Unified Memory vs Discrete VRAM for AI Coding: What Buyers Need to Know

Buyers building a machine for local AI coding keep hitting the same fork in the road: a big pool of slower unified memory, or a smaller pool of much faster discrete VRAM. Unified memory versus discrete VRAM for AI coding is not a simple better-or-worse call. It is a trade between how large a model you can load and how fast that model generates tokens, and the right answer depends entirely on the size of the model you actually want to run.

Quick Answer

Discrete VRAM wins on raw speed for any model that fits inside it, because cards like the RTX 5090 push roughly 1,792 GB/s of memory bandwidth against around 614 GB/s on a 128GB unified-memory Mac. Unified memory wins on capacity, holding 70B-plus coding models that simply will not load on a 32GB card. For most local AI coding, pick the platform that fits your target model first, then worry about speed.

What Each Memory Type Actually Does

A discrete GPU has its own dedicated VRAM soldered next to the chip, and inference speed during token generation is governed almost entirely by how fast the model weights can be streamed out of that memory. The RTX 5090 carries 32GB of GDDR7 at very high bandwidth, so anything that fits in 32GB runs quickly. The catch is the ceiling: once a model is larger than the card, weights spill into system RAM and every layer has to cross the PCIe bus, which collapses the speed advantage.

Unified memory, used in Apple Silicon and the newer integrated AI platforms, shares one large pool between CPU and GPU. A 128GB configuration lets the GPU address the whole pool, so it can hold a model four times larger than a 32GB card can. The bandwidth is lower, so a model that fits both will run faster on the discrete card, but unified memory can load things the card cannot touch at all.

Speed vs Capacity in Real Coding Workloads

Models that fit on both

For a dense 32B coding model quantised to roughly 19 to 20GB, both platforms load it comfortably. Here the discrete card stretches its legs: an RTX 5090 commonly lands in the mid-40 tokens per second range while a 128GB unified machine sits closer to the mid-20s. If your daily driver is a 7B to 32B model and responsiveness matters, the fast VRAM is the better experience.

Models only one platform can hold

A 70B model at a usable 4-bit quantisation needs more than 32GB. On the card those weights spill to system RAM, and the per-layer PCIe synchronisation drags throughput down to a level comparable with, or slower than, a unified-memory machine that holds the whole model in one place. At that size the capacity advantage flips the result, and the big memory pool becomes the practical winner.

The shift toward Mixture-of-Experts

A lot of strong 2026 coding models are Mixture-of-Experts designs, where a model might total 80B parameters but only activate around 3B per token. Because only the active experts are read each step, an 80B MoE coder can generate several times faster than a small dense model on the same hardware, while being a larger and generally stronger model. These designs reward large memory pools, which plays directly to unified memory's strength.

The PCIe Spill Problem, in Numbers

When a model overflows VRAM on a discrete card, the weights that do not fit must cross the PCIe bus on each token generation pass. The RTX 5090 moves data internally at roughly 1,792 GB/s. A PCIe 5.0 x16 slot tops out near 64 GB/s in ideal conditions, and real-world LLM workloads often see effective throughput well below that. The result is a two-orders-of-magnitude drop in effective bandwidth for the overflow layers, which turns a machine that was generating 40 tokens per second into one struggling to reach single digits. A unified-memory system with 128GB never hits this cliff, because the model fits entirely in a single addressable pool.

Practical Configuration Guide for SA Builders

For most South African developers building a local coding rig, the starting questions are: which models do I actually run today, and which will I want in 12 months? If the answer is firmly in the 7B to 32B range, a high-bandwidth discrete GPU keeps every response snappy and leaves budget for more system RAM and storage. If 70B-class reasoning or large MoE coders are the target, plan around unified memory or a multi-GPU build rather than discovering the VRAM cliff after the rig is already specced.

A middle path many local builders land on is a fast GPU for everyday work, paired with the acceptance that very large models stay out of reach until an upgrade. Whichever direction you go, the AI PC range at Evetech lays out machines configured for local model work, and you can sanity-check pricing against the best-selling PCs at Evetech to see where the value sits right now.

Frequently Asked Questions

Is 32GB of VRAM enough for AI coding?

For 7B to 32B coding models at 4-bit, yes, and a 32GB card runs them fast. The limit shows up at 70B and above, where 32GB is not enough and weights start spilling to system RAM, which sharply reduces speed.

Why is unified memory slower per token if it can hold bigger models?

Token generation speed tracks memory bandwidth, and discrete VRAM has much higher bandwidth than unified memory. Unified memory trades that bandwidth for a far larger capacity, so it loses on small models that fit both but wins when a model is too big for the card.

Do Mixture-of-Experts models change the maths?

Yes, significantly. An MoE model only reads its active experts each token, so a large MoE coder can run several times faster than its total size suggests. This makes large unified-memory pools especially attractive for the strongest current coding models.

Will a model that spills to system RAM still work on a discrete GPU?

It will run, but slowly. Every layer that lives in system RAM has to cross the PCIe bus during inference, and that synchronisation overhead usually erases the card's bandwidth advantage, sometimes dropping below a unified-memory machine of equivalent capacity.

Which should I buy for general use beyond coding?

A discrete GPU is more flexible if you also game, render, or train, since that ecosystem is mature and the card is fast at everything that fits. Choose a large unified-memory machine when running very large models locally is the main job and you can accept lower per-token speed.