All Eyes on NVIDIA

Aside from CEO Jensen Huang, the DGX GB200 NVL72 was the star of the GTC 2024 keynote. The rackscale system integrates 72 next-generation Blackwell GPUs connected by NVLink to form “1 Giant GPU.” Jensen’s description of the NVLink passive-copper “backplane” caused a brief panic among investors that believed it somehow replaced InfiniBand, which it does not. The NVL72 represents next-generation AI systems, but Nvidia also revealed new details of its deployed Hopper-generation clusters. Next-generation 800G (XDR) InfiniBand won’t reach customers until 2025, so early Blackwell systems will use 400G (NDR) InfiniBand instead.

Source: NVIDIA

Jensen said the Hopper-generation EOS supercomputer had just come online. This cluster uses 608 NDR switches with 64 ports each for a total of 38,912 switch ports. This system places the leaf switches in racks at the end of the row, so all InfiniBand links employ optical transceivers. We estimate the servers add 5,120 ports for a system total of 44,032 NDR ports. Because Nvidia uses what it calls “twin-port OSFP” 800G transceivers, each transceiver serves two NDR ports. Thus, we estimate the complete EOS system uses about 22,000 800G optical transceivers.

Blackwell-generation GPUs include 5th-generation NVLink, which doubles the interconnect bandwidth compared with Hopper. It does this by doubling the per-lane speed to 200Gbps, which results in 400Gbps of unidirectional bandwidth for each NVLink x2 port. Each Blackwell GPU includes 18 ports, which deliver 1.8TB/s (14.4Tbps) of aggregate bidirectional bandwidth. To connect 72 GPUs in the NVL72 rack, Nvidia developed the NVLink5 switch chip. The NVL72 rack includes nine NVLink switch trays with two ASICs each. Using 5,184 passive-copper (DAC) cables, the switches deliver all-to-all GPU connectivity within the rack.

Perhaps the biggest GTC 2024 disappointment for the networking ecosystem was the revelation that 800G InfiniBand is delayed until 2025. Despite the delay, the company disclosed the Quantum-X800 switch system and ConnectX-8 adapter (NIC). When available, these 800G InfiniBand products will double per-GPU bandwidth, as ConnectX-8 NICs will replace ConnectX-7 (400G) NICs one-for-one. They should also be first to handle optics with 200G lanes on the electrical (host) side, driving early demand for second-generation 200G/lambda DSPs.

LightCounting subscribers can access additional details, figures, and analysis in a full-length research note: www.lightcounting.com/login

Comments

Post a Comment

Popular posts from this blog

AMD Looks to Infinity for AI Interconnects

Image
With the formal launch of the MI300 GPU, AMD revealed new plans for scaling the multi-GPU interconnects vital to AI-training performance. The company's approach relies on a partner ecosystem, which stands in stark contrast with NVIDIA's end-to-end solutions. The plans revolve around AMD's proprietary Infinity Fabric and its underlying XGMI interconnect. Infinity Fabric Adopts Switching As with its prior generation, AMD uses XGMI to connect multiple MI300 GPUs in what it calls a hive. The hive shares a homogeneous memory space formed by the HBM attached to each GPU. In current designs, the GPUs connect directly using XGMI in a mesh or ring topology. Each MI300X GPU has up to seven Infinity Fabric links, each with 16 lanes. The 4th-gen Infinity Fabric supports up to 32Gbps per lane, yielding 128GB/s of bidirectional bandwidth per link. At the MI300 launch, Broadcom announced that its next-generation PCI Express (PCIe) switch chip will add support for XGMI. At last October...
Read more

NVIDIA Networks NVLink

Image
I attended several sessions at last week's Hot Chips, and NVIDIA's NVSwitch talk was a standout. Ashraf Eassa did a great job of covering the talk's contents in an NVIDIA blog, so I will focus on analysis here.  SuperPOD Bids Adieu to InfiniBand From a system-architecture perspective, the biggest change is extending NVLink beyond a single chassis. NVLink Network is a new protocol built on the NVLink4 link layer. It reuses 400G Ethernet cabling to enable passive-copper (DAC), active-copper (AEC), and optical links. The build its DGX H100 SuperPOD, NVIDIA designed a 1U switch system around a pair of NVSwitch chips. The pod ("scalable unit") includes a central rack with 18 NVLink Switch systems, connecting 32 DGX H100 nodes in a two-level fat-tree topology. This pod interconnect yields 460.8Tbps of bisectional bandwidth. NVLink Network replaces InfiniBand (IB) as the first level of interconnect in DGX SuperPODs. The A100 generation pod uses an IB HDR leaf/spine inte...
Read more

Ultra Ethernet Promises New RDMA Protocol

Image
This week saw the formal launch of the Ultra Ethernet Consortium (UEC), which aims to reinvent Ethernet fabrics for massive-scale AI and HPC deployments. An impressive list of founding members back this ambitious effort: hyperscalers Meta and Microsoft; chip vendors AMD, Broadcom, and Intel; OEMs Arista, Atos, and HPE; and Cisco, which straddles the chip and OEM camps. Absent this backing, we could easily write off this consortium as doomed to failure. Our skepticism is rooted not in the obvious need the UEC looks to serve but rather in the challenges of standardizing and implementing a full-stack approach. The effort plans to replace existing transport protocols as well as user-space APIs. Specifically, the Ultra Ethernet Transport (UET) protocol will be a new RDMA protocol to replace ROCE, and new APIs will replace the Verbs API from the InfiniBand heritage. UET will provide an alternative to RoCEv2 and Amazon’s SRD , both of which are deployed in hyperscale data centers. (Source: Ul...
Read more