Claude Says ...

[Digger88]

I'm doing a 30 day countdown of my 30 favorite Morgan Wallen drinking songs leading up to his concert in Baltimore on July 18th and sync up videos I make to the part of the song I like obviously I don't sleep much and these 10 minute posts are fun which get about 5,000 views daily.

Grok makes the raw video which I mute in edits on IG. Claude keeps track of the list and GPT makes the video description and cover poster. Then I sync up the corresponding song part in IG and clip it to fit. Sort of like your MARS regression model lmao

https://twitter.com/x/status/2073322536092291478

 

[Digger88]

 

[Digger88]

Hey Bert do you use any preferences in Claude? If so what do your day or anyone else for that matter and do you think it helps

 

[RCA]

No sir lol my specific dilemma is generating AI video content locally and there is nothing out there that can beat Nvidia 50 series gpus in this regard in ComfyUI or stable diffusion. Standard video editing a Mac shines but in generative AI which are mathematical calculations a chip will never beat vram that's why AI data centers are nothing but Nvidia gpus. An Nvidia 5080 GPU is about 85% of a 5090 and this thing screams to make 3 second 720p videos that's why Elon bought 200,000 GB 100 GB 200's for Colossus which I believe he leases to Google and Anthropic as his next Data center comes online or is already recently online. It was a fun little project but unless I scale back greatly I am never going to approach what I can get from Grok in this regard as that and hardcore coding is about the only thing it is good at

ChatGPT:


"An RTX 5090 is usually much faster for many generative-AI workloads that fit inside 32 GB of VRAM and are well supported by CUDA. A Mac Studio M3 Ultra with 256 GB unified memory is often better for running much larger local models that simply will not fit on a single RTX 5090.
So the RTX 5090 is not “better” in an absolute sense. It is faster but smaller. The M3 Ultra is slower but much larger in usable memory.

The core comparison​

Issue	RTX 5090	Mac Studio M3 Ultra, 256 GB
GPU memory	32 GB GDDR7	Up to 256 GB unified memory
Memory bandwidth	About 1,792 GB/s, based on commonly reported RTX 5090 specs	819 GB/s, per Apple
AI software ecosystem	Very strong: CUDA, PyTorch, TensorRT, ComfyUI, Stable Diffusion, video tools	Improving: MLX, Metal, llama.cpp, Ollama, some PyTorch/MPS
Best at	Fast image generation, video generation, smaller/medium LLMs, training/fine-tuning	Large local LLM inference, huge context/model memory, private all-in-one workstation
Weakness	32 GB VRAM limit	Lower raw GPU AI throughput and weaker AI software ecosystem
Power	High; RTX 5090 cards are around 575 W class in many reviews/spec discussions	Much more power-efficient as a complete workstation

NVIDIA’s own RTX 5090 page describes it as a Blackwell GPU with 32 GB of GDDR7 memory and emphasizes fifth-generation Tensor Cores and AI/creator workloads. Apple’s Mac Studio specs list the M3 Ultra configuration with up to an 80-core GPU, 32-core Neural Engine, 819 GB/s memory bandwidth, and unified memory configurable to 256 GB.

Where the RTX 5090 really is better​

For many generative-AI tasks, the RTX 5090 will be plainly superior.
That includes:

1. Stable Diffusion / Flux / image generation
2. ComfyUI workflows
3. AI video generation
4. CUDA-based AI tools
5. Small-to-medium LLM inference
6. LoRA fine-tuning
7. Training or experimentation in PyTorch
8. Anything using NVIDIA-optimized kernels

The reason is not just the hardware. It is the CUDA ecosystem. Most serious AI software is built first for NVIDIA. Even when Apple support exists, the NVIDIA path is usually more mature, better documented, and faster.
For example, with a model that fits comfortably in 32 GB—say many 7B, 8B, 14B, 32B, and some quantized 70B-class workflows—the RTX 5090 is likely to generate faster, fine-tune more easily, and run with fewer software compromises.
So if someone says:

“For Stable Diffusion, Flux, ComfyUI, AI video, and CUDA development, the RTX 5090 is much better than the Mac Studio M3 Ultra.”

