PCIe

Now the system start-up is expanding its portfolio with a product that replaces the Ethernet switches with optical ones. “You can compose [compute] nodes and drive them using our software,” says Bill Koss, CEO of Drut.

Only Google has demonstrated the know-how to make such a large-scale flexible computing architecture using optical switching.

Drut was founded in 2018 and has raised several funding rounds since 2021.

Jitender Miglani, founder and president of Drut, previously worked at MEMS-based optical switch maker, Calient Technologies.

Drut’s goal was to build on its optical switching expertise and add the components needed to make a flexible, disaggregated computing architecture. “The aim was building the ecosystem around optical switches,” says Miglani.

The company spent its first two years porting the PCIe protocol onto an FPGA for a prototype interface card. Drut showcased its prototype product alongside a third-party optical switch as part of a SuperMicro server rack at the Supercomputing show in late 2022.

Drut has spent 2023 developing its next-generation architecture to support clusters of up to 4,096 endpoints. These can be accelerators like graphics processing units (GPUs), FPGAs, data processing units (DPUs), or storage using the NVM Express (nonvolatile memory express).

The architecture, dubbed DynamicXcelerator, supports PCIe over optics to link processors (CPUs and GPUs) and RDMA (Remote Direct Memory Access) over optics for data communications between the GPUs and between the CPUs.

The result is the DynamicXcelerator system, a large-scale reconfigurable computing for intensive AI model training and high-performance computing workloads.

The core of the DynamicXcelerator architecture is a photonic fabric based on optical switches. This explains why Drut uses PCIe and RDMA protocols over optics.

Optical switches brings size and flexibility and by relaying optical signals, their ports are data-rate independent.

Another benefit of optical switching is power savings. Drut says an optical switch consumes 150W whereas an equivalent-sized packet switch consumes 1,700W. On average, an Infiniband or Ethernet packet switch draws 750W when used with passive cables. Using active cables, the switch’s maximum power rises to 1,700W. “[In contrast], a 32-64-128-144 port all-optical switch draws 65-150W,” says Koss.

Drut also uses two hardware platforms. One is the PCIe Resource Unit, dubbed the PRU-2000, which hosts eight accelerator chips such as GPUs. Unlike Nvidia’s DGX platform, which uses Nvidia GPUs such as the Hopper, or Google, which uses its TPU5 tensor processor unit (TPU), Drut’s PRU-2000 is an open architecture and can use GPUs from Nvidia, AMD, Intel, and others. The second class of platform is the compute node or server, which hosts the CPUs.

DynamicXcelerator’s third principal component are the FIC 2500 interface cards.

The iFIC 2500 card is similar to Drut’s current product’s iFIC 1000, which features an FPGA and four QSFP28s. However, the iFIC 2500 supports the PCIe 5.0 generation bus and the Compute Express Link (CXL) protocols. The two other FIC cards are the tFIC 2500 and rFIC 2500.

“The iFIC and tFIC are the same card, but different software images,” says Koss. “The iFIC fits into a compute node or server while the tFIC fits into our Photonic Resource Unit (PRU) unit, which holds GPUs, FPGAs, DPUs, NVMe, and the like.”

The rFIC provides RDMA over photonics for GPU-to-GPU memory sharing. The rFIC card for CPU-to-CPU memory transfers is due later in 2024.

Miglani explains that PCIe is used to connect the GPUs and CPUs, but for GPU-to-GPU communication, RDMA is used since even PCIe over photonics has limitations.

Certain applications will use hundreds and even thousands of accelerators, so a PCIe lane count is one limitation, distance is another; a 5ns delay is added for each metre of fibre. “There is a window where the PCIe specification starts to fall off,” says Miglani.

The final component is DynamicXcelerator’s software. There are two software systems: the Drut fabric manager (DFM), which controls the system’s hardware configuration and traffic flows, and the Drut software platform (DSP) that interfaces applications onto the architecture.

