AI モデル共設計：ハードウェアに優しい大規模言語モデルの設計

AI performance comes down to three dimensions:

• Accuracy: How well the model reasons and produces outputs

• Throughput: How many tokens per second a datacenter can generate

• Interactivity: How responsive the model feels to a user, dominated by latency

Deployments must balance all three: High accuracy is wasted if responses are slow, and raw throughput means little if each user’s experience is laggy. Practical systems therefore optimize accuracy, throughput, and interactivity together.

This post focuses on throughput and interactivity, and how model-design choices shape both without sacrificing accuracy (we flag accuracy trade-offs where they arise).

Holding accuracy fixed, the problem becomes a two-dimensional Pareto frontier: improving one usually costs the other. The goal is to push the whole frontier outward, maximizing the area under the curve (see Figure 1, below).

We will start with LLMs, today’s most prominent AI workload. The trade-off has two perspectives: the system deployer, who prioritizes fleet throughput (tokens/sec), and the user, who values lower first-token and inter-token latency. The inverse of inter-token latency is tokens/sec/user, where higher means greater responsiveness.

A single response looks like: prompt → [first-token latency] → first token → [inter-token latency] → next token, and so forth.

This is a practical primer for model developers who want their models to run well on modern hardware. The idea is simple: models that align with the hardware don’t just run faster; they scale better, cost less, and reach wider adoption.

First, consider the deployment landscape, which varies along two axes (see Figure 2, below).

Workloads range from short to long context, and service goals from throughput-oriented (maximize tokens/sec) to latency-oriented (minimize response time); many fall in between.

Each quadrant needs different optimizations: long-context, throughput-oriented serving spends most of its time in attention, while latency-oriented serving adds model parallelism to shorten attention and FFN, at the cost of communication and fixed overheads. Short-context, throughput-oriented serving splits time more evenly across attention and FFN, and can benefit from parallelism (such as expert parallelism) at scale.

Amdahl’s law applies: optimizing one part helps only as much as the time it occupies. If attention is 77% of runtime, tuning the feed-forward layers yields only marginal gains; the attention path is where effort pays off. Knowing your regime tells you where to focus.

This post gives simple rules of thumb to make these decisions early, for anyone who wants the most from the hardware without being a systems engineer. Each subsequent chapter of the series will target a different dimension of hardware-aware design, helping you avoid compute bottlenecks, deploy frictionlessly at datacenter scale, match design to your use case, and scale in practice.

Design smarter. Deploy faster. Scale wider. Let’s begin.

Hardware-friendly dimensioning of linear layers in LLMs

A key transformer design choice is its aspect ratio: the balance between model width and number of layers (). In decoder-style models, these factors dominate how computation and memory are distributed across the stack. Width itself is set by two dimensions, the hidden dimension () and the intermediate projection dimension () in the MLP layers.

Together, , , and also shape how cleanly the model maps onto parallelism strategies and how well it scales across GPUs. This post looks at how these three choices affect throughput, interactivity, and scalability, with a close focus on the linear layers.

Role of arithmetic intensity

On any hardware, achievable performance is bounded by the roofline model. Where a workload lands depends on its arithmetic intensity, defined as the number of compute operations performed per byte of memory moved.

Workloads with low arithmetic intensity are capped by memory bandwidth (memory-bound); those with high arithmetic intensity are capped by the device’s peak compute throughput, in FLOPS (compute-bound). When the goal is maximum throughput (tokens per second), you want to drive the workload into the compute-bound region so the hardware’s full math capacity is used.

Latency-sensitive decoding is the opposite, it runs at low concurrency and is memory-bound, so reducing memory-access time is what lowers response latency.

One straightforward way to raise operations per byte is to increase batch size, but model shape matters too. Let’s look at how and decide whether a GEMM is compute-bound or memory-bound. When running on a single device, and set the shape of a GEMM of the form:

Following the convention used in typical linear-algebra libraries, is a matrix of size ( rows, columns) and is . The product has outputs, each a dot product of length , requiring fused multiply-adds (FMAs). Since each FMA is 2 FLOPs (one multiply, one add), the costs are:

where are the per-element byte counts set by the precision used. Taking as the inputs and as the weights, Table 1, below, maps the token count (), , and to a GEMM’s , , and for each linear layer.

