Skip to main content

Rack-Scale AI Era: NVL72 vs Helios vs Groq 3 LPX vs Trn3 UltraServer — Four Major Solutions Compared

· 7 min read
Industry Research Team

2026 AI compute enters the "rack-scale" era. Single-chip comparisons have receded, and full-rack solutions have become the main battleground. This article provides an in-depth comparison of the five major rack-scale solutions: NVIDIA Rubin NVL72/NVL576, AMD Helios, Groq 3 LPX, AWS Trn3 UltraServer, and Google TPU 8t pod.

Why the Rack-Scale Era?

Limitations of Single-Chip Comparisons

Single-Chip Metric2018 (V100)2024 (H100)2026 (Rubin R200)2028 (Est.)
Compute125 TFLOPS989 TFLOPS25 PFLOPS80 PFLOPS
Memory32 GB80 GB288 GB1 TB
TDP300 W700 W1,800 W3,000 W

Single-chip TDP is about to exceed 3,000W — physical cooling, power delivery, and interconnect have all reached their limits.

Advantages of Rack-Scale

  • Unified cooling: Full-rack liquid cooling, high thermal efficiency
  • Unified power: Centralized power delivery, optimized energy efficiency
  • Unified interconnect: NVLink 6 / UALoF / GroqSync / NeuronLink
  • Unified management: Single-system software stack
  • Unified procurement: Single SKU purchase, simplified operations

Five Major Rack-Scale Solutions

1. NVIDIA Rubin NVL72 / NVL576

ItemRubin NVL72Rubin NVL576
GPU Count72576
CPU Count36288
Total HBM20.7 TB HBM4165 TB HBM4
Memory Bandwidth1.6 PB/s12.7 PB/s
NVLink Aggregate252 TB/s2,016 TB/s
FP4 Sparse Compute3.6 EFLOPS28.8 EFLOPS
FP8 Sparse Compute1.8 EFLOPS14.4 EFLOPS
DC NetworkConnectX-9 1152 portsConnectX-9 1152 ports
TDP (Rack)~130 kW~1 MW
CoolingLiquidLiquid
Suitability100B+ model training1T+ giant models
Price~$3-5M~$25-40M
Release2026 H22026 H2+

Rubin NVL576 = 28.8 EFLOPS FP4 = 1.5 ExaFLOPS FP8 = World's most powerful AI super node

2. AMD Helios Rack

ItemHelios
GPU Count72 MI400 GPUs
CPU Count36 EPYC Venice CPUs
Total HBM31.1 TB HBM4
Memory Bandwidth1.4 PB/s
Scale-up InterconnectUALoF 260 TB/s (open standard)
Scale-out NetworkPensando Vulcano 800G
FP4 Dense Compute2.88 EFLOPS
FP8 Dense Compute1.44 EFLOPS
TDP (Rack)~80 kW
CoolingLiquid
Suitability700B+ model training
Price~$2-3M
Release2026

Helios surpasses NVIDIA Rubin NVL72 in dense compute (2.88 vs 1.8 EF FP8 dense)

3. NVIDIA Groq 3 LPX Rack (Inference-Specialized)

ItemGroq 3 LPX
LPU Count256 Groq 3 LPUs
CPU CountNone (standalone)
On-chip SRAM128 GB aggregate
SRAM Bandwidth40 PB/s (SRAM, not HBM)
InterconnectGroqSync + NVLink-Network 640 TB/s
FP8 Compute~640 PFLOPS
INT8 Compute~640,000 TOPS
TDP (Rack)~80 kW
TTFT (Time to First Token)<20ms
TPOT<5ms
CoolingLiquid
SuitabilityUltra-low latency inference (Agentic AI)
Price~$8-10M
Release2026 H2

Groq 3 LPX is currently the only rack-scale LPU system designed specifically for Agentic AI

4. AWS Trn3 UltraServer

ItemTrn3 UltraServer
Chip Count144 Trainium 3 chips
Total HBM~20.7 TB
NeuronLink-v4Fully interconnected, >10 TB/s bidirectional
FP8 Dense Compute52 PFLOPS
BF16 Dense Compute~187 PFLOPS
TDP (Rack)~100 kW
CoolingLiquid
Suitability400B+ model training
Price (Est.)~$3-5M
Release2025-12 GA

Trn3 UltraServer = Best value large-scale training solution (2-3× performance per dollar vs NVIDIA)

5. Google TPU 8t pod

ItemTPU 8t pod
Chip Count9,216 TPU 8t chips
Total HBM~2 PB HBM
Interconnect3D Torus
Integrated CPUArm Axion (64 cores per node)
BF16 Dense Compute~32 PFLOPS × 9,216 = 295 EFLOPS
FP8 Dense Compute~590 EFLOPS
CoolingLiquid
SuitabilityGemini 3/4 training
PriceGoogle Cloud only
Release2026-04-22

TPU 8t pod = World's largest AI training cluster (9,216 chips × 9 PFLOPS ≈ 83 EFLOPS FP4 dense)

