NVIDIA RTX 3500 Ada Generation Laptop GPU vs Apple M3 Pro 14-Core GPU vs Apple M3 10-Core GPU

NVIDIA RTX 3500 Ada Generation Laptop GPU

The Nvidia RTX 3500 Ada Generation is a higher-end professional graphics card for use in laptops that sports 5,120 CUDA cores and 12 GB of ECC GDDR6 VRAM. Brought into existence in 2023, this graphics adapter leverages TSMC's 5 nm process and Nvidia's Ada Lovelace architecture to achieve higher-than-average performance combined with moderate power consumption. The Nvidia-recommended TGP range for the card is very wide at 60 W to 140 W leading to bizarre performance differences between different systems powered by what is supposed to be the same product.

Hardware-wise, the RTX 3500 is a cut-down GeForce RTX 4070 Desktop, as far as we can tell. Consequently, both make use of the AD104 chip and have little difficulty running triple-A games at QHD 1440p.

Quadro series graphics cards ship with a different BIOS and drivers than GeForce cards and are targeted at professional users rather than gaming. Commercial product design, large-scale calculations, simulation, data mining, 24 x 7 operation, certified drivers - if any of this sounds familiar, then a Quadro card will make you happy.

Architecture and Features

Ada Lovelace brings a range of improvements over older graphics cards utilizing the outgoing Ampere architecture. It's not just a better manufacturing process and a higher number of CUDA cores that we have here (up to 16,384 versus 10,752); under-the-hood refinements are plentiful, including an immensely larger L2 cache, an optimized ray tracing routine (a different wat to determine what is transparent and what isn't is used), and other changes. Naturally, these graphics cards can both encode and decode some of the most widely used video codecs, AVC, HEVC and AV1 included; they also support a host of Nvidia technologies, including Optimus and DLSS 3, and they can certainly be used for various AI tasks.

The RTX 3500 Ada features 40 RT cores of the 3rd generation, 160 Tensor cores of the 4th generation and 5,120 CUDA cores. Multiply those numbers by 1.15 and what you get looks exactly like a desktop RTX 4070: 46, 184 and 5,888, respectively. Elsewhere, the graphics card comes with 12 GB of 192-bit wide ECC GDDR6 memory for a very healthy throughput of ~432 GB/s. Error correction can be turned off if desired. The fact that error correction is present here proves that the RTX 3500 Ada is indeed targeted at professional users.

Just like Ampere-based cards, the RTX 3500 makes use of the PCI-Express 4 protocol. 8K SUHD monitors are supported, however, DP 1.4a video outputs may prove to be a bottleneck down the line.

Performance

With its power target set to ~105 W, the RTX graphics card delivers a Blender 3.3 Classroom CUDA score of 33 seconds which is a pretty good result. Most 2024 games are playable at 2160p on High with this GPU.

Nvidia's marketing materials mention "up to 23 TFLOPS" of performance, a 15% improvement over the 20 TFLOPS that the RTX 3000 Ada Generation is supposedly good for.

Your mileage may vary depending on how competent the cooling solution of your laptop is and how high the TGP power target of the RTX 3500 is. One other thing worth mentioning is that enabling error correction appears to reduce the amount of video memory that is available to applications and games by up to a gigabyte.

Power consumption

Nvidia no longer divides its laptop graphics cards into Max-Q and non-max-Q models. Instead, laptop makers are free to set the TGP according to their needs, and the range can sometimes be shockingly wide. This is the case for the RTX 3500, as the lowest value recommended for it sits at just 60 W while the highest is more than two times higher at 140 W (this most likely includes Dynamic Boost). The slowest system built around an RTX 3500 Ada can easily be 60% slower than the fastest one. This is the kind of delta that we've been seeing on consumer-grade laptops featuring the latest GeForce RTX cards.

Last but not the least, the improved 5 nm process (TSMC 4N) the RTX 3500 is built with makes for very decent energy efficiency, as of mid 2023.

Apple M3 Pro 14-Core GPU

The Apple M3 Pro 14-Core GPU is a self-designed graphics card in the Apple M3 Pro (with 11 CPU cores) and offers fourteen of the eighteen cores available on the chip. The graphics card uses a new architecture and now offers dynamic caching, mesh shading and ray tracing acceleration.

The GPU can access up to 24 GB LPDDR5-6400 unified memory via a 192-bit memory bus (max 150 GBit/s). Thanks to dynamic caching, the GPU only reserves as much memory as is required.

The entire chip is manufactured in the current 3nm process (probably N3B) at TSMC.

Apple M3 10-Core GPU

The 10-core Apple M3 GPU is an integrated graphics adapter designed by Apple that features 10 cores. This iGPU is built into the Apple M3 SoC and it uses the unified memory architecture (up to 24 GB LPDDR5-6400 with 100 GB/s bandwidth). Compared to M2 series GPUs, this new graphics adapter uses a new architecture with support for mesh shading and ray tracing. Dynamic caching is another new feature worth mentioning - it is coming to optimize shared memory usage.

In our testing, the M3 GPU (10 cores) mostly was 5% to 20% faster than the 10-core M2 GPU depending on the task.

Just like the rest of the Apple M3 chip, the graphics adapter is manufactured on a 3 nm TSMC process (possibly N3B). Our MBP 14 review revealed sustained power consumption figures of around 15 W for the iGPU.