{"id":3639,"date":"2026-11-17T13:29:50","date_gmt":"2026-11-17T20:29:50","guid":{"rendered":"https:\/\/yyc.ninja\/?p=3639"},"modified":"2026-11-17T13:29:50","modified_gmt":"2026-11-17T20:29:50","slug":"45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm","status":"publish","type":"post","link":"https:\/\/yyc.ninja\/index.php\/2026\/11\/17\/45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm\/","title":{"rendered":"A $45 GPU vs an M1 Pro, Mac Studio and RTX 5090: Local LLM Speed, Price and Power Compared"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1200\" height=\"630\" src=\"https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover.png\" alt=\"A $45 GPU vs an M1 Pro, Mac Studio and RTX 5090: Local LLM Speed, Price and Power Compared\" class=\"wp-image-3638\" sizes=\"auto, (max-width: 1200px) 100vw, 1200px\" srcset=\"https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover.png 1200w, https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover-300x158.png 300w, https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover-1024x538.png 1024w, https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover-768x403.png 768w, https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover-100x53.png 100w\" \/><\/figure>\n<h2 class=\"wp-block-heading\">A $45 GPU vs an M1 Pro, Mac Studio and RTX 5090: Local LLM Speed, Price and Power Compared<\/h2>\n<p class=\"wp-block-paragraph\">I paid $45 CAD for a used 8 GB RX 580. In an old HP Z640 workstation it runs Qwen3.6-35B-A3B, a 35-billion-parameter coding model, at 21 to 23 tokens per second on code. My 32 GB M1 Pro laptop runs the same model at 34 to 37. A Mac Studio with an M5 Max reaches about 130, and an RTX 5090 reaches 183 to 271.<\/p>\n<p class=\"wp-block-paragraph\">It is tempting to divide those speeds by the price of the card and declare the $45 GPU the winner. I did that in my first draft. It was wrong, because a GPU alone runs nothing. Once you price a complete computer and meter the power, the gap closes a lot. The old workstation is still the cheapest way in, but it is only level on speed per dollar, and it costs far more in electricity.<\/p>\n<h3 class=\"wp-block-heading\">Price, speed and power, side by side<\/h3>\n<p class=\"wp-block-paragraph\">All prices are Canadian dollars before tax. The RX 580 and M1 Pro figures are mine, from real coding use through opencode. The others are community or published benchmarks, mostly short runs, so treat them as best cases.<\/p>\n<table>\n<thead>\n<tr>\n<th>System<\/th>\n<th>Price<\/th>\n<th>35B decode (tok\/s)<\/th>\n<th>Prompt reading (tok\/s)<\/th>\n<th>Power under load<\/th>\n<th>Energy per token<\/th>\n<th>Electricity per million tokens<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>HP Z640 + RX 580, used parts<\/td>\n<td>about C$700 to C$900<\/td>\n<td>21 to 23<\/td>\n<td>130 to 190<\/td>\n<td>235 W on sensors<\/td>\n<td>10 to 11 J<\/td>\n<td>34 to 37 cents<\/td>\n<\/tr>\n<tr>\n<td>MacBook Pro M1 Pro 32 GB, refurbished<\/td>\n<td>about C$1,418<\/td>\n<td>34 to 37<\/td>\n<td>264 to 308<\/td>\n<td>about 50 W (estimate)<\/td>\n<td>about 1.4 J<\/td>\n<td>about 5 cents<\/td>\n<\/tr>\n<tr>\n<td>Mac Studio M5 Max, 40-core GPU<\/td>\n<td>C$4,399<\/td>\n<td>125 to 133<\/td>\n<td>about 3,700<\/td>\n<td>up to 200 W<\/td>\n<td>at most 1.6 J<\/td>\n<td>at most 5 cents<\/td>\n<\/tr>\n<tr>\n<td>RTX 5090, card only<\/td>\n<td>C$5,999 median, C$8,499 lowest today<\/td>\n<td>183 to 271<\/td>\n<td>2,892 to 9,240<\/td>\n<td>about 520 W for the system<\/td>\n<td>1.9 to 2.8 J<\/td>\n<td>6 to 9 cents<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p class=\"wp-block-paragraph\">Sources: <a href=\"https:\/\/omlx.ai\/benchmarks\/performance\/e58uc38u\">M1 Pro