Drut knew it would need to upgrade the iFIC 1000 card. DynamicXcelerator uses PCIe 5.0, each lane being 32 gigabit-per-second (Gbps). Since 16 lanes are used, that equates to 512 gigabits of bandwidth.

“That’s a lot of bandwidth, way more that you can crank out with four 100-gigabit pluggables,” says Koss, who revealed co-packaged optics will replace pluggable modules for the iFIC 2500 and tFIC 2500 cards.

The card for the iFIC and tFIC will use two co-packaged optical engines, each 8×100 gigabits. The total bandwidth of 1.6 terabits – 16×100-gigabit channels – is a fourfold increase over the iFIC 1000.

The system’s networking can be viewed as a combination of circuit switching and packet switching.

The photonic fabric, implemented as a 3D torus (see diagram), supports circuit switching. Using a 3D torus, three hops at most are needed to link any two of the system’s endpoints.

One characteristic of machine learning training, such as large language models, is that traffic patterns are predictable. This suits an architecture that can set the resources and the connectivity for a task’s duration.

Packet switching is not performed using Infiniband. Nor is a traditional spine-leaf Ethernet switch architecture used. The DynamicXcelerator does uses Ethernet but in the form of a small, distributed switching layer supported in each interface card’s FPGA.

The smallest-sized DynamicXcelerator would use two racks of stacked PRU-2000s (see diagram). Further racks would be added to expand the system.

“The idea is that you can take a very large construct of things and create virtual PODs,” says Koss. “All of a sudden, you have flexible and fluid resources.”

Koss says a system can scale to 16,384 units by combining four clusters, each of 4,096 accelerators. “Each one can be designated as an ‘availability zone’, with users able to call resources in the different zones,” he says.

Customers might use such a configuration to segment users, run different AI models, or for security reasons. “It [a 16,384 unit system] would be huge and most likely something that only a service provider would do or maybe a government agency,” says Koss.

Drut claims the architecture costs 30 per cent less than conventional systems, while operational cost-savings are 40 per cent.

The numbers need explaining, says Koss, given the many factors and choices possible.

The bill of materials of a 16, 32, 64 or 128-GPU design has a 10-30 per cent saving solely from the interconnect.

“The bigger the fabric, the better we scale in price as solutions using tiered leaf-spine-core packet switches involving Ethernet-Infiniband-PCIe are all built around the serdes of the switch chip in the box,” says Koss. “We have a direct-connect fabric with a very high radix, which allows us to build the fabric without stacked tiers like legacy point-to-point networks.”

There are also the power savings, as mentioned. Less power means less heat and hence less cooling.

“We can also change the physical wires in the network,” says Koss, something that can’t be done with leaf-spine-core networks, unless data centre staff change the cabling.

“By grouping resources around a workload, utilisation and performance are much better,” says Koss. “Apps run faster, infrastructure is grouped around workloads, giving users the power to do more with less.”

The system’s evolution is another consideration. A user can upgrade resources because of server disaggregation and the ability to add and remove resources from active machines.

“Imagine that you bought the DynamicXcelerator in 2024. Maybe it was a small sized, four-to-six rack system of GPUs, NVMe, etc,” says Koss. If, in mid-2026, Nvidia releases a new GPU, the user can take several PRU-2000s offline and replace the existing GPUs with the new ones.

“Also if you are an Nvidia shop but want to use the new Mi300 from AMD, no problem,” says Koss. “You can mix GPU vendors with the DynamicXcelerator.” This is different from today’s experience, where what is built is wasteful, expensive, complex, and certainly not climate-conscious, says Koss.

Drut has 31 employees, 27 of which are engineers. “We are going on a hiring binge and likely will at least double the company in 2024,” says Koss. “We are hiring in engineering, sales, marketing, and operations.”

Proof-of-concept DynamicXcelerator hardware will be available in the first half of 2024, with general availability then following.