*Table 1. GEMM dimensions (M, N, K) of the linear layers in a transformer block, expressed in terms of Tokens, H, and H′*

Consider a case where every GEMM is “square,” . A single GEMM then performs FLOPs while moving about elements of memory, so arithmetic intensity grows with . The practical consequence: when either or is small, the GPU spends proportionally more time moving data than doing math, even when the token dimension is large.

As a concrete example, take FFN-2 with a deliberately small and , using 4-bit inputs and 8-bit outputs on GB300. As Table 2 shows, below, even at large token counts (high GEMM-M) this layer stays memory-bound, dominated by data movement; the write cost dominates because the output is FP8 while the inputs are FP4, and the small reduction dimension keeps the GEMM memory-bound.

The plots in Figure 4, below, show this on GB300 silicon. With NVFP4 GEMMs and the token dimension (GEMM-M) fixed at a large 8192, we sweep the reduction dimension (GEMM-K) (shown on the left) and the projection dimension (GEMM-N) (shown on the right). In both cases throughput falls sharply when the swept dimension is small, confirming that a small or leaves the hardware underutilized. ( and correspond to or depending on the layer; see Table 1, above.)

This is an important point for model designers: model dimensions matter just as much as batch size for saturating compute.

*Guideline 1:* *For a fixed-parameter model, favor near-square weight matrices and avoid making either the projection or reduction dimension too small.*

But size alone isn’t enough. To reach high Tensor Core utilization, a GEMM’s dimensions must also map cleanly onto the underlying tiling geometry; poor alignment causes tile quantization, which reduces throughput even when arithmetic intensity is high.

How GPUs execute GEMMs

GPUs execute GEMMs by splitting the output matrix into tiles, each computed by a streaming multiprocessor (SM). On recent GPUs, SMs don’t have to work alone; they can cooperate on a single, larger tile: with clusterMMA, two neighboring SMs join forces on one tile, and a Cooperative Grid Array (CGA) lets a whole group of SMs work together as one cluster.

Cooperation improves data reuse, but it also enlarges the effective tile, so dimensions must be a multiple of a larger value to stay aligned. When a dimension is not a multiple of that effective tile, the edge tiles are only partially filled yet still launch and run a full tile’s worth of compute; the unused (padded) portion does no useful work, wasting cycles and lowering throughput. Figure 5 illustrates this: with a 256×128 base tile, clusterMMA, and a 4×2 CGA, sweeping GEMM-N in fine-grained steps produces local throughput maxima when N is a multiple of 256 (2×128 from clusterMMA) or 512 (2×256 from CGA).

To avoid this waste, choose model dimensions that are large multiples of the tile size and aligned with GPU cache-line widths. A multiple of 128 is a safe, portable floor; multiples of 256 (clusterMMA) or 512 (CGA) align with the larger cooperative tiles and capture the most throughput.

*Guideline 2: **Make model dimensions multiples of 128 at minimum to align with GPU tile sizes and cache-line widths, and prefer 256 or 512 to match the larger tiles formed by clusterMMA and CGA.*

Wider models are more hardware-friendly than deeper models

For a fixed-parameter budget, wider models offer higher arithmetic intensity and lower latency than deeper models through greater weight reuse and a shorter sequential critical path. This makes wider models advantageous for both throughput- and latency-oriented service goals.

That said, aspect ratio also affects model quality: depth contributes to representational power, so there is a useful width-to-depth band rather than “wider is always better.” Favor width only as far as accuracy holds, rather than stripping out layers just to make the model wider.

*Guideline 3:** When given a choice, prefer fewer, larger operations over many smaller ones; this maximizes arithmetic intensity and improves hardware utilization, benefiting both throughput and interactivity. In other words, wider transformer models are more hardware-friendly than deeper ones.*

Quantization as a performance lever

Quantization helps both compute-bound and memory-bound work, by raising math throughput and reducing memory traffic at once. Blackwell systems support NVFP4 along with other bit-width formats such as FP8 and FP16/BF16 (see Figure 6, below).