on oMLX<\/a>, <a href=\"https:\/\/omlx.ai\/benchmarks\/performance\/9e4yi6u4\">M5 Max on oMLX<\/a>, <a href=\"https:\/\/support.apple.com\/en-gb\/102027\">Apple&#8217;s Mac Studio power figures<\/a>, <a href=\"https:\/\/zenn.dev\/toki_mwc\/articles\/rtx5090-qwen36-35b-a3b-llmacpp-bench?locale=en\">183 tok\/s on a 5090<\/a>, <a href=\"https:\/\/huggingface.co\/datasets\/omegaprime669\/rtx-5090-benchmarks\">271 tok\/s on a 5090<\/a>, <a href=\"https:\/\/markaicode.com\/benchmarks\/rtx-5090-llama-benchmark\/\">5090 power<\/a>, <a href=\"https:\/\/gpuprix.com\/gpus\/geforce-rtx-5090\">5090 Canadian pricing<\/a>, <a href=\"https:\/\/kingy.ai\/blog\/m5-mac-studio-local-ai-m5-max-vs-m5-ultra\/\">Mac Studio pricing<\/a>, <a href=\"https:\/\/www.refurb.me\/ca\/stats\/macbook-pro\/macbook-pro-14-inch-m1-pro-late-2021\">M1 Pro refurbished price<\/a>.<\/p>\n<p class=\"wp-block-paragraph\">Some care is needed with those rows:<\/p>\n<ul class=\"wp-block-list\">\n<li>The Z640 price is what a dual-Xeon machine, RAM, an SSD and the card can cost in used parts when you shop around. It is not one current listing. My speeds also depend on all four memory channels of each CPU being filled, so a build with fewer channels populated will be slower.<\/li>\n<li>The M1 Pro&#8217;s 34 to 37 is from my own runs in May. Community runs show 48 at short context, falling to 31 at 32K. Its prompt speed is from those community runs; my own notes record 1,380, which I can no longer verify, so I use the lower figure.<\/li>\n<li>The M5 Max figures are at 1K context, so they will fall with long prompts. An independent 6-bit run reached 107 at 8K.<\/li>\n<li>Apple&#8217;s 200 W is the most it measured on a 32-core M5 Max under a stress test. LLM load is probably lower.<\/li>\n<li>The 5090&#8217;s 520 W comes from one run on a different model.<\/li>\n<\/ul>\n<h3 class=\"wp-block-heading\">Who is actually closest?<\/h3>\n<p class=\"wp-block-paragraph\">The closest Apple machine is an M1 Pro with 32 GB, which is a five-year-old laptop. On a 9B dense model it is about level with my card, at around 20 tokens per second against my 22. On the 35B it pulls ahead by about 1.6 times.<\/p>\n<p class=\"wp-block-paragraph\">The reason is where the model lives. My card holds only part of it, so 21 GB sits in DDR4 and the CPU reads it from there. The M1 Pro keeps everything in unified memory at about 200 GB\/s. That matters even more for reading prompts, and a 25,000-token prompt takes me 187 seconds against about 95 on the M1 Pro, computed from community figures. The M1 Pro can also run a dense 27B at 8 to 10 tokens per second, which my 8 GB card cannot hold at all.<\/p>\n<p class=\"wp-block-paragraph\">You do need the 32 GB. The 35B peaks at 19 to 23 GB of memory, so a 16 GB Mac is not a cheaper match, whatever its price.<\/p>\n<h3 class=\"wp-block-heading\">The per-dollar answer<\/h3>\n<p class=\"wp-block-paragraph\">Here is speed per C$100 for a complete computer, on this model:<\/p>\n<table>\n<thead>\n<tr>\n<th>System<\/th>\n<th>tok\/s per C$100<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Z640 + RX 580<\/td>\n<td>2.3 to 3.3<\/td>\n<\/tr>\n<tr>\n<td>M1 Pro 32 GB<\/td>\n<td>2.4 to 2.6<\/td>\n<\/tr>\n<tr>\n<td>Mac Studio M5 Max<\/td>\n<td>2.9 to 3.0<\/td>\n<\/tr>\n<tr>\n<td>RTX 5090, card only, host not included<\/td>\n<td>2.2 to 4.5<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p class=\"wp-block-paragraph\">They are all within about a factor of two, and the Z640 sits level with the M1 Pro and the Mac Studio. The market has priced this fairly well, and no option is a steal. My card looks like 49 tok\/s per C$100 only if you pretend the rest of the computer was free. For me it nearly was, because I already owned the workstation, so a cheap card is close to free speed. Even bought whole, the Z640 is the lowest entry price here, about C$500 to C$700 below the M1 Pro.