Apple192 GB VRAM800 GB/s$0/hr

GPT-OSS 20B (MoE) on Mac M2 Ultra (192GB unified): Does It Fit, How Fast, What It Costs

GPT-OSS 20B (MoE) at FP8 (E4M3) needs 20.9 GB for its weights and 22.6 GB in total at 8K context with 8 concurrent users. That means it fits on a single Mac M2 Ultra (192GB unified). Every figure is computed by tokcalc's open-source capacity model.

Fits?
Yes, 1 GPU
22.6 GB needed / 192 GB available
Decode speed
216.7
tok/s per stream
Aggregate
1875
tok/s @ 8 concurrent
Cost / 1M tokens
—
1× Mac M2 Ultra (192GB unified) · quote required
What changes as you add Mac M2 Ultra (192GB unified)s
Tensor-parallel degree vs VRAM, throughput, and cost. Cost scales linearly with GPU count; decode throughput scales near-linearly too, so cost per token stays roughly flat.
GPUsTotal VRAMVRAM neededDecode tok/sAggregate tok/sTTFTCost / 1M tokFits?
1× Mac192 GB22.6 GB216.718754.40 s—✓
2× Mac384 GB22.6 GB368.375002.22 s—✓
4× Mac768 GB22.6 GB736.7300001.12 s—✓
8× Mac1536 GB22.6 GB1473120000576.8 ms—✓
Concurrency vs context length
On 1× Mac M2 Ultra (192GB unified) at FP8 (E4M3). KV cache scales linearly with context, so this is the number of users your GPUs can actually serve.
ContextConcurrent usersKV per requestFits?
4K1648206.2 MB✓ fits
8K836412.3 MB✓ fits
33K2111.61 GB✓ fits
66K1063.22 GB✓ fits
131K536.44 GB✓ fits
Other GPUs that run GPT-OSS 20B (MoE)
Same workload, same quantization — ranked by value (tokens per dollar-hour).
GPUGPUsAggregate tok/sCost / 1M tok
RTX 5090 32GB1×5824$0.0238
RTX 3090 24GB1×2472$0.0281
RTX 4090 24GB1×3276$0.0339
AMD MI300X 192GB1×17225$0.0403
V100 SXM2 32GB1×2925$0.0475
A100 40GB SXM41×5038$0.0496
Frequently asked questions

Can GPT-OSS 20B (MoE) run on Mac M2 Ultra (192GB unified)?

Yes. GPT-OSS 20B (MoE) at FP8 (E4M3) needs 22.6 GB including the KV cache, and Mac M2 Ultra (192GB unified) has 192 GB — so it fits on a single GPU.

How fast is GPT-OSS 20B (MoE) on Mac M2 Ultra (192GB unified)?

216.7 tokens/sec per stream and 1875 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 4.40 s, and inter-token latency is 4.6 ms.

What does GPT-OSS 20B (MoE) on Mac M2 Ultra (192GB unified) cost per million tokens?

— per 1M output tokens at 1× Mac M2 Ultra (192GB unified). That is the marginal hardware cost — add idle capacity, ops, and egress for the real figure.

How much VRAM does GPT-OSS 20B (MoE) need on Mac M2 Ultra (192GB unified)?

20.9 GB for the weights at FP8 (E4M3), plus 1.6 GB of KV cache for 8 concurrent requests at 8K context — 22.6 GB total. Every extra concurrent user at long context costs more, so leave headroom.

Is Mac M2 Ultra (192GB unified) the right choice for GPT-OSS 20B (MoE)?

It works. Compare — per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 4× RTX 3080 10GB at $0.006615, and the fastest is 1× Cerebras CS-3 (WSE-3) at 65000 tok/s.

Adjust the workload yourself

Same formulas, your context length, concurrency, and quantisation.

Open in calculator