You can’t directly compare teraflops for GPU performance between different architectures unless you really only care about a single precision throughput, which is not a good metric. You do actually need real world graphics benchmarks to compare GPUs.
You also can’t compare NPU TOPs without knowing the baseline data type. Apple for the M3 uses FP16 whereas Intel uses INT8. You have to double the Apple number to get the raw data throughput (ignoring any other efficiencies for operations in different types).
It’s ~36 vs 48. So closer to 33% more for 100% more power use (impossible to measure just the NPU use though). The more comparable SOC for power use would be the M3 Pro
Gaming is essentially all done with FP32...so it's by far the best figure of merit (excluding issues with say RDNA3 dual issue which is rarely achieved in practice).
You will see gaming benchmarks come out soon, and Lunar Lake will be about 50% faster than the M3. (A secondary issue of course is how few latest gen games run on macOS....)
True, that's FP16, but it's not clear if M3's Neural Engine even supports INT8.
I'm sure M4 will make this much more competitive, but right now, Lunar Lake is overall a much more balanced architecture that most people would prefer, ceteris paribus....
Gaming is absolutely not all done with fp32. A lot of games actually target half precision , which is where most PowerVR based GPUs pull ahead. The majority of shaders and buffers are better suited for half.
It also ignores things like occupancy and memory throughput, among many other aspects of a GPU.
I think a 50% delta for GPU is very wishful thinking given even Intel are only claiming a 33% uplift versus meteor lake, which itself was behind the M3 line when compared against similar TDP.
Regarding the NPU, the M3 does support INT8. It’s just that between the M3 and M4 release, the rest of the industry started coalescing on INT8, hence the change in base type.
I expect the same will happen again now that NVIDIA are touting INT4 as their performance standard for marketing.
Intel Arc runs FP16 at 2:1 compared to FP32, and Battlemage on Lunar Lake is the same, and XMX FP16 is actually at 8:1. I don't think M3's GPU has a better ratio.
Of course there are many other aspects, but given it's Intel's latest architecture, which has improved efficiency tremendously (see https://cdrdv2-public.intel.com/824434/2024_Intel_Tech%20Tou... ) it's pretty unlikely M3 has any fundamental advantage.
Do you have any reference showing Neural Engine in M3 supports INT8 (and at 2x FP16? Just curious.)
I’m not saying the M3 has a fundamental advantage. I’m saying that it’s unlikely to be as high a difference as is being stated in real world use. I don’t think a SOC at half the power budget is going to be magically more powerful.
Possibly Battlemage running at 100% will use more power than M3's GPU running at 100%...it will take some detailed testing to track that+ Lunar Lake can be set at different TDP's (plus performance settings, on battery vs connected to power). Not to mention different "100%" GPU workloads.
At the end of the day though, users vastly prefer a more powerful built in GPU for the occasional game session...Intel is willing to pay for the transistors, and Apple reserves them for the M3 Pro instead.
You also can’t compare NPU TOPs without knowing the baseline data type. Apple for the M3 uses FP16 whereas Intel uses INT8. You have to double the Apple number to get the raw data throughput (ignoring any other efficiencies for operations in different types).
It’s ~36 vs 48. So closer to 33% more for 100% more power use (impossible to measure just the NPU use though). The more comparable SOC for power use would be the M3 Pro