Exactly one year ago, at GTC 2022, NVIDIA unveiled the Grace, its first data center CPU, and the Hopper, a next-generation high-performance computing GPU. It also used them to create two "Super chips," the Grace two-in-one CPU. Grace CPU+Hopper GPU in one.

A year passed.At the new GTC 2023 conference, Huang Yanxun presented a super CPU+GPU with his own hands, which was the first time we saw it in real life.

Unlike the glossy renderings, the real thing is a bit unpretentious.

The small PCB circuit board is arranged with two chips and a large number of power supply components, extremely compact, whichThe Grace CPU with the larger base board and smaller chip, and the Hopper GPU with the smaller base board and larger chip, are integrated and packaged with six HBMS of memory.

At the same time, Huang also releasedEffect of a server node equipped with a super CPU+GPU", but did not give further explanation, did not disclose more specifications, performance, just a mentionThe communication bandwidth between cpus and Gpus is up to 10 times that of traditional PCIe buses.

Huang also didn't say when exactly the product would ship.

At the same time, Huang also came up with a super CPU two-in-one, with the substrate plus chip taking up most of the PCB.

The packaged computing module was also unveiled for the first time, with dimensions of length and width8x5 inches (20.3 x 12.7 cm), two groups can be placed into the 1U air-cooled server rack.

Huang said,Grace delivers 30% better performance, 70% better energy efficiency, and twice the data center throughput than traditional x86 cpus.

Grace CPU is currently out of prototype and the partners are designing the system scheme, but there is no timetable for when it will be shipped.

According to previously announced data, the Hopper GPU uses a custom 4nm TCM-based architecture that includes 80 billion transistors, 18,432 CUDA cores, 576 Tensor cores and a 6,144-bit wide 80GB HBM3/HBM2e high bandwidth memory. It also supports PCIe 5.0 and fourth-generation NVLIink, which boasts four times the performance of the previous generation A100 and a maximum power consumption of 700W.

Grace CPU adopts dual-core design, a total of 144 Arm architecture cores, integrates 396MB cache, supports LPDDR5X ECC memory, also supports PCIe 5.0, internal bandwidth 3.2TB/s, inter-chip bandwidth 1TB/s, and power consumption 500W.