High-Fidelity GPU Performance Modeling
Quantifying the Networking Wall in Blackwell Clusters
The Thermodynamics of Compute Density
The performance of an AI cluster is not simply the sum of its parts. As we scale from 8 GPUs in a single server to 32,768 GPUs in a hyper-scale cluster, the efficiency of the Interconnect and Scale-Out Network becomes the dominant factor. Without a balanced Compute-to-IO ratio, even the most powerful B200 Blackwell cluster will suffer from significant synchronization stalls.
The Memory Wall
Modern LLMs often hit the HBM3e bandwidth limit before saturated TFLOPS. Quantifying this "Memory Bound" state is key to selecting the right GPU for inference-heavy vs. training-heavy workloads.
NVLink Fabric
NVLink provides 1.8TB/s of throughput, creating a "System-on-a-Cluster" environment. Moving beyond the node into the scale-out fabric (Ethernet/IB) is where the majority of performance degradation occurs.
GPU ROOFLINE PERFORMANCE MODELER
Arithmetic Intensity vs. Hardware Limits
The GPU is waiting on HBM3e bandwidth. Arithmetic logic is idle.
Hardware is operating at peak TFLOPS. Limited by total CUDA/Tensor Cores.
Scale-Out Efficiency at 800G
With Blackwell, NVIDIA has aligned the network rail capacity to match the 800G OSFP ecosystem. This double-bandwidth approach is designed to maintain the NCCL efficiency needed for massive mixture-of-experts (MoE) models, which require frequent All-to-All communication patterns that are notoriously sensitive to fabric latency.
Modeling the Compute Stack: From Silicon to System
A GPU performance model must capture four distinct bandwidth tiers that form a classic memory hierarchy pyramid. Tier 1: Register File and L1 Cache operates at over 100 TB/s aggregate within each SM (Streaming Multiprocessor). This is where tensor core instructions source their operands. Tier 2: L2 Cache and HBM3edelivers 3.35-8.0 TB/s depending on GPU generation. The H100 provides 3.35 TB/s across five HBM3 stacks; the B200 doubles this to 8 TB/s via eight HBM3e stacks using a 8192-bit memory interface. For a 175B parameter model using FP16, loading all weights from HBM once requires 350 GB of data movement—taking approximately 104ms on H100 and 44ms on B200. This is the fundamental "memory wall" bottleneck.
Tier 3: NVLink Intra-Node Fabric provides 900 GB/s (H100, bidirectional per GPU) or 1.8 TB/s (B200, bidirectional per GPU) , creating a shared memory domain across all 8 GPUs in an HGX baseboard. This 900 GB/s-to-1.8 TB/s bandwidth enables tensor parallelism where individual matrix multiplication columns are distributed across GPUs—each GPU computes a partial result and communicates with neighbors at effectively local HBM-latency levels. Tier 4: Scale-Out Network (InfiniBand NDR 400G or Ethernet 400G/800G) provides the lowest per-GPU bandwidth at 50 GB/s (400G) or 100 GB/s (800G). The performance modeler identifies at which scale the tier-4 bottleneck dominates, establishing the practical cluster size limit for your specific workload. A good rule of thumb: when tier-4 bandwidth drops below 5% of tier-2 bandwidth, training efficiency falls below 50%.
Roofline Model Analysis
The roofline model plots achievable TFLOPS against arithmetic intensity (FLOPs per byte). Workloads below the ridge point are memory-bound; above it are compute-bound. Most LLM inference falls in the memory-bound region, while large-batch training moves toward the compute-bound ceiling.
Bisection Bandwidth
The worst-case bandwidth when splitting a cluster into two equal halves determines how much cross-traffic can coexist during AllReduce. A fat-tree topology with full bisection bandwidth ensures no oversubscription; a 3:1 oversubscribed fabric will throttle collective operations by that same ratio.
