What if I had bought an RTX 5090 PC instead of a ₩8.6M Mac Studio?

₩8.6M — should I have bought an RTX PC?

I ordered a Mac Studio M5 Max with 128GB memory and a 1TB SSD. An education-discount route brought the payment to roughly ₩8.6M including AppleCare. AppleCare accounts for about ₩0.2M; the regular price for the same package was roughly ₩9.4M.

After paying, I started wondering about the alternatives. Would an RTX PC built with that money run local AI faster? I want to try big models and make images and video too. Was this a good purchase…?

But I am replacing more than an LLM server. I want Xcode, Simulator, several app projects, Hermes and local models in one development environment. So I looked at what each system can hold and what I can do on one machine, alongside speed.

What systems could that money build?

A GPU alone and a complete Mac are different purchases. A PC also needs a CPU, RAM, storage, power and cooling. My earlier examples added a GPU to roughly ₩2.79M in other parts. Instead of repeating part names, here are the amounts and what they mean.

Configuration	Recorded amount	How to read it
My Mac Studio M5 Max 128GB / 1TB	About ₩8.6M	Approximate actual payment · education discount + AppleCare
RTX 5090 32GB PC · 64GB system RAM / 1TB SSD	About ₩11.38M	New-parts sum, not a confirmed final purchase quote
Same PC with a different 5090 listing	About ₩7.08M	Hypothetical sum using a sold-out GPU price, not purchasable
RTX 5080 16GB PC	About ₩5.37M	New-parts example · final quote unconfirmed
RTX 3090 24GB PC	About ₩3.84M	Used-GPU asking price + other new parts · transaction/warranty unconfirmed
DGX Spark 128GB / 4TB	Domestic listing about ₩10.35M / import listing about ₩8.99M	Stock/final purchase price unverified · import excludes duties/VAT

Prices recorded on 2026-10-10. PC examples assume Linux and self-assembly, without a final quote including paid OS, assembly or shipping. Spark has 4TB; my Mac has 1TB and AppleCare. These are different packages. Sources are at the end.

The premium ₩16M-class 5090 prebuilt I initially found does not represent typical PC pricing. A sold-out card also cannot establish that ₩8.6M buys a working build today. I have not secured a purchasable complete 5090 quote within my budget. A 5080 or used 3090 could mean spending less, but the next question is whether the models fit their VRAM.

128GB unified memory vs 32GB VRAM

Mac and Spark share 128GB between CPU and GPU. A 5090 has 32GB dedicated VRAM, with system RAM separate. Adding PC RAM does not enlarge GPU VRAM.

Capacity is room to hold work; bandwidth is how quickly data can move. Separating them helps explain Spark’s appeal and why the 5090 is interesting for fast, smaller models.

Manufacturer specification	Memory structure	Memory bandwidth (GB/s)
Mac Studio M5 Max · 40-core GPU	128GB unified-memory configuration	614
RTX 5090	32GB dedicated VRAM	1,792
DGX Spark	128GB unified system memory	273

Repeated weight reads can matter during small-batch Decode. Bandwidth ratios are not token-speed ratios: architecture, kernels, caches and compute matter too. The Prefill/Decode distinction in EP01 returns here.

What about the models I want to load?

Model / workload	Mac Studio 128GB	RTX 5090 32GB PC	DGX Spark 128GB
20–30B Q4 LLM	Likely capacity fit · count development tools too	Likely capacity fit · check headroom for long context	Likely capacity fit · check runtime
70–100GB of weights alone	Check OS, KV and buffer headroom	Exceeds single VRAM · offload/partitioning and enough system RAM needed	Check OS, KV and buffers · verify Arm runtime
Large MoE	Total expert weights determine resident capacity	Small active parameter count does not ensure full weights fit VRAM	Check total weights and runtime support
Long context / multiple sessions	KV and workspace compete for shared memory	KV and buffers consume VRAM	KV and workspace compete for shared memory

20–30B at four bits gives roughly 10–15GB from weight-bit arithmetic alone. Actual files include quantization metadata; execution adds KV cache and buffers. This is not a measured requirement for a specific model. Count the whole budget, as in EP02. Fitting in memory and running at a useful speed are separate questions.