That is mostly true.

Where the Mac Studio M3 Ultra with 256 GB is better​

The Mac Studio’s great advantage is memory capacity.
A single RTX 5090 has only 32 GB of VRAM. That is a lot for consumer graphics, but not a lot for large language models. A 256 GB M3 Ultra can run models that a single 5090 simply cannot load entirely into GPU-accessible memory.
This matters for:

1. Large local LLMs
2. Large context windows
3. Multiple models loaded at once
4. Private document-analysis systems
5. Long-running local inference
6. Research, writing, coding, appraisal-analysis assistants
7. RAG systems with large embedding/index/context demands
8. Memory-heavy experiments where speed is less important than capacity

Apple’s MLX framework is specifically designed for Apple silicon and its unified memory architecture, and Apple describes it as optimized for efficient machine-learning research on that architecture. The MLX project also notes that arrays live in shared memory and can be operated on by supported device types without the usual explicit CPU/GPU transfer model.
That is the Mac’s real strength: the model can live in a very large shared memory pool.

The simple rule​

For generative AI, I would summarize it this way:
RTX 5090 wins when the model fits in 32 GB and speed matters.
M3 Ultra 256 GB wins when the model needs far more than 32 GB and local memory capacity matters.
That is the whole issue.

For local LLMs specifically​

This is the most nuanced case.
An RTX 5090 will usually beat the M3 Ultra on tokens-per-second for models that fit into VRAM. For example, smaller quantized models used for coding, chat, summarization, and agent workflows will usually run faster on the 5090.
But the RTX 5090 hits a wall when the model or context exceeds 32 GB. You can spill to system RAM, but performance usually drops sharply. You can add more GPUs, but then cost, heat, power, motherboard layout, driver issues, and multi-GPU software complexity enter the picture.
The Mac Studio does not have that same small VRAM wall. It has a much larger unified memory pool, although not all of it is equivalent to NVIDIA VRAM in performance. Still, for large local models, that can be decisive.
So for LLMs:

LLM use case	Likely better
7B–32B models, fast responses	RTX 5090
Quantized 70B models where speed matters	Usually RTX 5090, if it fits
Very large models exceeding 32 GB	Mac Studio M3 Ultra 256 GB
Huge local context	Often Mac Studio
Fine-tuning	RTX 5090
Private local assistant with large models	Mac Studio can be very attractive
Production AI development	Usually NVIDIA/Linux

For image and video generation​

Here the RTX 5090 is much more clearly favored.
Most image-generation and video-generation tooling is NVIDIA-first. Stable Diffusion, Flux, ComfyUI, AnimateDiff-style workflows, many upscalers, frame interpolation tools, and AI video pipelines generally run best on CUDA.
The Mac can do some of this, but the experience is usually less powerful, less flexible, and slower. Apple’s unified memory is not as useful here as it is for large LLMs, because many image/video models are more compute-bound and CUDA-optimized.
So if your primary interest is AI image/video generation, I would take the RTX 5090 system over the M3 Ultra.

For your likely use case​

Given your interests—appraisal analytics, USPAP compliance, MLS remarks, Prolog/R/Python workflows, local LLMs, report drafting, and possibly privacy-preserving document analysis—the Mac Studio M3 Ultra with 256 GB is not foolish at all.
In fact, for you, I would divide it this way:
Mac Studio M3 Ultra is attractive for:

• Running large local LLMs.
• Keeping appraisal data private.
• Document summarization.
• MLS remarks extraction.
• USPAP rule-checking assistants.
• R/Python/Prolog development.
• Quarto/report generation.
• General workstation use.
• Video editing with Apple’s media engines.

RTX 5090 workstation is attractive for:

• CUDA-based AI development.
• Stable Diffusion / Flux / ComfyUI.
• AI video generation.
• PyTorch experimentation.
• Fine-tuning.
• Maximum speed on models that fit in 32 GB.
• Better compatibility with open-source AI projects.