Using the Modeler for Procurement and Design
The modeler supports comparative analysis across GPU architectures. Enter your target workload parameters—model parameter count, batch size, sequence length, training steps—and the tool calculates effective TFLOPS utilization for each GPU candidate. This reveals which GPU is actually faster for your workload, which often differs from peak specification comparisons. For example, the MI300X has higher theoretical FP16 TFLOPS than the H100 (1,307 vs. 989 TFLOPS respectively), but its lower HBM bandwidth (5.3 vs. 3.35 TB/s, a 1.58x advantage vs. the H100's compute-to-bandwidth ratio) means that for memory-bound inference workloads, the performance gap narrows considerably from the raw compute numbers. The B200's 4.5 PFLOPS of FP4 with 8 TB/s HBM3e creates a new performance regime where the compute-to-bandwidth ratio actually improves compared to previous generations.
Scale-out sizing uses the modeler's efficiency curves to identify the "sweet spot" GPU count for a given interconnect speed. At 400G per GPU, training efficiency typically remains above 80% up to approximately 1,024 GPUs for dense transformer models. Beyond that count, the gradient synchronization time grows faster than compute time shrinks, pushing efficiency below 65%. At 800G per GPU, the sweet spot extends to roughly 2,048-4,096 GPUs. The modeler also accounts fortopology constraints: a 2-tier fat-tree supports fewer GPUs at full bisection bandwidth than a 3-tier design, and rail-optimized topologies that colocate GPUs participating in the same NCCL rings on the same leaf switches can reduce cross-switch traffic by up to 50%.
Common GPU Performance Modeling Mistakes
Using peak TFLOPS instead of sustained TFLOPS is the most pervasive error. GPU peak specifications assume 100% tensor core utilization with data perfectly resident in registers—a condition never achieved in real training. Sustained TFLOPS for typical transformer training averages 35-55% of peak due to memory stalls, kernel launch overhead, and wavefront quantization effects. The modeler applies workload-specific efficiency factors derived from published MLPerf benchmarks and internal profiling data.
Ignoring power and thermal throttling can overstate performance by 10-20%. An H100 draws 700W at full load; in a DGX H100 with 8 GPUs, the total system draw exceeds 10 kW. Without adequate facility cooling, GPUs thermally throttle, reducing clock frequencies and memory bandwidth. The modeler includes a thermal headroom parameter that reduces effective performance when ambient temperatures approach the throttling threshold (typically 85°C junction temperature).Overlooking memory fragmentation—KV cache allocation, optimizer state storage (AdamW requires 8-12 bytes per parameter for momentum and variance), and activation checkpointing all consume HBM capacity, reducing the effective batch size below theoretical maximums. The modeler validates memory capacity against requested batch sizes and flags configurations that would trigger out-of-memory errors.
Multi-Vendor Comparison
The modeler supports NVIDIA H100/H200/B200, AMD MI300X/MI250X, and custom silicon profiles. For each architecture, it captures SM/CU count, tensor core throughput per data type (FP64, FP32, TF32, FP16, BF16, FP8, FP4, INT8), HBM capacity and bandwidth, NVLink or Infinity Fabric topology, and PCIe or SXM form factor constraints. This enables true apples-to-apples comparisons for purchasing decisions rather than relying on vendor marketing numbers that cherry-pick favorable benchmarks.
Technical Standards & References
NVLink vs InfiniBand Bandwidth Amplification
In modern GPU-accelerated systems, the interaction between NVLink (intra-node GPU interconnect) and InfiniBand (inter-node fabric) defines the effective bandwidth amplification ratio available to distributed training workloads. Bandwidth amplification refers to the factor by which the inter-node fabric bandwidth appears magnified at the GPU memory level, because data does not need to traverse the PCIe bus multiple times. In a traditional architecture without NVLink, GPU-to-NIC data flows through the PCIe root complex: GPU → PCIe switch → NIC, consuming PCIe bandwidth for every network operation. With NVLink and GPUDirect RDMA, the NIC can write directly to any GPU's HBM memory via the NVLink fabric, bypassing the PCIe bottleneck. The effective bandwidth amplification ratio is: A = (NNVLink × BNVLink) / (NNIC × BNIC), where NNVLink is the number of NVLink connections per GPU, BNVLink is the per-link bidirectional bandwidth, and NNIC and BNIC are the NIC equivalents. For an H100 HGX with 8 GPUs, each offering 18 NVLink 4.0 links at 50 GB/s per direction, the aggregate intra-node GPU-to-GPU bandwidth is 900 GB/s, while a single 400 Gbps (50 GB/s) NIC provides inter-node bandwidth.
