TensorWave is expanding its AI cloud with the AMD Helios rackscale solution, a 72-GPU rack built around AMD Instinct MI455X GPUs for frontier-scale inference, training, and fine-tuning. For operators and platform architects, the key point is integration: Helios packages compute, networking, and software into a single rack-scale building block intended to scale from rack to cluster.
The AMD Helios rackscale solution integrates AMD Instinct GPUs, 6th Gen AMD EPYC CPUs, AMD Pensando networking, and AMD ROCm software. The rack-scale configuration includes 18 ORW-aligned 4-GPU compute trays, totaling 72 MI455X GPUs, and is designed to be virtualized for consistent deployment and expansion at cluster scale.
Rack-scale performance and memory numbers
At rack scale, Helios is specified at up to 2.9 exaFLOPS peak 4-bit (OCP MXFP4) and 1.4 exaFLOPS peak 8-bit (OCP MXFP8), with 31 TB of HBM4 and 1.7 PB/s of memory bandwidth. Per MI455X GPU, AMD lists up to 40 PFLOPs peak 4-bit, 20 PFLOPs peak 8-bit, 432 GB of HBM4, and 23.3 TB/s of memory bandwidth.
Those memory and bandwidth figures are the part data center teams will likely fixate on. For large-model training and high-concurrency inference, memory capacity and fabric behavior can become the gating factors just as quickly as raw FLOPs, so a rack that’s explicitly sized around HBM4 and network integration is aimed at the real constraints operators hit in production.
Networking and software stack
On the fabric side, Helios uses UALink over Ethernet (UALoE) within the rack and Ultra Ethernet Consortium (UEC)-aligned Ethernet across racks. The rack also includes the AMD Pensando Vulcano 800 AI NIC for high-bandwidth, low-latency connectivity, with AMD calling out support for multi-trillion-parameter training and rapid hyperscale growth.
AMD also lists “defense-in-depth security,” including Confidential Computing, device-level attestation, and encrypted multi-GPU scaling. On the software side, AMD ROCm is described as providing unified, high-throughput inference and training across all 72 GPUs and beyond, with optimized support for PyTorch, TensorFlow, JAX, ONNX Runtime, vLLM, and SGLang.
“AMD Helios rackscale solution brings the best of AMD compute and networking technology together into a single rack-scale building block,” said Andrew Dieckmann, corporate vice president and general manager, Data Center GPU Business Unit at AMD.
TensorWave Chief Growth Officer and co-founder Jeff Tatarchuk said, “AMD Helios brings compute, memory, and networking together at rack scale, enabling our customers to build AI faster, more efficiently, and with better economics at scale.”
Eugene Cheah, CEO and co-founder of Featherless AI, said, “We chose TensorWave because they understand AMD infrastructure at a level most clouds simply don’t.”
Source: TensorWave