A small active parameter count in MoE does not mean all experts occupy little memory. Two 5090s provide 64GB in aggregate, but not automatically one contiguous pool. Without NVLink, a supported runtime must partition the model and communicate between GPUs. Using EP04’s expert streaming adds SSD read costs to the capacity discussion.

How fast is it? — Prefill and Decode

Capacity does not tell me how long I will wait. I brought back the published llama.cpp results: pp512 processes 512 input tokens; tg128 generates 128 tokens.

Published LLM benchmarks · two stages, small model

Llama 2 7B Q4_0 · GPU offload 99. Independent community submissions, not controlled for hosts, drivers, runtime, file hashes or power. NVIDIA: CUDA / FA off; Mac: Metal / FA setting unconfirmed. pp512: 512 input tokens; tg128: 128 generated tokens. Separate zero-origin scales.

Speed references for a small model that fits memory. Not a value ranking accounting for whole-system cost, large-model capacity, Xcode or energy.

Prefill · pp512

0 — 15,000 tokens/s

• RTX 5090
  14,073.41
• RTX 4090
  11,992.70
• RTX 5080
  8,297.36
• RTX 3090
  5,174.69
• RTX A6000
  4,913.93
• M5 Max · MacBook Pro
  3,347.44
• DGX Spark · 128GB
  3,062.31

Decode · tg128

0 — 300 tokens/s

• RTX 5090
  290.02
• RTX 4090
  186.21
• RTX 5080
  181.99
• RTX 3090
  158.16
• RTX A6000
  138.73
• M5 Max · MacBook Pro
  119.08
• DGX Spark · 128GB
  57.21

Mac reference: 16-inch MacBook Pro, M5 Max / 40 GPU cores / 128GB, submitted 2026-06-22. Not my Mac Studio. Cooling, power, runtime and file hashes are unmatched; the shared chip does not establish Studio speed.

Llama 2 7B Q4_0 · Values, reported ± and sources · scroll horizontally

Measured device / backend	Prefill pp512 tok/s	Decode tg128 tok/s	llama.cpp Commit · source
RTX 5090 CUDA · FA off	14073.41 ± 115.16	290.02 ± 1.10	8cf6b42 @totaldev
RTX 4090 CUDA · FA off	11992.70 ± 107.99	186.21 ± 0.13	2241453 @lhl
RTX 5080 CUDA · FA off	8297.36 ± 9.50	181.99 ± 0.42	8a4280c @Hedede
RTX 3090 CUDA · FA off	5174.69 ± 21.83	158.16 ± 0.21	c76b420 @m18coppola
RTX A6000 CUDA · FA off	4913.93 ± 6.79	138.73 ± 2.75	4795c91 @Hedede
M5 Max · MacBook Pro 16″ · 40 GPU cores · 128GB · Metal/BLAS · 6 threads · FA ?	3347.44 ± 1.53	119.08 ± 0.69	52b3df0 @michaelasper
DGX Spark · 128GB CUDA · FA off	3062.31 ± 11.02	57.21 ± 0.06	5acd455 @ggerganov

Source and runtime commits

llama.cpp · CUDA scoreboard (no FA) · 2026-10-10

llama.cpp · Apple Silicon · @michaelasper · 2026-06-22 · Checked 2026-10-10

• RTX 5090 · 8cf6b42 · @totaldev
• RTX 4090 · 2241453 · @lhl
• RTX 5080 · 8a4280c · @Hedede
• RTX 3090 · c76b420 · @m18coppola
• RTX A6000 · 4795c91 · @Hedede
• M5 Max · MacBook Pro · 52b3df0 · @michaelasper
• DGX Spark · 128GB · 5acd455 · @ggerganov

Curated newer/stronger submissions introduce selection bias. A6000 here means Ampere.

For this small model that fits memory, the 5090’s throughput stands out. The 4090, 5080 and 3090 references also give reasons to consider less expensive cards. Spark has 128GB, yet that capacity does not turn into proportionally faster small-model Decode. Holding a large model and running a small model quickly are different strengths.

I also lack matched three-system results for large MoE or 70–100GB weights, so I have not filled those gaps with estimated speeds.

Long-input Prefill, Decode and time to first token are different metrics. EP03’s acceleration changes the balance again through verification cost and accepted length. Once the machine arrives, I want to start with short and long answers using my own models.

Which system fits my work?