<\/p>\n<h3 class=\"wp-block-heading\">What the $45 does not buy<\/h3>\n<p class=\"wp-block-paragraph\">The card is not doing the job alone. The 35B&#8217;s file is 22 GB and my GPU has 8. llama.cpp keeps the attention layers on the GPU and all 21 GB of experts in system RAM, where the CPU reads them. That only works because the host is an HP Z640 with two Xeons and 144 GB of DDR4 across eight channels. Those experts are read at about 54 GB\/s, and that bandwidth sets the speed limit, not the GPU.<\/p>\n<p class=\"wp-block-paragraph\">A cheap card has odd limits too. Mine exposes only 256 MB of its VRAM to the CPU. For weeks, any model loaded after the GPU had sat idle ran about 18% slower until it was reloaded, because 209 MB of its buffers landed in system memory instead of on the card. Old platforms come with old-platform problems. That one turned out to have a one-line fix, below.<\/p>\n<p class=\"wp-block-paragraph\">I also tested the lore about memory channels. The standard advice is to fill every channel with identical sticks, and my Z640 had one 32 GB stick sitting beside three 16 GB ones, which put a slow section of memory above the 64 GB mark. I swapped it for a 16 GB stick, so all four channels are now equal, and measured again: no change at all in speed, 21 tokens per second before and after. The slow section had held about 11% of the model, and the model was not waiting on memory bandwidth to begin with, so removing it bought nothing measurable. Filling every channel still matters when you build, because an empty channel is a much bigger loss than a mismatched stick.<\/p>\n<p class=\"wp-block-paragraph\">So $45 is the price of the card, not the price of the speed. A 5090 needs a host too, but a much less special one, because the whole model sits in its 32 GB of VRAM.<\/p>\n<p class=\"wp-block-paragraph\">Two more things the RX 580 cannot buy at any price:<\/p>\n<ul class=\"wp-block-list\">\n<li><strong>Dense models above about 8 GB.<\/strong> A dense 27B needs about 17 GB. A 5090 runs one at 77 tok\/s. Mine cannot run it at all.<\/li>\n<li><strong>Fast prompt reading.<\/strong> This is the largest gap of all, and decode speed hides it.<\/li>\n<\/ul>\n<h3 class=\"wp-block-heading\">Did I need a modded BIOS to fix it?<\/h3>\n<p class=\"wp-block-paragraph\">No. The fix was one environment variable, but I nearly went the long way first.<\/p>\n<p class=\"wp-block-paragraph\">The usual suggestion for that kind of slowdown is Resizable BAR, which lets the CPU see all of the card&#8217;s VRAM instead of a 256 MB slice. You will see it described as a way to enable full PCIe bandwidth. It does not do that. My link already runs at full speed, PCIe 3.0 x16. Resizable BAR changes how much of the card the CPU can reach, not how fast the link is.<\/p>\n<p class=\"wp-block-paragraph\">My HP Z640 cannot give the card a bigger window anyway. It hands the card 384 MB of address space, all below the 4 GB mark, and an 8 GB window cannot fit there. I tried the stock BIOS option, PCIE MMIO Assignment Mode set to 64-bit: it opened a 2 GB window above 4 GB, the card moved into it, and it still got only 256 MB, because the driver asks for all 8 GB or nothing. The other option was a <a href=\"https:\/\/github.com\/bibikalka1\/HP_Z440_Z640_Z840_BIOS_mod\">community-modded BIOS<\/a> with no success report for an RX 580 or on Linux, and one <a href=\"https:\/\/github.com\/bibikalka1\/HP_Z440_Z640_Z840_BIOS_mod\/issues\/14\">Z640 owner&#8217;s flash<\/a> that left the machine stuck in a power cycle.