Vendors & EquipmentAmdAI RanVirtual RanForward Error Correction
AMD Telco Chief Warns of GPU 'Backlash' in AI-RAN
AMD's telco lead predicts a "big, big backlash" against data center GPUs in the RAN, betting instead on 84-core CPUs running FEC in software, with Samsung as its likeliest route to market.
Vendors & EquipmentWhy it matters
- AMD telco lead Piotr Weglicki predicts a "big, big backlash" against power-hungry data center GPUs in the RAN
- AMD's latest CPU advertises 84 cores, exceeding the 72 in Intel's most advanced Granite Rapids product
- Nvidia helped set up the AI-RAN Alliance almost three years ago; AMD has shown no GPU-RAN response
- Samsung, the world's fifth-largest RAN vendor, is promoting AMD's 84-core accelerator-free platform
- AMD acquired Xilinx in early 2022; its FPGA-based T2 Telco Accelerator Card failed to attract sufficient interest
The story
AMD's telco chief expects a "big, big backlash" against the use of power-hungry data center GPUs in the radio access network, effectively ruling the company out of the GPU-RAN race that Nvidia has championed for nearly three years. Piotr Weglicki, the telco lead for AMD's server business, delivered the verdict at a recent AI-RAN symposium organized by Omdia, a Light Reading sister company.
The stance is notable because AMD, headquartered in Santa Clara, looks like the natural second source for almost everything telcos might want. It is the only major alternative to Intel for x86-based CPUs, now a pillar of virtual RAN. Excluding Chinese vendors, it is also the only serious rival to Nvidia in the data center GPU market — chips Nokia is using to develop its future 5G and 6G products, backed by Nvidia's $1 billion stake in the Finnish vendor.
Why not GPUs in the RAN?
Weglicki's objection is energy. Data center GPUs are simply too power hungry for mobile network sites. His comments would not entirely rule out AMD building a lighter-weight GPU for the RAN's energy constraints. But almost three years have passed since Nvidia helped found the AI-RAN Alliance and began promoting its GPU-based RAN vision, and AMD has shown no response. That response now looks less likely than ever.
What is AMD selling instead?
Like Intel, AMD prefers a virtual RAN built on general-purpose CPUs. The argument is partly about silicon economics: custom RAN chips, developed at considerable expense, cannot be sold outside the telecom sector, which weakens the return on investment and may force more frequent redesigns to accommodate new features.
"What it gives telcos, actually, is access to data-center economics with much bigger R&D budgets with much faster development cycles," said Weglicki. He added that general-purpose compute is absorbing the RAN acceleration workload: "With each of the generations, we bring more CPU performance within the same power envelope."
The execution details diverge sharply from Intel's. Intel's latest products pair the CPU with an integrated hardware accelerator for forward error correction (FEC), an energy-hungry Layer 1 function. AMD executives have criticized that integration as "vendor lock-in" — an impediment to porting virtual RAN software between platforms.
AMD's first move was a separate FEC card, the T2 Telco Accelerator, built on FPGAs from Xilinx, which AMD acquired in early 2022. Telco executives complained those FPGAs were too power hungry, and AMD failed to attract sufficient interest.
Its latest approach drops the accelerator entirely. By raising core counts — a model advertised earlier this year reaches 84 cores, against 72 in Intel's most advanced Granite Rapids product — AMD claims the CPU can now run FEC entirely in software.
"Forward error correction increasingly could be done on software on general-purpose cores, and the consequences of that are huge," said Weglicki. Operators gain flexibility, he argued: they can decide how much compute goes to L1, L2 and L3 functions, how much to AI-RAN models and how much to edge applications.
Who is buying?
Commercial traction remains thin. Geopolitics rules out Huawei and ZTE as customers. Nokia is committed to its Nvidia GPU strategy, and Ericsson remains heavily invested in its own custom silicon; its only commercial virtual RAN partner is Intel.
Samsung, the world's fifth-largest RAN vendor, is the most promising route to market. A longtime virtual RAN advocate that previously used Intel CPUs, it has recently promoted AMD as an option. "We've been doing quite a bit of work with AMD around their platform that does not have acceleration but can support meaningful multicell capacity," said Alok Shah, Samsung Networks' vice president of network strategy. "It's an 84-core platform and so that's a path that operators can certainly choose to go down."
A rollout beyond Samsung's footprint would require other vendors to develop software-only FEC and resolve power concerns. No operator will adopt the technology if it cannot match custom silicon and mainstream virtual RAN on consumption. But as a processor alternative in a market Intel has long dominated, the 84-core approach has a plausible opening — and on GPU RAN, the industry still faces a choice of Nvidia or nothing.
Also reported
Source: Light Reading
More from Rebecca Stone
Show full bio
Correspondent covering technology and product news in Carrier business, 5G and 6G, fiber and broadband at Telecom Gazette.
539 articles