Mac M4 vs RX 9070 XT — same Qwen, three numbers

· 3 min read

Same model, same quant tier, two machines, two completely different regimes. Qwen3.6-27B at Q4 runs at 9.8 t/s on a MacBook Pro M4 48 GB and 32.8 t/s on an RX 9070 XT 16 GB — the GPU is 3.3× faster in single-stream. But pile 8 concurrent requests on the Mac and total throughput climbs to 26.9 t/s, while the GPU stays stuck on single-stream. And the 9070 XT pulls a card the Mac bench never played: 125 t/s on a 35B MoE.

Single user, GPU wins. Multi-user, Mac wins. MoE, GPU wins again — by a lot.

Setup#

                MacBook Pro M4              RX 9070 XT
─────────────────────────────────────────────────────────────────
Memory          48 GB unified               16 GB GDDR6
Runtime         oMLX                        llama.cpp Vulkan/RADV
Quant           UD-MLX-4bit (Unsloth)       Q4_K_S (bartowski)
Price           ≈ 2 700 €                   ≈ 600 € (GPU only)

Same model family (Qwen3.6-27B), same quant tier (4-bit), different runtimes. Mac numbers come from the Qwen3.6 Mac bench; 9070 XT runs llama.cpp build 8401 on NixOS unstable, kernel 6.19.9, mesa 26.0.3.

Single request — Qwen3.6-27B Q4#

Metric                M4 48 GB     9070 XT 16 GB     Δ
─────────────────────────────────────────────────────────
tg128 (t/s)                9.8              32.8     +235%
pp (t/s)                 124.1             971.0     +682%
peak memory             25.2 GB           14.8 GB      —

In a one-user, one-prompt world, the 9070 XT wins outright. Both machines are memory-bound on inference; GDDR6 just moves bytes faster than M4 unified memory. Smaller and faster beats bigger and cooler when nothing else is fighting for the bus.

(pp rates measured at different prompt lengths — 1024 on Mac, 512 on GPU. Treat the comparison as directional.)

Plot twist: the Mac scales#

Throw 8 concurrent prompts at the M4 and total throughput jumps from 9.8 to 26.9 t/s — 2.74× scaling on a single machine, no extra hardware. The 9070 XT bench above is single-stream only; on a 16 GB card, the KV cache eats the VRAM headroom well before you reach 4× concurrency.

Concurrency       M4 48 GB (t/s)     9070 XT 16 GB (t/s)
──────────────────────────────────────────────────────────
1×                          9.8                    32.8
8×                         26.9                      —    ← VRAM-limited

This is the architectural plot twist most reviewers miss. Unified memory is the Mac’s batching superpower: KV cache and weights share the same 48 GB pool, so 8× context fits without spillover. A 16 GB GPU has no equivalent move. On aggregate t/s under load, the Mac closes most of the gap. For agents, evals, or any multi-tenant workload, that’s the number that matters — not the single-stream headline.

Plot twist the other way: MoE rewrites the math#

Same GPU, swap the dense 35B for a Mixture-of-Experts 35B with ~3B active parameters per token, and the 9070 XT runs 125 t/s — 3.8× the dense, on the same VRAM budget.

Model                            Size GB    tg128 (t/s)
────────────────────────────────────────────────────────
Qwen3.6-27B (dense)               14.76          32.8
Qwen3.6-35B-A3B (MoE, ~3B act.)   15.45         125.4    ← 3.8× the dense
gpt-oss-20B (MoE, ~3.6B active)   10.90         186.1    ← absurd

Compute is paid on the active subset (~3B params); only weight movement still has to load every expert from VRAM. The dense heuristic tg128 ≈ 560 / size_GB predicts ~36 t/s for a 15 GB model — the MoE blows through that ceiling by 3.5×.

What it means in practice: a 16 GB GPU runs a 35B model at the speed of a 4B. The Mac’s 48 GB context advantage stops mattering when the GPU is 4× faster on the same MoE — and MLX can run MoE on Apple Silicon in theory, but the dense gap (3.3×) suggests it won’t flip in MoE territory either.

The call#

Same Qwen, two regimes, plus one architectural unlock.

  • Single-user, anything dense → 9070 XT. 3.3× faster, ~4× cheaper. End of debate.
  • Multi-user, batched workloads → M4. Unified memory turns one Mac into eight. 8× concurrency, 26.9 t/s aggregate.
  • MoE-first → 9070 XT. The math the Mac doesn’t have a benched answer for yet.

The Mac is the better laptop and the better batch server. The 9070 XT is the better single-user box and the better MoE platform. They’re for different jobs — pick the one that matches yours.