My opinion​

I would not say, “The RTX 5090 is much better at generative AI than the Mac Studio M3 Ultra 256 GB.”
I would say:
The RTX 5090 is much better at fast CUDA-based generative AI. The Mac Studio M3 Ultra 256 GB is better at large-memory local AI.
For a general AI hobbyist or researcher doing image/video generation and PyTorch work, the RTX 5090 is probably the better AI machine.
For a professional who wants a quiet, integrated Mac workstation that can run very large local language models and also do serious non-AI work, the M3 Ultra 256 GB is a very defensible choice.
For your situation, Bert, I would be reluctant to replace the Mac Studio with a 5090 PC. I would think of the 5090 PC as a specialized CUDA accelerator box, not as an outright replacement for the Mac Studio. The ideal setup may be: Mac Studio as the main workstation, plus a separate NVIDIA/Linux machine or cloud GPU for CUDA-heavy work."

 

[RCA]

Continued:

"For complex statistical work, I would generally favor the Mac Studio M3 Ultra with 256 GB RAM over a single RTX 5090 workstation, unless your statistical work specifically depends on CUDA/GPU-accelerated machine learning.

The reason is simple: most serious statistical work is constrained by memory, CPU throughput, storage speed, numerical libraries, reproducibility, and workflow stability more than by raw GPU tensor speed.

For classical and advanced statistics, the Mac Studio is probably better​

By “complex statistical work,” I mean things like:

• large regression systems;
• MARS models;
• GAMs;
• spatial statistics;
• kriging;
• bootstrapping and cross-validation;
• censored models;
• mixed models;
• Bayesian modeling;
• simulation;
• large data wrangling;
• report generation;
• R / Python / Quarto workflows;
• appraisal analytics;
• model diagnostics;
• reproducible research pipelines.

For that kind of work, 256 GB unified memory is extremely valuable. Many statistical tasks do not fit neatly into GPU pipelines. They involve irregular data structures, formula interfaces, model matrices, repeated fitting, sparse/dense matrix operations, and lots of data movement. A large-memory, fast CPU workstation is often more useful than a very fast GPU with only 32 GB of VRAM.

So for your appraisal/statistical work, the Mac Studio is a very strong machine.

Where the RTX 5090 helps statistical work​

The RTX 5090 becomes attractive when your “statistical work” is really closer to machine learning or deep learning, such as:

• neural networks;
• transformer models;
• PyTorch / TensorFlow training;
• deep learning image models;
• GPU-accelerated gradient boosting;
• large-scale embedding generation;
• vector-search pipelines;
• CUDA-accelerated Bayesian sampling;
• simulation code specifically written for GPU;
• model fine-tuning;
• AI video/image generation.

In those cases, the RTX 5090 can be much faster, provided the task fits inside its GPU memory.

But for ordinary R statistical modeling, an RTX 5090 often does surprisingly little. Most R packages do not automatically become dramatically faster because a powerful NVIDIA card is installed. You usually have to use GPU-aware libraries or rewrite parts of the workflow.

The key distinction​

The question is not really:

Which machine is better for AI?

It is:

Is my workload GPU-native?

For complex statistics, the answer is often no.

Many statistical workflows are:

• CPU-bound;
• RAM-bound;
• disk-I/O-bound;
• single-thread-limited in parts;
• dependent on BLAS/LAPACK;
• dependent on package implementation details;
• dominated by preprocessing and diagnostics rather than raw matrix multiplication.

A 5090 is magnificent when the workload is shaped like large tensor operations. But a lot of serious statistics is not shaped that way.

For R specifically​

For R, I would not buy an RTX 5090 expecting ordinary R code to become much faster.

R benefits more from:

• high RAM;
• fast SSD;
• strong single-core CPU speed;
• good multicore performance;
• optimized BLAS;
• careful parallelization;
• efficient data structures;
• using data.table, arrow, duckdb, Rcpp, or database-backed workflows where appropriate.

A 256 GB Mac Studio gives you room to hold large datasets, model matrices, cached objects, simulation results, and reports in memory. That is often more valuable than GPU acceleration.

For Python statistics and machine learning​

Python is more favorable to the RTX 5090 because the NVIDIA ecosystem is stronger there.

