Nvidia unveiled high-performance computing (HPC) reference designs and new water-cooling know-how for its GPUs on the annual Computex tradeshow in Taipei, Taiwan.
The reference designs make use of Nvidia’s forthcoming Grace CPU and Grace Hopper Superchips, due subsequent yr. Grace is an Arm-based CPU – Nvidia’s first for the server market. Hopper is Nvidia’s subsequent era of GPU processors.
There are two superchips: the Grace Superchip, which mixes two Grace CPU dies linked with the chipmaker’s tremendous high-speed NVLink-C2C interconnect tech; and the Grace Hopper Superchip, which options one Grace CPU linked to 1 Hopper GPU, additionally linked on to the CPU by NVLink-C2C.
These are a part of Nvidia’s HGX line for giant HPC deployments, the place compute density is the order of the day. They arrive in 1U and 2U designs with two HGX Grace-Hopper nodes or 4 HGX Grace nodes in a single chassis. All 4 include BlueField-3 networking processors.
The 4 designs are:
- Nvidia HGX Grace Hopper methods for AI coaching, inference and HPC
- Nvidia HGX Grace methods for HPC and supercomputing with a CPU-only design
- Nvidia OVX methods for digital twins (digital copies of bodily objects) and collaboration workloads
- Nvidia CGX methods for cloud graphics and gaming
“These new reference designs will allow our ecosystem to quickly productize the servers which might be optimized for Nvidia-accelerated computing software program stacks, and [they] will be certified as part of our Nvidia-certified methods lineup,” stated Paresh Kharya, director of datacenter computing, in a convention name with journalists.
Nvidia already has six associate distributors lined as much as launch methods within the first half of subsequent yr: Asus, Foxconn, Gigabyte, QCT, Supermicro, and Wiwynn.
Liquid cooling for GPUs
On the cooling entrance, Nvidia is acknowledging what players have recognized for a while: Its GPUs run very popular, and water cooling is a viable different to air cooling.
Nvidia’s present air-cooled 3000 sequence GPU is as much as three inches thick. The precise board with the GPU on it’s only some millimeters thick, with the heatsink and fan making up the distinction and making them extraordinarily unwieldy and troublesome to insert in a decent house.
So Nvidia is taking a cue from hard-core players’ playbooks: It’s eradicating the heatsink and fan and changing it with a a lot slimmer water block, just like the water-cooling know-how used on CPUs. Water is piped in, cools the water block over the processor, after which is pumped away to be cooled. (See an instance of that right here.)
The result’s a liquid-cooled PCIe card for Nvidia’s present flagship server GPU, the Ampere A100, that will likely be accessible within the third quarter of this yr. Nvidia has an analogous plan for the A100’s just lately introduced successor, the Hopper-powered H100, due subsequent yr. Greater than a dozen server makers are anticipated to help the liquid-cooled A100 PCIe card later this yr, together with ASUS, Gigabyte, Inspur, and Supermicro.
The water block is far slimmer than the warmth sink, and that’s key. The warmth sinks are so giant they block the PCI Specific slot proper subsequent to it. With the water block, a second card will be positioned within the system.
“Liquid cooling has been round for the reason that mainframe period, and it is extensively used at the moment to chill the world’s quickest supercomputers. We sometimes help these deployments with our high-performance HDX platforms. Nevertheless, now even the mainstream enterprise knowledge facilities are choices for liquid cooling of their knowledge heart infrastructures,” stated Kharya.
Kharya claimed that liquid cooling will enable knowledge facilities to offer the identical degree of efficiency as knowledge facilities with air cooling on the identical playing cards whereas consuming as much as 30% much less energy, in response to the outcomes of latest assessments Nvidia carried out with Equinix.
Jetson for the sting
The corporate additionally introduced broader deployment of its low-power Jetson AGX Orin platform with methods from greater than two dozen associate distributors focused at edge and embedded functions. Jetson is a low-power edge design with AI processing in thoughts.
Jetson AGX Orin was introduced earlier this yr throughout Nvidia’s GTC present. It’s a developer package with a single-board laptop primarily based on an Arm CPU with 12 Cortex-A78AE cores and an Ampere A100 GPU.
Nvidia claims 6,000 clients are constructing Jetson platforms, and to handle the rising demand for them the corporate introduced 4 new reference designs of various reminiscence configurations, accessible later this yr.
Copyright © 2022 IDG Communications, Inc.