NVFP4 is designed specifically for balancing model accuracy and speed; it applies a fine-grained FP8 (E4M3) scale to each 16-value micro-block, plus a second-level FP32 scale per tensor. This hierarchical scaling sharply reduces quantization error while keeping the speed of 4-bit computation, letting NVFP4 closely match higher-precision accuracy across a wide range of LLM workloads (see Figure 7, below, DeepSeek-R1).

NVIDIA provides end-to-end tooling to make this practical: TensorRT Model Optimizer and LLM Compressor support post-training quantization (PTQ), quantization-aware training (QAT), and advanced calibration, quantizing models to NVFP4 with minimal accuracy loss.

As a result, most matrix-multiplication-heavy operations, such as the linear layers, can fully exploit NVFP4 for higher throughput while keeping model fidelity. To learn more about training with lower bit-width, check out Pretraining Large Language Models with NVFP4 (NVIDIA Research, 2025).

*Guideline 4: **When introducing computationally expensive operations into a network, consider whether they can be quantized at deployment. Designing layers that benefit from low-precision execution is key to getting the most out of modern GPUs.*

Large expert parallelism boosts throughput

In throughput-oriented serving the goal is simple: Serve as many users as possible, as fast as possible, on the fewest GPUs. Since most state-of-the-art LLMs are now Mixture-of-Experts (MoE), a highly effective way to do this is expert parallelism (EP): run attention data parallel and distribute the FFN experts across GPUs.

Using data parallelism (DP) for attention avoids a key limitation of tensor parallelism (TP), the expensive AllReduce needed to combine partial results. That overhead grows with concurrency and drags down throughput, making TP a poor fit here. DP attention scales more naturally: adding GPUs directly raises global concurrency, and higher concurrency means a larger GEMM-M in the MoE FFNs.

Assuming uniform token routing:

where global concurrency is concurrent tokens per GPU × number of GPUs, top-k is the* *experts activated per token (8 for DeepSeek-R1), and #experts is the* *total number of experts (256 for DeepSeek-R1). Because GEMM-M grows with concurrency and shrinks with model sparsity, pushing concurrency is the key to high GEMM utilization in sparse MoE models, and since per-GPU concurrency is capped by the KV cache footprint, widening EP across more GPUs is the main lever for raising it.

Even when each expert’s effective batch (its GEMM-M) is small, whether from limited concurrency or imbalanced token routing, widening EP still helps in two ways:

• Faster execution via aggregate bandwidth: Distributing experts across more GPUs increases the effective memory bandwidth available for loading expert weights, reducing end-to-end FFN latency.

• Reduced per-GPU memory footprint: Each GPU stores only a subset of experts, freeing memory to increase effective concurrency per GPU and better saturate compute.

EP does bring two challenges, all-to-all communication overhead and expert load imbalance, but NVIDIA’s Wide-EP feature in TensorRT-LLM addresses both:

• High-performance all-to-all kernels that scale EP across Blackwell multi-node NVLink systems, and

• An adaptive load balancer that redistributes expert traffic based on real-time load patterns to stay stable under skewed token and expert distributions.

For details, see the TensorRT-LLM documentation on expert parallel scaling. For throughput-oriented MoE deployments, the key lever is expert parallelism.

*Guideline 5: **Complex, sparse MoE models with large numbers of experts can still deliver high throughput when deployed with the right parallelism strategy, most notably by scaling expert parallelism wide.*

Design for pipeline parallelism

Pipeline parallelism becomes relevant when you disaggregate prefill and decode, which can be an effective serving strategy for certain model sizes, traffic patterns, and latency targets, as we analyzed in Beyond the Buzz: A Pragmatic Take on Inference Disaggregation (NVIDIA Research, 2025): running prefill and decode with different model-partitioning lets each be optimized and scaled on its own, raising overall throughput.

For prefill, maintaining high throughput while reducing first-token latency (FTL), especially on long contexts, takes aggressive parallelization. Chunked Pipeline Parallelism (CPP) is well suited here: it splits both the model layers across GPUs and the input context into chunks that flow through the pipeline (see Figure 8, below).

This lets prefill workers handle long sequences within tight FTL budgets without resorting to wide tensor parallelism (see Figure 9, below).

![Chart showing that as pipeline-parallel size grows from 1 to 32, first-token latency drops steadily while tokens/s/GPU stays roughly constant.

](https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-625x451.jpg)

<figcapt