<\/p>\n<p class=\"wp-block-paragraph\">So I looked at what was actually going wrong. After an idle spell, that 256 MB slice was already full when the model loaded, and llama.cpp&#8217;s Vulkan backend put its buffers in system memory instead, reached over PCIe at about a twentieth of the card&#8217;s own speed. The backend has a switch that tells it never to use that slice at all:<\/p>\n<pre><code class=\"language-ini\"># \/etc\/systemd\/system\/llama-gpu-router.service.d\/vk-vidmem.conf\n[Service]\nEnvironment=GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM=1\n<\/code><\/pre>\n<p class=\"wp-block-paragraph\">With it, the buffers go into the card&#8217;s normal memory every time. Loaded after an idle GPU, the 35B went from 17 to 18 tokens per second on code to 21.5 to 22, in four tries out of four. Every other model I run got faster or stayed level:<\/p>\n<table>\n<thead>\n<tr>\n<th>Model<\/th>\n<th>Before<\/th>\n<th>After<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Qwen3.6 35B-A3B, writing<\/td>\n<td>19.2 tok\/s<\/td>\n<td>23.8 tok\/s<\/td>\n<\/tr>\n<tr>\n<td>Qwen3.5 9B, writing<\/td>\n<td>18.2 tok\/s<\/td>\n<td>21.8 tok\/s<\/td>\n<\/tr>\n<tr>\n<td>Qwen3.5 4B, writing<\/td>\n<td>25.2 tok\/s<\/td>\n<td>31.3 tok\/s<\/td>\n<\/tr>\n<tr>\n<td>Gemma E4B, reading a 20K-token prompt<\/td>\n<td>182 tok\/s<\/td>\n<td>252 tok\/s<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p class=\"wp-block-paragraph\">Those are single runs, but the pattern was the same on every model, and nothing crashed. The cost is a few hundred MB more VRAM per model. The warm-up trick I had been trying, a short GPU job just before loading, worked only two times out of three.<\/p>\n<p class=\"wp-block-paragraph\">The wider lesson is simple: find out exactly what is blocking you before you flash anything, because the fix may be one setting. Here it was.<\/p>\n<h3 class=\"wp-block-heading\">Where the real gap is<\/h3>\n<p class=\"wp-block-paragraph\">Decode is what people quote. Prefill, the time spent reading the prompt, is what you feel with a coding agent, because an agent sends tens of thousands of tokens before the model writes a word.<\/p>\n<table>\n<thead>\n<tr>\n<th>Task<\/th>\n<th>RX 580<\/th>\n<th>RTX 5090<\/th>\n<th>Gap<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Writing the answer<\/td>\n<td>21 to 23 tok\/s on code<\/td>\n<td>183 to 271 tok\/s<\/td>\n<td>8 to 13x<\/td>\n<\/tr>\n<tr>\n<td>Reading the prompt<\/td>\n<td>130 to 190 tok\/s<\/td>\n<td>2,892 to 9,240 tok\/s<\/td>\n<td>15 to 70x<\/td>\n<\/tr>\n<tr>\n<td>First request of a 25K-token agent session<\/td>\n<td>187 s<\/td>\n<td>3 to 9 s (computed)<\/td>\n<td>22 to 58x<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p class=\"wp-block-paragraph\">A real opencode first prompt on my setup is 24.8K tokens, and even a one-word reply took 210 seconds cold. A prompt cache removes most of that cost after the first request, and later turns take 17 to 27 seconds. I also save the prompt state to disk after each long answer. On a real 21K-token opencode session, the next turn after a model swap took 30 seconds with the restore and 180 seconds without it. That is software narrowing the gap. It does not close it.<\/p>\n<h3 class=\"wp-block-heading\">What it feels like with a real coding agent<\/h3>\n<p class=\"wp-block-paragraph\">I ran opencode against the 35B on five small repositories, each with bugs across several files or behavior that has to be worked out from a written spec. Hidden tests did the grading, and each attempt had a 30-minute limit. Twenty attempts, half with thinking on and half with it off, took 5.1 hours through a real agent, not a raw API call. Finished attempts took 4 to 21 minutes, and five of the twenty hit the limit without finishing.