Mac Studio M5 Ultra vs RTX 5090 Build

Verdict

Two machines that cost Rs 2.2 million and Rs 4.6 million, asked to do one job: generate images and video for 20+ AI models, every day, for five years.

🟢 The RTX 5090 build. It is not close — and the Mac costs twice as much.

Performance: on your real pipelines (Forge + ControlNet + ADetailer + LoRAs, FP8 models, Wan / LTX video, LoRA training) the 5090 lands between 3× and 8× faster. On plain vanilla SDXL the Mac is surprisingly close — but nothing you run is vanilla.

Longevity: the Mac is frozen on the day you buy it. The PC swaps its GPU in 2028 and keeps going. Every AI video model that ships next year ships CUDA-first; the Mac gets a port months later, slower, if at all.

Scaling: at 20 models the currency is clips per hour per rupee. One 5090 plus your existing 3090 gives two render nodes for less than half the Mac's price.

The one thing the Mac does that the PC cannot: hold a 200B-parameter LLM in 256 GB of memory. You don't need that — your chat brains already run on the VPS.

The two contenders

🍎 Mac Studio (as configured)

M5 Ultra

36-core CPU · 80-core GPU with Neural Accelerators · 256 GB unified memory · 1.2 TB/s · 2 TB SSD · 6× Thunderbolt 5 · $16,499 · ships in 16–18 weeks · sealed, nothing user-replaceable

🟩 Zestro high-end build + RTX 5090

7800X3D + 5090

Ryzen 7 7800X3D · MSI X870E-S EVO · 64 GB DDR5-6000 · Samsung 9100 Pro 2 TB Gen5 · MSI A1250GL 1250 W ATX 3.1 · Lian Li O11 Vision Compact · MSI A13 360 AIO · RTX 5090 32 GB GDDR7 · 1.79 TB/s · ≈ Rs 1.75–2.25 M depending on which 5090 · every part replaceable

What you are really comparing The Mac is one integrated chip whose GPU shares the system's 256 GB. The PC is a dedicated 32 GB GPU with 5.6× the raw tensor compute, attached to 64 GB of system RAM. The Mac wins on memory capacity. The PC wins on memory speed and on everything that does arithmetic.

Raw silicon

Before software, before ecosystems — what the two chips can physically do per second.

Metric	Mac Studio M5 Ultra	RTX 5090	Edge
FP16 matrix compute (dense)	≈ 40 TFLOPS (M5 Max measured ~20; Ultra = 2 dies)	≈ 210 TFLOPS	5090 · ~5×
FP8 compute	not native — emulated / community patch	≈ 420 TFLOPS native	5090 · ~10×+
FP4 compute (Blackwell)	none	≈ 840 TFLOPS	5090 only
Memory bandwidth	1,200 GB/s	1,792 GB/s	5090 · 1.5×
Memory available to the GPU	256 GB (shared)	32 GB dedicated + 64 GB system	Mac · 8×
Power at full load	~270 W whole machine	~575 W GPU alone, ~900 W system	Mac
Apple's own claim	"4.3–4.5× the peak AI compute of M3 Ultra" — true, and it still lands at roughly one-fifth of a 5090 in FP16, with no FP8/FP4 path at all.

Why FP8 and FP4 decide this Every modern image and video model — FLUX, Krea 2, Wan 2.2, LTX-2.3, HunyuanVideo — ships in FP8 because it halves memory and doubles speed on tensor cores. Blackwell adds FP4 on top. Apple Silicon has no hardware FP8: ComfyUI on a Mac either falls back to bf16 (double the memory, slower) or leans on a community monkey-patch that emulates it. That single gap is most of the 3–8× you'll feel day to day.

Your actual stack — what runs where

You are not buying a benchmark. You are buying a home for Diana HD, Diana FAST, Bella, Big Lust, and 16 more like them.

Tool in your pipelines	On the 5090 (Windows / CUDA)	On the Mac (MPS)
Forge WebUI (your generator)	Native, full speed	Runs on MPS, poorly maintained there — expect breakage on updates
ControlNet (mandatory on every image)	Native	Works; preprocessors slow, some CUDA-only
ADetailer face-bake (Diana FAST, Bella)	Native	Works, slower — no FP8 path
inpaint.py face-apply (Diana HD)	Native	Port needed
DMD2 4-step distilled checkpoints	Native	Works
Krea 2 / FLUX refiner (ComfyUI, FP8)	Native FP8	bf16 fallback or ComfyUI-AppleSilicon-FP8 patch; some custom nodes hard-require `cuda`
Joy Caption (vision captioning)	Native	Works via MPS, slower
SageAttention / TeaCache / torch.compile	Native — the video speed multipliers	Not available
kohya LoRA training (persona LoRAs)	Native	Works, 3–6× slower
Wan 2.2 / HunyuanVideo in ComfyUI	Native, FP8/GGUF	"No stable MPS path yet" — Draw Things is the only working route
Your automation: .bat launchers, WMI Forge restart, Chrome :9222, bundle stager, Windows scheduled tasks	Unchanged	All rewritten for macOS

The hidden cost nobody prices Two years of pipeline work — the caption trigger, the evaluate loop, the crop-tag UI, the Forge watchdog, the bundle stager — is Windows + CUDA. Moving to the Mac is not "install ComfyUI." It is a port, a re-verification of every step, and living permanently one release behind on every tool that matters. That work has a cost, and it is not zero even if the Mac were free.

