← All posts
v1.3 Technical Blog

GPU-to-CPU Deployment Ratios in
Modern Server Deployments

Published · 6 min read · MANISH AI

GPU:CPU planning is becoming workload-specific again. For long-context, retrieval-heavy, and orchestration-heavy inference, CPU capacity, memory bandwidth, and data movement can become first-class constraints.

This post examines where and why CPU importance rises in specific inference regimes. It is not a universal claim about all AI workloads, and it should not be read as a replacement for workload profiling.

Scope & Definitions

This analysis focuses on inference infrastructure. Training ratios often remain in the 4:1-8:1 range on NVLink systems. Unless stated otherwise, ratios here are physical GPU:CPU socket counts at the node level, not logical threads or vCPUs.

[Observed: Shipping Systems] [Model: Back-of-envelope] [Projection: Architectural Trend]

Training

Observed

8:1 – 4:1

H100/H200 NVL8. GPU-bound matmul. CPU feeds data loaders and NCCL.

Short-context Inference

Observed

8:2 – 4:2

DGX H100, < 8k tokens. KV cache fits in HBM. CPU for tokenization + scheduling.

Long-context / RAG

Projection

1:2 – 2:4

128k-1M tokens. CPU memory bandwidth and capacity start to matter for retrieval, paging, and KV placement. GPU still dominates dense compute, but system balance shifts.

Platform Generations & The Generational Shift

Platform Typical GPU:CPU Interconnect Classification Dominant Constraint
Hopper DGX H100 8:2 NVLink 900 GB/s die-to-die
PCIe 5.0 128 GB/s to host
Observed GPU HBM3 capacity, NVLink BW
GH200 Grace-Hopper 1:1 NVLink-C2C 900 GB/s
Cache-coherent
Observed Balanced HBM + LPDDR5X
GB200 Grace-Blackwell 2:1 NVL72 NVLink-C2C 1.8 TB/s Observed LPDDR5X BW for CPU
Rubin Ultra 1 1:2 (projected for some inference profiles) NVLink-C2C > 1.8 TB/s Projection CPU memory BW, socket count

1 This row is a forward-looking workload projection, not a shipping product specification. It reflects how coherent CPU-GPU links, larger host-memory pools, and orchestration-heavy inference may push some future systems toward more CPU capacity per GPU.

Why the balance shifts for affected workloads

Training-era sizing rules assume dense GPU compute is the primary limiter. For long-context inference, the roofline can shift toward memory movement and cache placement. KV-cache size grows with batch, layers, hidden size, and context length. Once working-set size stops fitting comfortably in local HBM, architects are forced to choose among sharding, paging, offload, or more aggressive batching trade-offs.

Coherent CPU-GPU fabrics make host-memory participation far more practical than it was in PCIe-only designs. That does not make CPUs “the new accelerator”; it means some inference stacks become constrained by memory bandwidth, page movement, retrieval, tokenization, and orchestration before raw dense compute is exhausted. In those cases, adding GPU FLOPS alone may not improve end-to-end throughput or tail latency.

Workload Sizing Calculator

Back-of-the-envelope estimator

Assumes a simplified roofline-style model and ignores NUMA effects, scheduler behavior, fragmentation, batching policy, quantization details, and software overheads. Use it as directional intuition, not procurement guidance. Any real sizing exercise should be validated with traces and production-like load tests.

Inputs

Estimate: Outputs Model

Run calculation to see estimates...

Section 5: Interconnect Nuance & Coherent Memory

NVLink-C2C at 900GB/s-1.8TB/s is cache-coherent

Coherent CPU-GPU links materially change the programming model. In PCIe-only systems, host participation often means explicit copies, managed-memory penalties, or page-fault-driven migration. In coherent C2C designs, some host-resident structures become cheaper to access and reason about. That does not eliminate locality concerns, but it reduces the software friction of hybrid memory use.

Observed: Shipping Systems On PCIe-attached systems, host-to-device interaction is still much more constrained than on coherent packages. Managed-memory or spill-heavy designs can incur painful migration behavior and large latency penalties once access patterns become irregular. That is one reason long-context inference and memory-heavy retrieval pipelines often degrade sharply when they outgrow local HBM.

Observed: Shipping Systems On GH200- and GB200-style coherent systems, the CPU and GPU can participate in a much tighter memory model. The practical effect is that host memory becomes far more usable for certain classes of inference support work, even though local HBM is still the preferred home for the hottest data. That distinction matters: coherence improves what is feasible, but it does not erase the gap between “accessible” and “optimal.” [1]

// Hopper: Explicit copy required
void* cpu_buf = malloc(1<<30);
void* gpu_buf;
cudaMalloc(&gpu_buf, 1<<30);
cudaMemcpy(gpu_buf, cpu_buf, 1<<30, cudaMemcpyHostToDevice); // PCIe BW
// Grace-Blackwell: Direct access
void* unified_buf = malloc(1<<30); // CPU NUMA node
gpu_kernel<<<...>>>(unified_buf); // Coherent access path; locality still matters

Projection: Architectural Trend This opens the door to cleaner hybrid-memory designs. Tokenization, retrieval, safety checks, schedulers, and some queueing/orchestration logic can live closer to the CPU while the GPU stays focused on dense compute. The ratio implication is simple: if your bottleneck is on the host side, adding GPUs alone will not fix it.

Section 8 & 9: Power, Cooling & Facility Impact

Thermal density note

Power and facility planning do not get simpler just because per-node GPU count may fall in some inference-oriented designs. A more CPU-heavy node can trade one bottleneck for another: socket density, DIMM count, memory-channel utilization, and cooling complexity all become more important. The architectural point is not that every future rack will be easier to cool; it is that planners may need to think about whole-node balance rather than GPU thermals in isolation.

Section 8: Rules of Thumb for Capacity Planning

Condense previous verbose list to three sizing rules. For detailed analysis, use calculator.

Section 9: Facility Manager Checklist

Hopper-era DC

  • • 6-8kW per 8U node
  • • 4x 200GbE NICs
  • • 2 CPU sockets, 24-32 DIMMs
  • • GPU cold plates, air for CPU

Rubin-era DC Projection

  • • 2.5-3.5kW per 2U node
  • • 2x 400GbE/800GbE NICs
  • • 4 CPU sockets, 48-64 DIMMs
  • • Full-node liquid, rear door HX

References, Methodology & Confidence Levels

[1] NVIDIA public architecture materials and GTC sessions covering Grace Hopper, Grace Blackwell, and NVLink-C2C. These establish the existence and direction of coherent CPU-GPU memory fabrics used throughout this post.

[2] Vendor architecture briefs, product pages, and public conference talks for H100, GH200, and GB200 platforms. These support the observed system-level ratios and interconnect discussion.

[3] All calculator outputs are first-pass analytical estimates derived from simple working-set and bandwidth assumptions. They are included to illustrate directionality, not to claim benchmark-equivalent precision.

Anything explicitly labeled [Projection] should be read as architectural interpretation rather than established field consensus. Where exact public benchmarking is unavailable, the post intentionally uses softer language and separates observed platforms from forward-looking projections.