<\/p>\n<p class=\"wp-block-paragraph\">Waiting is only part of the story. Scores swung wildly between attempts at the same task with the same settings: one scored 9 of 9 and another 1 of 9. Thinking did not settle it either: 58 of 82 checks with it off, 61 with a small budget, 52 with a larger one, all inside the noise. The smaller 9B model scored only 24 to 26 of 82 on the same tasks, so a faster small model is no shortcut here. The weak attempts I read still reported that all tests passed, because the agent ran only the one easy test in the repository. On a 5090 you would get the same model and the same mistakes, only sooner. A fast card gives you more attempts per hour, and with results this variable, that has real value.<\/p>\n<h3 class=\"wp-block-heading\">The model choice matters more than the card<\/h3>\n<p class=\"wp-block-paragraph\">After those first runs I added two more models to the same bench: another 35B called Ornith-1.5, and an 8B called LFM2.5 that fits entirely on the card. The results change the advice in this post more than any hardware change did.<\/p>\n<table>\n<thead>\n<tr>\n<th>Model<\/th>\n<th>Decode<\/th>\n<th>Thinking off, checks passed<\/th>\n<th>Thinking on, checks passed<\/th>\n<th>Attempts passing every check<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Ornith-1.5 35B<\/td>\n<td>18 tok\/s<\/td>\n<td>67 of 82<\/td>\n<td><strong>80 of 82<\/strong><\/td>\n<td>8 of 10<\/td>\n<\/tr>\n<tr>\n<td>Qwen3.6 35B<\/td>\n<td>18 tok\/s<\/td>\n<td>58 of 82<\/td>\n<td>52 of 82<\/td>\n<td>4 of 10 at best<\/td>\n<\/tr>\n<tr>\n<td>Qwen3.5 9B<\/td>\n<td>22 tok\/s<\/td>\n<td>24 of 82<\/td>\n<td>26 of 82<\/td>\n<td>0 of 10<\/td>\n<\/tr>\n<tr>\n<td>Gemma E4B<\/td>\n<td>25 tok\/s<\/td>\n<td>14 of 66<\/td>\n<td>18 of 74<\/td>\n<td>0 of 10<\/td>\n<\/tr>\n<tr>\n<td>LFM2.5 8B<\/td>\n<td><strong>76 tok\/s<\/strong><\/td>\n<td>10 of 74<\/td>\n<td>always thinks<\/td>\n<td>0 of 10<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p class=\"wp-block-paragraph\">Two things stand out.<\/p>\n<p class=\"wp-block-paragraph\">The fastest model is the worst coder. LFM2.5 runs at 76 tokens per second, four times the 35B, loads in 12 to 33 seconds, and needs no CPU for its weights because it fits on the card. It also scored 10 of 74, failing every task in about the same way and never once timing out. It is the right model for quick drafts and for turning speech into text. It is not an agent.<\/p>\n<p class=\"wp-block-paragraph\">The second 35B is the best coder by a wide margin, and thinking is what does it. Ornith-1.5 with thinking on passed 80 of 82 checks and finished 8 of 10 attempts perfectly, against 52 of 82 for the Qwen model with the same setting. Its decode speed is identical, 18 tokens per second, so the gain is not throughput. The same model with thinking off scores 67 of 82, which says the reasoning is worth 13 points here and nothing for the other model. The cost is time: a median 869 seconds per attempt against 489, and 2 of 10 attempts ran out the 30-minute clock.<\/p>\n<p class=\"wp-block-paragraph\">So on a $45 card, choosing the right model was worth 22 points of coding accuracy, and buying a card 10 times faster would have been worth 0 points of accuracy and 10 times more attempts. If you are doing this at all, try the models before you spend on the hardware.<\/p>\n<h3 class=\"wp-block-heading\">What a long idle actually costs<\/h3>\n<p class=\"wp-block-paragraph\">One last thing I could not answer from short tests: what a real overnight idle does. I blocked every port into the machine for 10 hours, logged the memory every 10 minutes, and measured again in the morning.