Image generation

Honest first: on plain SDXL the Mac is closer than anyone expects. Then the gap opens.

SDXL 1024², 30 steps, one image, no extras — seconds (shorter = better)

M5 Ultra (Draw Things, est. from M5 Max 7 s)

~4–5 s

RTX 5090 (ComfyUI / Forge)

~3.5 s

Your RTX 3090 today

~10–12 s

Your Diana FAST pipeline — DMD2 + ControlNet + ADetailer face-bake + 2 LoRAs, end-to-end per image

M5 Ultra (est. — no FP8, no fused attention)

~25–35 s

RTX 5090

~8–10 s

Your 3090 today (measured)

22–27 s

FLUX / Krea 2 refiner, FP8, 1024²

M5 Ultra (bf16 fallback)

~40–60 s

RTX 5090 (native FP8)

~10–13 s

Read this carefully On the one test the Mac is designed to win — a single SDXL image in Apple's own app — it reaches ~80% of a 5090. The moment you add the things every one of your images needs (ControlNet pose lock, face bake, LoRAs, FP8 refiner) it drops to a quarter to a third of the 5090's throughput. And it would run through Draw Things, not Forge — a different tool, different presets, different bugs.

Video generation — the part that decides the next five years

You said video for every model. This is where the two machines stop being comparable.

Model · task	RTX 5090	Mac Studio M5 Ultra	Ratio
Wan 2.2 14B · I2V 5 s @ 720p	~125 s per clip (warm), FP8 / GGUF, 32 GB is enough	Draw Things only. Extrapolated from LTX on M-series at ~⅛ of a 4090: ~8–15 min per clip	5090 · ~4–7×
LTX-2.3 distilled · I2V with synced audio	~22 s per clip	est. 90–180 s	5090 · ~4–8×
HunyuanVideo 1.5	Native	No stable MPS path	5090 only
SageAttention + TeaCache (the 2–3× video multipliers)	Yes	No	5090 only
Next year's video model, day one	CUDA release	Wait for a port	5090

What 32 GB actually buys you for video Wan 2.2 14B's transformer is ~14 GB in FP8 or ~8 GB in GGUF Q5. LTX-2.3 fits with room for 4K. The 5090 is the strongest single consumer card for local video today, and every published guide reaches the same conclusion. The Mac's 256 GB is not the bottleneck it solves — compute is the bottleneck, and the Mac has one-fifth of it.

Throughput math at 20 models Say 5 clips per model per day = 100 clips/day. Wan 2.2 on the 5090: ~3.5 hours. On the Mac: ~16–25 hours — i.e. it can't be done in a day. LTX-2.3 on the 5090: 37 minutes. The 5090 does in an afternoon what the Mac can't finish overnight.

LoRA training

Twenty models means twenty persona LoRAs, and each one gets retrained as the look evolves.

M3 Max 96 GB (published)

2–3 h

M5 Ultra (est., 2× M5 Max GPU)

~60–90 min

RTX 4090 (published)

20–30 min

RTX 5090

~15–20 min

SDXL LoRA, 30–40 images, kohya_ss. Mac numbers via the MPS backend, which works but has no fused optimisers, no bf16 tensor cores, and no bitsandbytes 8-bit Adam.

Scaling to 20+ models

At this size you stop thinking "which computer is faster" and start thinking "how many clips per hour per rupee, and can I add a second box next year."

5090 build

2 nodes

5090 as the primary + your existing 3090 kept alive as render node #2 (a standing decision). Same Windows automation, same Forge presets, batch.json fan-out already works across boxes.

Mac Studio

1 node

The 3090 rig can't share work with it — different OS, different toolchain. The Mac becomes an island, and the second node is another Rs 4.6 M Mac.

Cloud burst (both)

RunPod

When 20 becomes 40, you rent 5090/H100 by the hour. Every cloud image is CUDA. A CUDA workflow moves to the cloud unchanged; a Mac workflow has to be rebuilt to burst.

Rupees per clip Mac: Rs 4.6 M ÷ ~7 Wan clips/hour. 5090 build: Rs 2.25 M ÷ ~29 clips/hour. The Mac costs roughly 8× more per clip of throughput. That gap does not shrink at scale; it compounds.

Which one lasts longer

Hardware lifetime and useful lifetime are different questions, and the answer flips between them.