Five-Solution Horizontal Comparison

MetricNVIDIA NVL72AMD HeliosGroq 3 LPXTrn3 UltraServerTPU 8t pod
Form FactorTraining rackTraining rackInference rackTraining rackTraining pod
Chip Count72 GPU72 GPU256 LPU144 chip9,216 chip
Total Memory20.7 TB HBM31.1 TB HBM128 GB SRAM20.7 TB HBM~2 PB HBM
InterconnectNVLink 6 252 TB/sUALoF 260 TB/sGroqSync 640 TB/sNeuronLink-v43D Torus
FP4 Compute3.6 EF (sparse)2.88 EF (dense)
FP8 Compute1.8 EF (sparse)1.44 EF (dense)640 PF52 PF (dense)590 EF (dense)
TDP130 kW80 kW80 kW100 kW~10 MW (pod)
TTFT~100ms~100ms<20ms~100ms~100ms
EcosystemCUDA 13ROCm 8Groq SDKNeuron 3JAX 0.5+
Price$3-5M$2-3M$8-10M$3-5MInternal use
CustomersAll clouds + customersCustomers + cloudCustomers + cloudAWS CloudGoogle Cloud
Standardization❌ NVLink proprietary✅ UALoF open❌ GroqSync❌ NeuronLink❌ Torus
Release2026 H220262026 H22025-12 GA2026-04

Selection Recommendations

Large-Scale Training

ScenarioRecommended SolutionReason
100B-700B model trainingNVIDIA Rubin NVL72Single rack fits 100B, strongest FP4 compute
700B-1T model trainingNVIDIA Rubin NVL576 or AMD HeliosMulti-rack interconnect
1T+ giant model trainingNVIDIA NVL576 (8 units)28.8 EFLOPS × 8 = 230 EFLOPS
Hyperscale (Gemini class)Google TPU 8t pod (9,216 chip)Google Cloud only
AWS internal trainingTrn3 UltraServerBest value
Open ecosystem preferenceAMD HeliosUALoF open interconnect

Ultra-Low Latency Inference

ScenarioRecommended SolutionReason
Agentic AI (1000+ calls/sec)Groq 3 LPXTTFT <20ms, only choice
Real-time Code Gen (Copilot)Groq 3 LPXSub-100ms response
Trillion-parameter inferenceNVIDIA Rubin R200 + Groq 3 LPX coordinatedGPU training + LPU inference
70B single-model inferenceTPU 8i (288GB HBM)Single card fits FP16 70B
Multi-modal real-time inferenceTPU 8i (air-cooled)Flexible cooling

Cost-Sensitive Training

ScenarioRecommended SolutionReason
100B parameter trainingAWS Trn3 UltraServer2-3× performance per dollar vs NVIDIA
Hyperscale (non-Gemini)AWS Trn3 UltraServer × N$3-5M/rack
70B fine-tuningAMD Helios single rackValue + open ecosystem
100B+ parameter trainingTrn3 UltraServer × 3144 × 3 = 432 chips

1. Per-Rack Compute Continues Growing

YearPer-Rack ComputeMainstream Solution
2024~0.2 EFLOPS FP8GB200 NVL72
20261.8-3.6 EFLOPS FP8Rubin NVL72 / Helios
20288-15 EFLOPS FP8Rubin Ultra NVL72 / MI500
203030-50 EFLOPS FP8Feynman era

2. Multi-Rack Interconnect Standards Competition

StandardVendorStatus
NVLink NetworkNVIDIAProprietary, 2026 primary
UALoFAMD/Broadcom/IntelOpen, 2026 Helios debut
UALinkAllianceUALoF evolution
NeuronLinkAWSPrivate
GroqSyncGroq (NVIDIA)Private, ultra-low latency

3. Software Ecosystem Layering

  • Training frameworks: PyTorch + JAX + Megatron
  • Inference engines: vLLM + TensorRT-LLM + SGLang
  • Resource scheduling: Slurm + Kubernetes + Ray
  • Multi-rack management: NVIDIA Base Command / AMD ROCm RunTime

Detailed Product Pages

Summary

The primary battleground for AI compute in 2026 is rack-scale solutions:

  1. NVIDIA Rubin NVL72/NVL576 — Strongest training, FP4 3.6/28.8 EFLOPS
  2. AMD Helios — Open ecosystem, leading dense compute
  3. Groq 3 LPX — Ultra-low latency inference, TTFT <20ms
  4. AWS Trn3 UltraServer — Best value, 2-3× per dollar
  5. Google TPU 8t pod — Hyperscale, 9,216 chip cluster

There is no best, only the most suitable. Selection should consider:

  • Model size (100B / 700B / 1T+)
  • Training vs inference
  • Latency requirements (normal vs Agentic)
  • Ecosystem preference (CUDA / ROCm / JAX / Neuron)
  • Budget ($2-10M/rack)
  • Deployment location (on-prem / cloud)