<\/p>\n<p class=\"wp-block-paragraph\">Nothing was lost. The model&#8217;s 21 GB stayed locked in RAM on the right CPU, video memory did not change, and there were no GPU resets. From the second request of the morning the speed matched the evening exactly.<\/p>\n<p class=\"wp-block-paragraph\">The first request paid about 14 seconds more before its first token, 18.8 seconds against 4.8. I do not know exactly why: the GPU had gone into deep sleep, and the process&#8217;s enabled memory may have been pushed out. It happens once, then normal speed returns. That is a real cost for an assistant you use twice a day, and nothing for one you use all evening.<\/p>\n<h3 class=\"wp-block-heading\">When the expensive option is the cheap one<\/h3>\n<p class=\"wp-block-paragraph\">Here is the counterargument, and it is a strong one. Take the 5090&#8217;s median price minus my card: about C$5,950. Divide by what an hour of your time is worth. At C$60 an hour, the card pays for itself after about 100 hours of waiting removed. If you work with a local agent all day and truly sit blocked while it runs, you reach that in a few months. Waiting is rarely fully blocked, so count only the time you could not use for something else.<\/p>\n<p class=\"wp-block-paragraph\">Power runs against intuition too. I read my machine&#8217;s own sensors during a real coding session on the 35B: 83 W for the GPU, 60 W for the CPU running the model, 75 W for the second CPU, which is mostly serving other things, and 17 W for memory. That is 235 W before fans, drives and power supply losses, because I have no wall meter. The share that belongs to the model alone is about 153 W.<\/p>\n<p class=\"wp-block-paragraph\">That makes my setup the least efficient in the table by a wide margin, at 10 to 11 joules per token against about 1.4 on the M1 Pro. At about 12 cents per kWh in Alberta, before delivery charges, a million tokens costs me about 35 cents and the Mac about 5. The sum is small under load, but idle is where it adds up. Dual-Xeon workstations reportedly idle around 100 to 170 W (I have not metered mine), against 7 W for a Mac Studio. That is roughly C$137 a year against C$7, though my box runs other services too.<\/p>\n<p class=\"wp-block-paragraph\">I did not compare hosted APIs here. For many people they are the real alternative to all three of these.<\/p>\n<h3 class=\"wp-block-heading\">Which one I would buy<\/h3>\n<ul class=\"wp-block-list\">\n<li><strong>You already own a workstation with lots of RAM channels:<\/strong> a cheap card is close to free speed. It suits overnight jobs, private documents, home automation and anything that can wait. That is my situation, and the RX 580 is the right answer for it.<\/li>\n<li><strong>You are starting from nothing on a small budget:<\/strong> the choice is close. A used-parts Z640 at C$700 to C$900 is the cheapest way in, leaves room for more RAM and a better GPU later, and can double as a home server. A used 32 GB M1 Pro costs C$500 to C$700 more, runs the 35B about 1.6 times faster, reads prompts about twice as fast, and is silent and portable. If the Z640 stays on all year, its idle power costs about C$130 a year more, which uses up that saving in four to five and a half years. Pick the Z640 if the budget is firm or you want a server. Pick the M1 Pro if you will sit and wait on it.<\/li>\n<li><strong>You code with a local agent for a living:<\/strong> buy speed. Prefill alone justifies it, and the payback above is short. Before you buy, though, spend an evening trying models on what you already have. On my card the model choice moved coding accuracy from 52 to 80 of 82 checks, which no amount of extra speed would have done.<\/li>\n<li><strong>You want one quiet machine that runs large models:<\/strong> the Mac Studio costs less than a 5090 at today&#8217;s prices, includes the computer, and uses far less power.