5 ms·
Right, so a device with 5000+ CUDA Cores, 312 TFlops of Tensor Cores, hundreds of dedicated RT Cores, and 1+ Tb/s memory bandwidth to 8gb of dedicated VRAM, and
by cromwellian 6y ago
Right, so a device with 5000+ CUDA Cores, 312 TFlops of Tensor Cores, hundreds of dedicated RT Cores, and 1+ Tb/s memory bandwidth to 8gb of dedicated VRAM, and 54 billion transistors, is going to lose to a SoC? Not likely.
Apple's view of what "discrete" performance is, is probably 2-generation old cut-down mobile variants. For example, their top-of-the-line uber-expensive Mac Pro ships with Radeon Vega II architecture from 2017. Meanwhile, AMD RDNA2/Navi and NVidia Ampere are about to ship.
From my view, they're at least 2-3 generations behind in performance, and their focus on mobile means they're always going to be behind given thermals. NVidia and AMD are focused on maximum performance, and they assume gaming takes place plugged into a wall socket. And because they're focused on maximum performance, especially for triple-A title games and DCC, they're putting in features like DXR (DirectX Raytracing) with support for hardware accelerated ray-intersection. How long until I can run Minecraft with ray-tracered shaders at 2k or 4k on Apple's GPU? (https://www.youtube.com/watch?v=AdTxrggo8e8 https://www.youtube.com/watch?v=AdTxrggo8e8)
I mean, Apple's engineers on ARM are good and they did a fantastic job improving the PA Semi IP, but I don't think they're going to take what's essentially a PowerVR architecture which was never competitive with discrete, and leapfrog NVidia and AMD.
- m0xte 6y agoAverage gamer doesn’t need and cant afford that stuff. Some have just got used to having every bell and whistle. Average gamer is playing fortnite which works perfectly nicely on an A12Z. I tried it. Plugged my monitor into my iPad Pro and paired an Xbox controller with it and got a pleasant surprise. The key is consistent performance and that is something they can manage with full control over API and silicon end to end. They are already doing this well on cores that are two generations old. I expect to see a good compromise on performance and power which is what we need for general purpose computing. To be clear I have actually canned my gaming pc recently which was a ryzen 3700x and GTX 1660. I haven’t missed it, the big titles or the pain in the arse getting it stable and built to start with.
- ClumsyPilot 6y agoPain in getting it stable? What does it even mean, it was unbalabced and falling off your desk? What did you have, crashing games? It has been like 7 years since I had any stability issues on windows for gaming.
- m0xte 6y agoJust annoying little issues like the six months it wouldn’t power up first time. Turned out to be a GPU initialisation problem. But when you only have one type of each component and the only way of debugging is swapping parts until you find the problematic one, you have to live with it. For some people it never works and then you have to deal with several vendors you bought the parts from all of whom point at each other. And if you buy off the shelf then you really have no idea what crap you’re being sold unless you want to pay a very high premium. Every PC build takes a lot of risk on. Not one I’m willing to take any more. Commercial desktops are either junk or too expensive. I’d rather use a single sourced well integrated appliance. Edit: also the dubious nature of some parts is a worry as demonstrated here https://www.reddit.com/r/WatchPeopleDieInside/comments/g0420s/pc_build_does_not_go_as_planned/ https://www.reddit.com/r/WatchPeopleDieInside/comments/g0420...
- Veedrac 6y agoApple's GPUs have exceptional render performance. Consider how close they get to a GTX 1060 laptop already[1], and then consider that their upcoming chips can comfortably double GPU core counts and will be 50% stronger per core with even conservative generational improvements. They do less well in compute, but even there they do well, and a 16 core A14 GPU is likely to bat with a 1060. [1] https://images.anandtech.com/graphs/graph13661/103805.png https://images.anandtech.com/graphs/graph13661/103805.png (not a perfect benchmark because of CPU bottlenecking, but there are no good benchmarks to choose)
- fomine3 6y agoNot exceptional and impressing for comparing 14nm/7nm GPUs. (yes it's fanless.)
- Veedrac 6y agoA 1060 is 200mm² of pure GPU at ~100W, an A12X's GPU is ~26mm² (approximated) running at ~10W. Of course it's impressive. Even if that was somehow all down to the better node, their next generation SoCs will still use a newer node than NVIDIA's next generation GPUs.
- fomine3 6y agoYou comparing 1060's entire die size including NVEnc, memory controller, PCIe controller, and so on vs A12X's GPU core only die size. And I expect 1060 is optimized for perf but A12X is optimized for power(temp). There's no reason NVIDIA should use a process behind Apple forever because both uses TSMC (and Samsung) and process improves slowing down.
- Veedrac 6y agoCropping to the inner part still gives ~140mm², or comfortably over 5x the size of Apple's GPU. I don't really get your argument. Apple customers buying Apple Silicon Macs—which, again, will probably have a GPU over three times as fast as the A12X—aren't going to let hypotheticals detract from their powerful and power-efficient GPUs. ‘But NVIDIA didn't optimize for efficiency’ and ‘but NVIDIA hypothetically could have used a newer node than they did’ don't count for squat.