The GPU-to-NIC affinity mapping is the critical architectural variable that determines whether NVLink bandwidth amplification can be realized. In a 4-GPU, 4-NIC configuration where each GPU is co-located on the same PCIe root complex as a dedicated NIC (1:1 mapping), the amplification ratio is effectively 1.0—no NVLink benefit for inter-node traffic because each GPU communicates directly with "its" NIC via the local PCIe bus. The maximum amplification is achieved in an NVSwitch fully-connected topology where all 8 (or 16) GPUs are interconnected through a non-blocking NVSwitch fabric, and the NICs are connected to the NVSwitch rather than to individual GPUs. In this configuration, any GPU can communicate with any NIC through the NVSwitch at the full NVLink bandwidth, and the amplification ratio approaches the number of GPUs per NIC. For an 8-GPU NVSwitch system with 4 NICs, the effective inter-node bandwidth per GPU is: (BNIC × 4) / 8 = BNIC / 2, but because each GPU can use all 4 NICs simultaneously via the NVSwitch fabric, the aggregate cluster-level bandwidth is 4 × BNIC, regardless of which GPU submits the RDMA operation. Our modeler captures this topology-dependent amplification effect and computes the effective per-GPU bandwidth for any GPU-to-NIC mapping, enabling architects to optimize the trade-off between NIC count and NVSwitch port utilization.
At the software collective level, the interaction between NVLink and InfiniBand is orchestrated by NCCL's hierarchical algorithm. NCCL decomposes the all-reduce operation into two phases: an intra-node reduction over NVLink (typically using NVLink SHARP for in-network aggregation) and an inter-node exchange over InfiniBand. The intra-node phase on an 8-GPU NVSwitch system completes in approximately 2-4 μs for the FP32 gradient of a 1,000-parameter tensor, while the inter-node phase requires at least (2 × α + message_size / (NNIC × fractional_bandwidth)) depending on the InfiniBand subnet topology. The optimal chunk size for the inter-node exchange is determined by the NVLink-to-InfiniBand bandwidth ratio: larger chunks amortize the InfiniBand latency but increase the memory pressure on the intra-node NVLink accumulation buffers. Our model applies the Logistic Bandwidth Model where the effective inter-node bandwidth as a function of chunk size S follows: Beff(S) = Bpeak × (1 - exp(-S / S₀)), where S₀ is the characteristic chunk size at which 63% of peak bandwidth is achieved. For a 400 Gbps InfiniBand link with RoCEv2 overhead, S₀ is typically 32-64 KB, meaning that gradient tensors below 32 KB achieve less than 63% of the nominal link bandwidth due to protocol header overhead and interrupt coalescing inefficiency.
For large-scale clusters exceeding 512 GPUs, the bandwidth amplification ceiling is ultimately determined not by the NVLink or InfiniBand specification but by the bisection bandwidth ratio of the inter-node fabric. The NVSwitch non-blocking domain covers only the intra-node GPUs; once the collective operation requires cross-rack or cross-pod communication, the InfiniBand fabric's oversubscription ratio becomes the dominant constraint. In a typical three-tier Clos fabric (Edge → Aggregation → Core), the oversubscription ratio is 1:1 for a fully non-blocking design or 3:1 for a cost-optimized design. The effective bandwidth amplification Aeff = Aintra / Ointer, where Aintra is the NVLink-to-NIC amplification within the node and Ointer is the inter-node oversubscription ratio at the Spine-to-Leaf link layer. For an Aintra of 8× (typical for 8-GPU node with 1 NIC) and an Ointer of 2.5× (a moderately oversubscribed cluster), the effective amplification across the cluster is only 3.2×, meaning the NVLink advantage is significantly diluted at cluster scale. Our model provides fabric dimensioning recommendations: to preserve a minimum Aeff of 4× for the target model size and parallelism strategy, the InfiniBand oversubscription ratio must not exceed the NVLink amplification ratio divided by 4.
"You are our partner in accuracy. If you spot a discrepancy in calculations, a technical typo, or have a field insight to share, don't hesitate to reach out. Your expertise helps us maintain the highest standards of reliability."
Contributors are acknowledged in our technical updates.