<\/li>\n<li><strong>You want the middle:<\/strong> a used RTX 3090 at about C$1,200 to C$1,800 has 24 GB of VRAM, enough to hold the 22 GB model. I have not measured one, so that is a suggestion and not a result.<\/li>\n<\/ul>\n<p class=\"wp-block-paragraph\">The $45 card proves that an old workstation can still earn its keep, and used parts make it the cheapest way in. It does not win on speed per dollar or on running cost. The cheapest speed-up I found was not hardware at all: a different 35B model, on the same card, at the same speed, was 22 points more accurate at coding. The useful questions are what you already own and what your waiting costs, and only you know those numbers.<\/p>\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Get the next one in your inbox<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Enjoyed this? I write about local AI, self-hosting, and cloud cost ninjutsu. Join the Ninja Dispatch for occasional, high-signal notes. No fluff.<\/p>\n\n\n\n\t\t\t<form id=\"sib_signup_form_2\" method=\"post\" class=\"sib_signup_form\">\n\t\t\t\t<div class=\"sib_loader\" style=\"display:none;\"><img\n\t\t\t\t\t\t\tsrc=\"https:\/\/yyc.ninja\/wp-includes\/images\/spinner.gif\" alt=\"loader\"><\/div>\n\t\t\t\t<input type=\"hidden\" name=\"sib_form_action\" value=\"subscribe_form_submit\">\n\t\t\t\t<input type=\"hidden\" name=\"sib_form_id\" value=\"2\">\n                <input type=\"hidden\" name=\"sib_form_alert_notice\" value=\"Please fill out all required fields.\">\n                <input type=\"hidden\" name=\"sib_form_invalid_email_notice\" value=\"Please enter a valid email address.\">\n                <input type=\"hidden\" name=\"sib_security\" value=\"ec07c95cd0\">\n\t\t\t\t<div class=\"sib_signup_box_inside_2\">\n\t\t\t\t\t<div style=\"\/*display:none*\/\" class=\"sib_msg_disp\">\n\t\t\t\t\t<\/div>\n                    \t\t\t\t\t<p class=\"sib-email-area\">\n    <input type=\"email\" class=\"sib-email-area\" name=\"email\" placeholder=\"you@example.com\" required=\"required\">\n<\/p>\n<p class=\"sib-FIRSTNAME-area\">\n    <input type=\"text\" class=\"sib-FIRSTNAME-area\" name=\"FIRSTNAME\" placeholder=\"First name (optional)\">\n<\/p>\n<input type=\"hidden\" class=\"sib-SOURCE_DOMAIN-area\" name=\"SOURCE_DOMAIN\" value=\"yyc.ninja\">\n<input type=\"hidden\" class=\"sib-SIGNUP_SOURCE-area\" name=\"SIGNUP_SOURCE\" value=\"post\">\n<p>\n    <input type=\"submit\" class=\"sib-default-btn\" value=\"Join the Ninja Dispatch\">\n<\/p>\n<p class=\"ninja-fineprint\">No spam. Unsubscribe anytime.<\/p>\t\t\t\t<\/div>\n\t\t\t<\/form>\n\t\t\t<style>\n\t\t\t\tform#sib_signup_form_2{background:#192a3d;border-radius:10px;padding:16px;font-family:inherit}form#sib_signup_form_2 input[type=text],form#sib_signup_form_2 input[type=email],form#sib_signup_form_2 select{width:100%;box-sizing:border-box;margin:0 0 10px 0;padding:10px 12px;border:1px solid #33465c;border-radius:8px;background:#0f1c29;color:#e8eef5;font-size:14px}form#sib_signup_form_2 input::placeholder{color:#8a9bb0}form#sib_signup_form_2 .sib-default-btn{width:100%;margin:2px 0 0 0;padding:11px 14px;background:#e63946;color:#ffffff;border:0;border-radius:8px;font-size:15px;font-weight:600;cursor:pointer}form#sib_signup_form_2 .sib-default-btn:hover{background:#d02f3c}form#sib_signup_form_2 p{margin:0 0 10px 0}form#sib_signup_form_2 .ninja-fineprint{margin:8px 0 0;font-size:11px;color:#8a9bb0;text-align:center}form#sib_signup_form_2 p.sib-alert-message {\n    padding: 6px 12px;\n    margin-bottom: 20px;\n    border: 1px solid transparent;\n    border-radius: 4px;\n    -webkit-box-sizing: border-box;\n    -moz-box-sizing: border-box;\n    box-sizing: border-box;\n}\nform#sib_signup_form_2 p.sib-alert-message-error {\n    background-color: #f2dede;\n    border-color: #ebccd1;\n    color: #a94442;\n}\nform#sib_signup_form_2 p.sib-alert-message-success {\n    background-color: #dff0d8;\n    border-color: #d6e9c6;\n    color: #3c763d;\n}\nform#sib_signup_form_2 p.sib-alert-message-warning {\n    background-color: #fcf8e3;\n    border-color: #faebcc;\n    color: #8a6d3b;\n}\n\t\t\t<\/style>\n\t\t\t\n","protected":false},"excerpt":{"rendered":"<p>I paid $45 CAD for a used 8 GB RX 580. In an old HP Z640 workstation it runs Qwen3.6-35B-A3B, a 35-billion-parameter coding model, at 21 to 23 tokens per second on code. My 32 GB M1 Pro laptop runs the same model at 34 to 37. A Mac Studio with an M5 Max reaches about 130, and an RTX 5090 reaches\u2026<\/p>\n","protected":false},"author":1,"featured_media":3638,"comment_status":"","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[12],"tags":[],"class_list":["post-3639","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-how-to"],"featured_image_urls_v2":{"full":["https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover.png",1200,630,false],"thumbnail":["https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover-150x150.png",150,150,true],"medium":["https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover-300x158.png",300,158,true],"medium_large":["https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover-768x403.png",768,403,true],"large":["https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover-1024x538.png",1024,538,true],"RoboGalleryMansoryImagesCenter":["https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover-600x630.png",600,630,true],"RoboGalleryPreload":["https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover-100x53.png",100,53,true],"1536x1536":["https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover.png",1200,630,false],"2048x2048":["https:\/\/yyc.ninja\/wp-content\/uploads\/2026\/11\/cg-45-dollar-rx-580-vs-rtx-5090-vs-mac-studio-local-llm-cover.png",1200,630,false]},"post_excerpt_stackable_v2":"<p>I paid $45 CAD for a used 8 GB RX 580. In an old HP Z640 workstation it runs Qwen3.6-35B-A3B, a 35-billion-parameter coding model, at 21 to 23 tokens per second on code. My 32 GB M1 Pro laptop runs the same model at 34 to 37. A Mac Studio with an M5 Max reaches about 130, and an RTX 5090 reaches\u2026<\/p>\n","category_list_v2":"<a href=\"https:\/\/yyc.ninja\/index.php\/category\/how-to\/\" rel=\"category tag\">How To Ninja<\/a>","author_info_v2":{"name":"YYC Ninja","url":"https:\/\/yyc.ninja\/index.php\/author\/admin\/"},"comments_num_v2":"0 comments","_links":{"self":[{"href":"https:\/\/yyc.ninja\/index.php\/wp-json\/wp\/v2\/posts\/3639","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/yyc.ninja\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/yyc.ninja\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/yyc.ninja\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/yyc.ninja\/index.php\/wp-json\/wp\/v2\/comments?post=3639"}],"version-history":[{"count":1,"href":"https:\/\/yyc.ninja\/index.php\/wp-json\/wp\/v2\/posts\/3639\/revisions"}],"predecessor-version":[{"id":3648,"href":"https:\/\/yyc.ninja\/index.php\/wp-json\/wp\/v2\/posts\/3639\/revisions\/3648"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/yyc.ninja\/index.php\/wp-json\/wp\/v2\/media\/3638"}],"wp:attachment":[{"href":"https:\/\/yyc.ninja\/index.php\/wp-json\/wp\/v2\/media?parent=3639"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/yyc.ninja\/index.php\/wp-json\/wp\/v2\/categories?post=3639"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/yyc.ninja\/index.php\/wp-json\/wp\/v2\/tags?post=3639"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}