	Mac Studio M5 Ultra	RTX 5090 build
Physical lifetime	Excellent — sealed, cool, ~270 W, Apple supports macOS ~7 years	Good — 1250 W ATX 3.1 PSU, AIO, quality board; GPU is the hot part
Upgrade path	None. Memory, storage, GPU are all soldered. What you buy is what you have in 2031.	GPU swaps to a 6090 in 2028. AM5 takes a Zen 6 CPU drop-in. RAM to 128 GB. PSU already sized for it.
Software relevance in 2028	Every new model ships CUDA-first; Apple's port lands months later, without FP8/FP4. The gap widens each generation.	Day-one on every release. FP4 already in the silicon for models that haven't been written yet.
Resale	Strong — Macs hold value	GPU holds value (a 3090 still fetches Rs 250–300k six years on); rest depreciates normally
Failure mode	One part dies → whole machine to Apple, 16–18 weeks if replaced	One part dies → replace that part in Karachi the same week

The honest framing The Mac will physically outlast the PC. It will also be less useful every year it is alive, because the field moves on CUDA and the Mac cannot follow. The PC's GPU may be swapped twice in that time — and that is the point. "Lasts longer" for AI work means "can keep up," and only one of these can.

Where the Mac genuinely wins

So this isn't a hit piece — the Mac has real advantages. They just aren't yours.

256 GB unified memory

200B+ LLMs local

Runs DeepSeek-class or Llama-405B-class models entirely on device. The 5090 caps out around 70B at Q4. But: your chat brains run DeepSeek on the VPS — you don't run frontier LLMs locally, and Sam's rule is no paid Claude API for the brain, not "must be local."

Power & noise

~270 W

A third of the PC's draw, near silent. Real money on a Karachi bill, and it would sit comfortably behind your 600 W Riello UPS — which the 5090 build cannot.

Plain SDXL in Draw Things

~80% of 5090

Genuinely impressive. Apple's Neural Accelerators + Metal FlashAttention made the M5 line a real image-gen machine for hobbyists.

Reliability & support

Sealed box

No PSU sizing, no BIOS switches, no cable daisy-chaining, no thermal pads. It just runs.

If your work were different A one-person operation doing a few SDXL images a day in Draw Things, plus local LLM chat, plus video editing in Final Cut — the Mac would be the right call. That is not an AI-OFM agency running Forge pipelines for 20 personas with video on the roadmap.

Cost

Item	Mac Studio M5 Ultra 256 GB / 2 TB	Zestro build + RTX 5090
Sticker	$16,499 US ≈ Rs 4.6 M	Platform ≈ Rs 695k + 5090 Rs 1.05–1.55 M = Rs 1.75–2.25 M
Landed in Karachi	PakMac "coming soon" — expect Rs 5 M+ after duty and dealer margin	All local, all in stock (Zestro / Redtech / Czone)
Wait	16–18 weeks from Apple US; PakMac TBA	Days
Financing	Apple Card 7.99% APR — US residents only	Cash / bank transfer; UK Scan 48-mo if a relative buys the card
Second render node	Another Mac — Rs 4.6 M+	Your existing 3090 — Rs 0
Electricity (8 h/day at load, Rs 45/kWh est.)	~Rs 3,000 / month	~Rs 9,700 / month
5-year total (hardware + power, no upgrades)	≈ Rs 4.8 M	≈ Rs 2.8 M — and a 6090 swap in 2028 is optional, not required

The Mac costs 2× more, delivers ⅓ to ⅛ the throughput on your work, and can't be upgraded. Per unit of output that is roughly a 10× price difference. There is no reading of the numbers where it wins on value.

How to buy the 5090 build right

Which 5090

Air-cooled, local

The ROG Astral LC at $6,999 US ≈ Rs 1.96 M before shipping and duty — more than Karachi's Rs 1.55 M, with no warranty here. Buy locally: Redtech Rs 1,049,000 (confirm stock) · Zestro ASUS TUF Rs 1,550,000 · Czone MSI Gaming Trio. An air-cooled Astral / TUF / Suprim performs identically; the LC's 360 mm radiator would have to share the O11 Vision Compact with the A13 360 CPU AIO — possible, but a plumbing headache for zero speed.

Power

1250 W ATX 3.1

Already in the Zestro spec (MSI A1250GL). Native 12V-2×6 cable — no adapters. Your Riello 600 W UPS cannot carry this machine under load; keep it on the 3090 rig or set nvidia-smi -pl 400 when load-shedding is likely.

Memory

64 GB, 2×32

Matches what you have today. Never 32 GB. Confirm the kit is 2 sticks — dual channel.

Keep the 3090

Node #2

Standing decision. It keeps captioning, cropping and running Diana FAST while the 5090 does video. Two boxes, one automation.

Sequence 1) Platform from Zestro (board, CPU, RAM, PSU, case, AIO, SSD). 2) 5090 from whoever has stock at the lowest price that week — chase Redtech first. 3) Clone your Forge install; verify Diana FAST end-to-end on the new box before touching the old one. 4) Only then move Big Lust and video work over.