During this year’s SIGGRAPH festivities, NVIDIA took the stage to present impressive new superchip, GH200 with HBM3e memory, a breakthrough solution that follows on from the Grace Hopper GH100 model we saw about a year ago, and which reaffirms NVIDIA’s position as one of the largest, most advanced and most important companies in the world of specialized high-performance computing in AI and data.

The GH200 is NVIDIA’s first superchip to feature HBM3e high-bandwidth memory. This type of memory offers a speed increase of up to 50% compared to the HBM3 standard and allows it to offer bandwidth up to 5 TB/s per chip and 10 TB/s per system. In comparison, the NVIDIA H100 accelerator in the SMX format with 80 GB HBM3 offers a bandwidth of up to 1.68 TB/s.

The difference is overwhelming, but we must not forget the total memory capacity either. When we have to work with large AI models or processes associated with them, the total amount of memory available can drastically affect performance. NVIDIA knows this and has confirmed that the GH200 can offer up to 282GB of HBM3e memory per system.

NVIDIA has also confirmed that this superchip will feature 72 high-performance Neoverse ARM processor cores (144 cores per system) and that it will be capable of delivering AI performance overall. 8 PETAFLOPS per system. Yes, it is simply impressive and as the green giant confirms you will be able to work with models 3.5x larger than the previous generation. It will be available from the second quarter of 2024.

Now let’s dive into talking about the new NVIDIA RTX Ada Lovelace, which was also presented at SIGGRAPH 2023. Overall, the green giant presented three models that are placed directly in the high-end and mid-rangeNVIDIA RTX 5000 with 32 GB of graphics memory, RTX 4500 with 24 GB of graphics memory and RTX 4000 with 20 GB of graphics memory.

The RTX 5000 is the most powerful of the three as it stands 12,800 shaders, 400 texture units, 100 3rd gen RT cores, 400 4th gen tensor cores, 256-bit bus, 32 GB GDDR6 18 GHz memory and has a performance at FP32 of 65.3 TFLOPs. The RTX 4500 is a mid-range model 7,680 shaders240 texture units, 60 3rd gen RT cores, 240 4th gen tensor cores, 192-bit bus, 24 GB GDDR6 18 GHz memory and has a performance of 39.6 TFLOPs at FP32.

The RTX 4000 is the most modest of the three as it comes with the configuration 6,144 shaders, it has 192 texture units, 48 ​​3rd gen RT cores, 192 4th gen tensor cores, a 160-bit bus, 20 GB GDDR6 18 GHz memory and FP32 performance of 26.7 TFLOPs. All of these graphics cards integrate ECC (memory error correction technology) and are priced in that order $4,000, $2,250 and $1,250.