjmlp 2 hours ago
                  Extension spaghetti is hardly any better when each vendor does its own way.
                • imbusy111 23 hours ago
                  I figured there is a need for generating a lot of samples and building a predictive model per game for best results. Documentation confirms:

                  > Most of these corner cases can be resolved by providing the model with enough training data without increase the complexity and cost of the technique. This also enables game developers to train the neural upscalers with their content, resulting in a completely customized solution fine-tuned for the gameplay, performance, or art direction needs of a particular title.

                  Source: https://developer.arm.com/documentation/111019/latest/

                  • ksec 13 hours ago
                    At this point, IMG / PowerVR isn't even used by MediaTek. Which means GPU on Mobile is just Apple, Qualcomm Adreno, ARM Mali. Still wish ARM had rebranded their Mali range.

                    Samsung Exynos uses AMD RDNA but I am not even sure if they are being used at all. Nvidia seems to have no interest in the market.

                    • ryao 13 hours ago
                      Nvidia has the Tegra line, but the market is not interested in it outside of game consoles.
                      • msh 11 hours ago
                        Or Qualcomm used their monopoly to keep nvidia out of phones.
                        • dagmx 1 hour ago
                          Qualcomm wouldn’t have to try to do that. Tegra was basically a side project for Nvidia that they barely care about till the switch came along.
                          • msh 1 hour ago
                            The launch platform for android tablets was tegra chips. They were also popular in automotive.
                          • TiredOfLife 10 hours ago
                            Before Switch basically every company made 1-2 Tegra products only to newer use Nvidia again. Tegra was late and bad.
                        • TiredOfLife 10 hours ago
                          Apple is basically PowerVR with serial numbers partially filed off
                        • flakiness 23 hours ago
                          hardware-wise, this seems like a NVIDIA TensorCore? via https://huggingface.co/Arm/neural-super-sampling/blob/main/2...

                          - https://github.com/KhronosGroup/Vulkan-Docs/blob/5d386163f25... Adding tensor ops to the shader kernel vocaborary (SPIR-V). Promising.

                          - https://github.com/KhronosGroup/Vulkan-Docs/blob/5d386163f25... Adding TenforFlow/NNAPI/-like graph API. Good luck.

                        • cubefox 22 hours ago
                          There are now at least three ways to accelerate machine learning models on consumer hardware:

                            - GPU compute units (used for LLMs)
                            - GPU "neural accelerators"/"tensor cores" etc (used for video game anti-aliasing and increasing resolution or frame rate)
                            - NPUs (not sure what they are actually used for)
                          
                          And of course models can also be run, without acceleration, on the CPU.
                          • colejohnson66 22 hours ago
                            An "NPU" is a matrix multiplier accelerator. It removes some general-purpose stuff that GPUs provide in favor of more "AI"-useful units, like support for values a byte or smaller (i.e., FP4, INT4, etc.).
                            • MobiusHorizons 1 hour ago
                              I think NPUs are often aimed at efficient matmul performance. Not all implementations are significantly faster than vector units in the CPU, but they use much lower power. Gpu acceleration is typically much faster than the CPU, but also higher power.
                              • cubefox 22 hours ago
                                All three of them accelerate matrix multiplications actually.
                                • almostgotcaught 20 hours ago
                                  any thing that computes matmul faster than by hand technically accelerates matmul - so what's your point?
                              • atq2119 20 hours ago
                                At least for desktop gaming, the tensor cores are in the GPU compute units (SM), same as for the big data center GPUs.

                                It seems ARM believe it makes sense to go a different route for mobile gaming.

                                • catgary 19 hours ago
                                  From what I can tell, NPUs are mostly being used by Microsoft to encourage vendor lock-in to the MicrosoftML/ONNX platform (similar to their DirectX playbook).
                                  • pjmlp 2 hours ago
                                    Extension spaghetti is hardly any better, just because it says the same API name on the tin.

                                    Google and Apple have been doing NPUs for a while now.

                                    • jonas21 18 hours ago
                                      They're used a lot on mobile. Apple uses their "neural engine" NPU to power their on-device ML stuff and Samsung does something similar in their Exynos processors. Apple also exposes the NPU to developers via CoreML.
                                  • Roark66 6 hours ago
                                    ARM adds... Since I saw the first arm based soc (rockchip rk3566) every so came with npu accelerator. Usually pretty small ones. 0.5 Tops (int8) etc.

                                    The novel thing seems to be that they will make it a part of the GPU? Really? Even my Samsung Galaxy S7 (quite few years old by now) supported Vulcan and run neural nets pretty well with Vulcan etc.

                                    Where is the novelty?

                                  • ltbarcly3 20 hours ago
                                    "Arm neural technology is an industry first, adding dedicated neural accelerators to Arm GPUs"

                                    HiSilicon Kirin 970 had an NPU in like 2017. I think almost every performance-oriented Arm chip released in the last 5 years has had some kind of NPU on it.

                                    I suspect they are using Arm here to mean "Arm-the-company-and-brand" not "Arm the architecture", which is both misleading and makes the claim completely meaningless.

                                    • atq2119 20 hours ago
                                      The marketing speak isn't exactly clear, but I believe the point is that this is like an NPU inside of the GPU instead of next to it as a separate device. That would indeed be new, and I can see how it'd be beneficial to integration with games.
                                      • ginko 13 hours ago
                                        “ARM” is the architecture. “Arm” is the company.
                                        • adrian_b 8 hours ago
                                          Not really, i.e. not any more.

                                          In all recent documents issued by the Arm company, "Arm" is used for the architecture, i.e. the architecture variants are named "Armv6", "Armv7", "Armv8", "Armv9".

                                      • westurner 13 hours ago
                                        How many TOPS/WHr?