Axelera AI Challenges Nvidia With Europa Chip For High-Performance Edge Computing

In an interview with CRN, Alexis Crowell, Axelera AI’s CMO and general manager of Americas, says that the company decided to use last-generation PCIe Gen 4 connectivity for its upcoming Europa AI chip because it wants to ‘be able to plug into as many systems as we can.’

Axelera AI on Tuesday said its newly revealed Europa AI chip can outperform certain Nvidia products using less energy and at lower cost for high-performance edge computing use cases, including large language models and computer vision.

The Europa chip marks a new tier of edge AI performance above the Netherlands-based chip startup’s Metis lineup, providing 629 trillion operations per second (TOPS) in a 45-watt power envelope in a big step up from its predecessor, which can reach a maximum of 214 TOPS while using four to eight watts depending on the form factor.

Set for release in the first quarter of next year, Europa was announced a few months after Axelera AI launched a global partner program and promoted CMO Alexis Crowell to grow the company’s business in North America and South America, with a major U.S. emphasis.

In an interview with CRN last Friday, Crowell said Axelera AI will continue to sell and support its Metis products because company plans to offer Europa for a “very different” part of the edge computing market that requires significantly higher performance.

Europa’s targeted use cases include robotics, such as humanoids and automated guided vehicles, as well as computer vision for surveillance and analytics purposes where there are large deployments of cameras such as stadiums, according to Crowell.

The executive said Axelera AI also sees a play for Europa in “AI factories,” referring to data centers being used for advanced AI development. According to Crowell, the company has already seen AI factory proposal requests “that are specific for our technology.”

With these use cases in mind, the company is marketing Europa for workstations, edge servers, enterprise servers and rack-mount servers. By contrast, Axelera AI positions Metis for mini PCs, regular-sized PCs, workstations and small edge servers.

Why Europa Uses PCIe Gen 4, Not Gen 5

At launch, Europa will be available in PCIe form factors that can scale from one to four cards in a system using last-generation PCIe Gen 4 connectivity. The company plans to expand Europa to other form factors in the future.

Crowell said Axelera AI decided to use PCIe Gen 4 and not the latest PCIe Gen 5 connectivity, which offers twice the transfer speed of the previous generation, because the company wants to “be able to plug into as many systems as we can,” pointing to the large footprint of PCIe Gen 4 devices across the world.

“What we’re intentionally trying to do is make sure that it’s not just net new designs. It’s, you have some hardware, you have got all your [capital expenditures spent], but you’ve got some money that you can [use to] go upgrade. Everybody’s got a PCIe slot, so if I can plug into those, then it makes the sale, candidly, easier,” she said.

This tracks with the plug-and-play ethos that Axelera AI has touted in previous interviews, with its CEO and founder, Fabrizio Del Maffeo (pictured), telling CRN earlier this year: “We want to sell cards that integrators can plug inside their computer, install the software and just run or develop their algorithm.”

How Europa Competes Against Nvidia Products

In graphs provided by Axelera AI, the company claimed that Europa provides higher performance per dollar compared to two unnamed rival products by a magnitude of three to six times for Llama 3 models. It saw the highest advantage with a 70-billion-parameter Llama 3 model, particularly a four-chip system, and smaller but still impressive margins over the competing products with the 8B Llama 3 and 11B Llama 3.2 vision models.

The company also claimed that Europa provides roughly two to three times higher performance per watt on the same Llama 3 models, with the 70B version giving Axelera AI once again the highest advantage, this time with the one-chip system at the top.

In comparison to a third unnamed rival product, Axelera AI said that Europa can provide varying levels of performance advantages for frames per second across different versions of computer vision models, including ResNet-50 and various Yolo models.

While Axelera did not name the competing products in the graphs, Crowell told CRN that they are all recent Nvidia products used for edge computing purposes, including the L40 GPU and unspecified Jetson system-on-chip products.

“We stack up really nicely against L40s,” she said.

Europa will rely on the company’s same Voyager software development kit that is used for its Metis chips. Crowell emphasized the importance of software to Axelera, saying that half of its workforce of more than 200 people is focused on that area.

“So if companies want to deploy both [Metis] and [Europa] because they’ve got mixed use cases [and] they’ve got mixed solutions, they absolutely can. It’s one tool, and that, to me, is really important, because ease of use is the only way any of these partners are going to actually go adopt [Axelera’s chips],” she said.

How Europa Packs More Performance Than Metis

With Europa, Axelera AI has doubled the number of AI processing units to eight and quadrupled the maximum memory capacity to 64 GB per chip over Metis. The chip also features 200 GBps of DRAM bandwidth and 128 MB of L2 SRAM.

Another big difference with Europa is the chip’s ability to fully do pre- and post-processing thanks to the integration of 16 RISC-V vector processing cores. By contrast, systems using Metis chips must offload pre- and post-processing to the system’s CPU.

“We now can take a lot of that on ourselves directly in a way that provides a much better [total cost of ownership] for the customers,” Crowell said.

Europa also features an on-board video decoder, which can help optimize the performance of computer vision models running on the chip.

“We’re trying to make it even easier as customers are bringing more and more of their pipelines into AI and allow it to then just run more effectively, more cost effectively and with higher performance,” Crowell said.