During this year’s SIGGRAPH celebration NVIDIA has taken advantage of the stage to present an impressive new superchip, the GH200 with HBM3e memory, a groundbreaking solution that follows the lines of the Grace Hopper GH100 model that we saw about a year ago, and that reaffirms NVIDIA’s position as one of the largest, most cutting-edge and most important companies in the world of specialized high-performance computing in AI and data.
The GH200 is the first NVIDIA superchip to feature high-bandwidth HBM3e memory. This type of memory offers an increase of up to 50% in speed compared to the HBM3 standard, and allows it to offer a bandwidth of up to 5 TB/s per chip and 10 TB/s per system. For comparison, an NVIDIA H100 accelerator in SMX format with 80 GB of HBM3 offers up to 1.68 TB/s bandwidth.
The difference is overwhelming, but we must not forget about the total capacity of memory. When we have to work with large AI models, or with processes associated with it, the total amount of memory available can drastically affect performance. NVIDIA knows this, and that is why it has confirmed that the GH200 can offer up to 282 GB of HBM3e memory per system.
NVIDIA has also confirmed that this superchip will feature a high-performance 72 Neoverse ARM CPU cores (144 cores per system), and that in total it will be able to deliver AI power of 8 PETAFLOPs per system. Yes, it is simply impressive, and as confirmed by the green giant you will be able to work with models 3.5 times bigger than the previous generation. It will be available from the second quarter of 2024.
We now jump to talking about the new NVIDIA RTX Ada Lovelace that have also been presented at SIGGRAPH 2023. In total, the green giant has presented three models that are placed directly in the high-end and mid-rangethe NVIDIA 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 has 12,800 shaders, 400 texture units, 100 3rd generation RT cores, 400 4th generation tensor cores, 256-bit bus, 32GB of 18GHz GDDR6 memory and has a power at FP32 of 65.3 TFLOPs. For its part, the RTX 4500 is a mid-range model with 7,680 shaders240 texture units, 60 3rd generation RT cores, 240 4th generation tensor cores, 192-bit bus, 24GB of 18GHz GDDR6 memory and has a power at FP32 of 39.6 TFLOPs.
The RTX 4000 is the most modest of the three, as it comes configured with 6,144 shaders, it has 192 texture units, 48 3rd generation RT cores, 192 4th generation tensor cores, 160-bit bus, 20 GB of 18 GHz GDDR6 memory and a power at FP32 of 26.7 TFLOPs. All these graphics cards integrate ECC (memory error correction technology), and are respectively priced at $4,000, $2,250, and $1,250.
