Dual-GPU AI Workstations Without NVLink: Planning a 192 GB Tower

Overview

A dual-GPU mid-tier workstation — two 96 GB professional Blackwell cards for 192 GB aggregate — is the largest AI machine that still fits under a desk, and the point where planning becomes engineering. The critical fact buyers miss: current-generation workstation Blackwell cards have no NVLink; the two GPUs talk over PCIe 5.0 x16 at roughly 128 GB/s bidirectional per card, versus 1.8 TB/s within each card’s own memory. That asymmetry decides which multi-GPU strategies work well. The 300 W Max-Q edition exists precisely for this build: NVIDIA rates it for up to four GPUs in one chassis, trading some peak clocks for thermal sanity.

Max-Q or full power: pick by GPU count

The full 600 W Workstation Edition maximises single-card throughput; the Max-Q variant holds the same 96 GB, same 24,064 CUDA cores, at a 300 W cap. Independent testing, including Puget Systems’ comparison, shows Max-Q giving up a modest slice of peak performance for half the power — a trade that flips decisively in its favour the moment a second card enters the chassis. Two 600 W cards plus platform exceeds what standard tower cooling and many office circuits handle gracefully; two Max-Q cards total 600 W of GPU and stay boring. Rule of thumb: one GPU, buy full power; two or more, buy Max-Q.

Life without NVLink

Tensor parallelism splits every layer across both GPUs and synchronises constantly — it is the strategy most sensitive to interconnect bandwidth, and over PCIe it pays a visible latency and throughput tax. Pipeline parallelism assigns whole layer blocks to each card and passes activations once per stage boundary — far lighter traffic, well suited to PCIe, at the cost of some pipeline bubble inefficiency. For inference, serving frameworks (vLLM, TensorRT-LLM) support both; on PCIe-only towers, pipeline-parallel or per-GPU model replicas are usually the better answer. For fine-tuning, FSDP/DeepSpeed-style sharding works over PCIe but profits from gradient-accumulation settings that reduce sync frequency. None of this is exotic — it just needs configuring with the interconnect in mind rather than copying H100-server recipes; our cluster networking article covers the same physics at rack scale.

Platform engineering: lanes, power, air

Two x16 Gen5 slots need CPU lanes that desktop sockets don’t have — this build wants a workstation platform (Threadripper PRO / Xeon W-class) with 128+ lanes, so both GPUs, NVMe and networking run unshared. Power: two Max-Q cards (600 W) plus a high-core CPU and storage lands near 1,000–1,200 W sustained; a 1,600 W-class PSU with headroom is the safe spec, and Indian buyers should verify the office circuit and UPS sizing — a loaded dual-GPU tower approaches a 1.3 kW wall draw. Cooling: blower/flow-through professional cards are designed to stack; consumer-style triple-fan cards are not. This validation burden is exactly what pre-integrated workstation SKUs exist to absorb.

What scales to two cards — and what doesn’t

Near-linear: batch image/video generation (independent jobs per GPU), parallel fine-tuning experiments, serving two different models, embedding pipelines, and unquantised 70B FP16 inference via pipeline parallelism where the alternative was “doesn’t fit at all.” Sub-linear: single-stream latency on one model split across cards — expect real gains but not 2×, with PCIe sync as the ceiling. Not at all: workloads bound by a single GPU’s compute on models that already fit in 96 GB — a second card adds capacity, not speed, unless you parallelise jobs. Buyers whose growth path is concurrency and users, not model size, should compare a proper server instead; see workstation vs GPU server.

Dual-GPU configurations at a glance

Frequently asked questions