I move between Cookly, KnitLens, Delsa, Sobok and Sound Journey. On my M1 iMac with 8GB RAM and 256GB storage, IDEs, Xcode and Simulator make both resources tight. A Samsung T7 1TB helps with storage, but putting simulators on it has been slow and inconvenient. That is where EP00 started.

An RTX PC or Spark can take the AI workload. Xcode and Simulator would still run on the Mac, though. Buying a GPU server does not upgrade the 8GB iMac’s development environment. Windows/Linux PCs and Spark cannot directly replace Xcode, so my comparison becomes one Mac Studio versus a development Mac plus an AI server.

CUDA is appealing for image/video generation and LoRA. ComfyUI also supports Apple Silicon, but the specific models, custom nodes and training backends need checking. Spark’s CUDA and large shared memory likewise do not guarantee every RTX workflow works unchanged on Arm Linux.

My plan is to use Hermes and a medium LLM frequently, then switch to a larger model or vision, speech or image models when needed. I do not intend to keep everything resident. Changing budget, purpose and model size helps expose the trade-offs.

Where would I start for my workload?

A workload checklist, not a performance predictor or purchase ranking. Prices checked 2026-10-10. Sold-out or unverified quotes are not purchasable candidates.

BudgetUp to ₩5M₩5M–₩10M₩10M and aboveMain workloadLarge LLMs and long contextsImage/video generation and trainingApp development + multiple AI modelsLocal LLM size20–30B Q4 · medium LLM70–100GB weights · large LLMModel / memory requirement undecided

Configuration to investigate

Compare Mac Studio with a Mac plus a separate GPU server

Advantage

A Mac can host Xcode and local AI in the same development environment.

Constraint

CUDA-only jobs may need another environment. A PC alone cannot replace Xcode. For 20–30B Q4, roughly 10–15GB is only weight-bit arithmetic. Count actual files, KV and buffers; check headroom particularly on 16GB GPUs.

Check next

Check memory with development tools, existing Mac condition and two-machine costs.

Price and stock status

My Mac was about ₩8.6M; new-5080 parts about ₩5.37M, opened/unused-4090 parts about ₩9.18M. Spark import listing: ₩8.99M before import duties/VAT, final purchase price unverified. Eligibility, stock and final quotes need checking. The sold-out-GPU ₩7.08M 5090 is excluded from purchasable candidates.

This is not a quotation tool that automatically picks purchasable complete systems. I am starting with the development environment plus memory capacity, then revisiting a Mac plus GPU server if CUDA work becomes important. OpenCode Go and Codex will remain part of the setup; one local model does not need to carry my entire budget.

Why I chose the Mac Studio

I did not buy it after proving it was the fastest LLM machine. I wanted a better app-development environment and enough memory to experiment with local AI together. It is an attempt to fix two recurring frustrations on my 8GB iMac.

If a CUDA-only video workflow or KnitLens training hits a wall on Mac, an NVIDIA server or cloud service may fit better. For development tools beside agents and models, with occasional large-model experiments, macOS plus a large shared space appeals to me. Spark remains an interesting large-model server candidate, but I would still need a separate development Mac.

I also want to measure idle power for always-on Hermes and energy per completed workload. GPU TGP or adapter capacity is not whole-system electricity use. Actual power and usage time are needed before operating-cost conclusions.

I will judge the ₩8.6M decision through development waits, model switching and the image/video workflows I actually use. After delivery, I want to return with “Was the ₩8.6M choice worth it?”

First, the machine has to arrive.

References

Checked October 10, 2026. Mac prices are approximate amounts from my purchase. Retail amounts refer to particular observed configurations, not market minimums/averages or confirmed purchase quotes.

• Specifications/development: Apple Mac Studio, RTX 5090, DGX Spark, Xcode requirements.

• Published performance/execution: llama.cpp CUDA submissions, Apple Silicon submissions, ComfyUI, MLX, Spark Arm porting guide, multi-GPU inference.

• Pricing: 5090 card, sold-out conditional card example, specific premium prebuilt, 5080 card, 3090 used asking price, Spark Korean listing, Spark import listing. PC totals add the roughly ₩2.79M non-GPU parts observed at the time.

• Memory bandwidth: Apple Studio specifications, NVIDIA Blackwell architecture, DGX Spark specifications. Checked 2026-10-10.