If you use:

• PyTorch;
• JAX;
• TensorFlow;
• RAPIDS;
• XGBoost GPU mode;
• LightGBM GPU mode;
• CuPy;
• Numba CUDA;

then an NVIDIA machine becomes much more compelling.

But if your Python work is mostly:

• pandas;
• NumPy;
• SciPy;
• statsmodels;
• scikit-learn;
• matplotlib;
• geospatial processing;
• report generation;

then the Mac Studio remains very competitive, and its memory capacity may matter more.

​

For your appraisal analytics​

For your particular kind of work, I would value the Mac Studio highly.

Your likely workload includes:

• MLS data cleaning;
• regression;
• MARS;
• KMeans;
• HPI adjustments;
• residual analysis;
• spatial modeling;
• text extraction from MLS remarks;
• report production;
• Quarto documents;
• R, Python, SQL, and Prolog;
• compliance logic;
• local LLM assistance.

That is not primarily a CUDA workload. It is a large-memory professional analytics workload.

For that, the Mac Studio M3 Ultra with 256 GB would be excellent.

The RTX 5090 would become useful as an auxiliary system if you start doing heavy:

• local LLM fine-tuning;
• neural-network training;
• image/video generation;
• CUDA-based document AI;
• very high-speed embedding generation;
• GPU-native machine learning.

My opinion​

For someone doing complex statistical work, I would not say the RTX 5090 is “much better.” I would say the opposite in many cases:

A Mac Studio M3 Ultra with 256 GB RAM is probably the better primary workstation for serious statistical computing, especially in R, mixed R/Python workflows, reproducible research, and large-memory analytics.

The RTX 5090 is better as a special-purpose accelerator for CUDA-native AI and deep learning.

For you, Bert, the most sensible arrangement would be:

Mac Studio M3 Ultra 256 GB as the main statistical workstation.

Then, only add an RTX 5090 Linux/Windows box if you clearly need CUDA for deep learning, image/video generation, model fine-tuning, or GPU-native AI experiments."

 

[Digger88]

good read and it shows why each use is perfect for each of us.

My only rebuttal for your suggestion I consider a Mac is, but can it run Crysis?

and given this is a hilariously old joke about PC gaming I feel I owe you an explanation lol



My PC is a best for gaming which I like to do from time to time but I'm also a graphics snob. AI and the fun I get came much later. Some gamers love to play natively I otoh want every bell and whistle, frame gen ray tracing and upscaling DLSS has to offer. My digital library is 300+ deep so the RTX offers me the best of both worlds

 

[RCA]

good read and it shows why each use is perfect for each of us.

My only rebuttal for your suggestion I consider a Mac is, but can it run Crysis?

and given this is a hilariously old joke about PC gaming I feel I owe you an explanation lol

View attachment 110250

My PC is a best for gaming which I like to do from time to time but I'm also a graphics snob. AI and the fun I get came much later. Some gamers love to play natively I otoh want every bell and whistle, frame gen ray tracing and upscaling DLSS has to offer. My digital library is 300+ deep so the RTX offers me the best of both worlds

The issue here, implicitly, is the advice WE are giving to someone who is deciding what to buy. You are very focused on your own use, which is, it appears to me, very specialized and has little to do with appraisal. Clearly, a Mac can handle generative AI, but it is not as fast or as easy to use as an Nvidia board on a PC. For the typical appraiser, that would likely be a NON-ISSUE. In my case, at present, that is certainly the case. When they day comes that I start doing a lot of neural network processing, I may want to consider upgrading to something else - but I will most likely use the cloud, e.g RunPod.com, before spending $15K on an Nvidia RTX 6000. Consider I have Terabytes of SF Bay Area photos on my Synology. I don't know - maybe I might buy a 5090X, - but then again I could rent that on the cloud as well.

To spend a lot of money on a low-memory special-use piece of hardware that will soon be out of date, superseded by other hardware - should be a hard decision to make with the cloud as an alternative - or something like an Apple Mac Studio M3 Ultra (which, however, is currrently only available with 96G).

For your special case, you know best. But not many appraisers will fall into your special case.