Which machine for which person
Links in this section may be affiliate links that earn us a commission. We also make Akhna, a Mac app. See how we test and how we earn.
Each class wins on a different number. This is the number to look at for each. Every figure is from the spec table or computed with the rules; derived figures are bounds.
RTX desktop (RTX 5090, 24 GB cards, RTX PRO)
The deciding numbers are 1,792 GB/s and 209.5 TFLOPS dense FP16/BF16 on one 32 GB card spec [S32]. For models that fit in 24-32 GB (8B dense, 30B-A3B MoE, ~50B dense at 4-bit), the 5090's bandwidth is 2.9x an M5 Max's and 6.6x a DGX Spark's derived; its prefill floor for the 8B example is 1.2 s against 4.1 s-4.2 s for the AMD chips derived (the Spark, the Macs and the RTX PRO cards publish no dense FP16/BF16 peak to put next to it); and CUDA is the default target of common fine-tuning and serving software. The limits are memory per card, power (575 W per 5090 plus the PC spec [S33]), and price: in September 2026 the three cheapest 5090 listings on eBay averaged $5,581 a day cited [S37], about 2.8x its $1,999 MSRP (new or used not stated).
-
GeForce RTX 5090
$1,999: card, MSRP (2025-01-06) spec [S36]$5,581: card, eBay: average of the 3 lowest listings per day, Sept 2026, new or used not stated (2026-09) cited [S37]NVIDIA RTX 5090 (maker page)
-
RTX PRO 6000 Blackwell
96 GB · 1,792 GB/s spec [S40]$16,000: Workstation Edition, NVIDIA US Marketplace (2026-08-12) cited [S49]NVIDIA RTX PRO 6000 (maker page)
-
GeForce RTX 4090 (previous generation)
24 GB · 1,008 GB/s spec [S32]$1,599: launch price, Oct 2022 (2022-10) cited [S38]NVIDIA RTX 4090 (maker page)
-
RTX PRO 4000 Blackwell
24 GB · 672 GB/s spec [S39]$3,134: lowest new US listing (tracker of Newegg / Best Buy / B&H) (2026-10-15) cited [S54]$2,999: PNY list price at a reseller (2026-10-15) cited [S55]NVIDIA RTX PRO 4000 (maker page)
DGX Spark-class GB10 machines
The deciding numbers are 128 GB under CUDA in a 240 W box, and 200 Gb/s clustering up to four units spec [S14] [S15]. This is the machine for people who develop for NVIDIA's data-centre software and want the same stack on the desk, or who fine-tune mid-size models with CUDA tools. The limit is 273 GB/s: decode ceilings on large dense models are about 23% of an M5 Ultra's and about equal to AMD's derived. $3,999 at launch (2025-10-13), $4,699 from the week of 2026-02-23, $6,950 from 2026-10-02; OEM 128 GB prices not verified cited [S50] [S51] [S21] [S17].
-
DGX Spark, 128 GB (NVIDIA and OEM GB10 versions)
128 GB · 273 GB/s spec [S14]$6,950: NVIDIA's own unit (Founders Edition) (2026-10-02) cited [S17]NVIDIA DGX Spark (maker page)
-
DGX Spark, 64 GB (partners only)
64 GB · 273 GB/s cited [S17]NVIDIA DGX Spark (maker page)
AMD Ryzen AI Max mini PCs
The deciding number is cost per GB. For 128 GB with 96 GB usable by the GPU: $27-31 per GB derived. For 192 GB with 160 GB usable (PRO 495): $35-39 per GB derived; that machine holds a 235B-A22B MoE model (132 GB resident) a 128 GB machine cannot. Bandwidth (256-273 GB/s) and therefore decode ceilings are about equal to the Spark's. AMD publishes 60-61 TFLOPS FP16 for the GPU; NVIDIA publishes no FP16/BF16 figure for GB10, so the two cannot be compared on prefill from specs. The caveat is software: AMD lists the chip as supported under ROCm 7.2.1, with production PyTorch support, and has officially validated only FP16 so far spec [S26].
-
Ryzen AI Max+ PRO 495 systems (AMD Ryzen AI Halo next-gen, Framework Desktop 192 GB)
192 GB · 273 GB/s spec [S24]Framework Desktop (maker page)
-
Ryzen AI Max+ 395 systems (Framework Desktop, AMD Ryzen AI Halo, others)
128 GB · 256 GB/s spec [S22]$3,449: Framework Desktop DIY (2026-10-15) spec [S27]$3,999.99: AMD Ryzen AI Halo, 2 TB SSD, Windows 11 Pro or Linux (Micro Center only) (2026-07-06) cited [S53] [S31]Framework Desktop (maker page)
Apple Silicon Macs
We make a Mac app; read this section with that in mind. The deciding numbers are bandwidth at large capacity: 614 GB/s at up to 128 GB (M5 Max) and 1.2 TB/s at up to 512 GB (M5 Ultra) spec [S1]. For a model over 100 GB, these are the options that hold 235B-A22B (132 GB resident, 12.4 GB read per token):
| Option | Decode ceiling, 235B-A22B derived | Price |
|---|---|---|
| Mac Studio M5 Ultra, 256 GB, one unit | 97 tokens/s | $10,799 |
| Ryzen AI Max+ PRO 495, 192 GB, one unit | 22 tokens/s | $6,799 |
| Two DGX Spark 128 GB, tensor-parallel, before link costs | 44 tokens/s | $13,900 |
| Two RTX PRO 6000 (192 GB), tensor-parallel, before PCIe link costs, cards only | 290 tokens/s | $32,000 |
Prices: M5 Ultra 36/80-core, 256 GB / 1 TB, derived from Apple's base price and configurator steps (2026-08-25) derived [S1] [S57] [S58]; PRO 495 Framework DIY, pre-order spec [S27]; two Sparks at NVIDIA's $6,950 each cited [S17]; two RTX PRO 6000 at NVIDIA's $16,000 listing, out of stock, plus a host PC cited [S49].
So on one unit the M5 Ultra's ceiling is 4.4x the PRO 495's at about 1.6x its DIY price, and 2.2x that of two Sparks at about 0.8x their price; two RTX PRO 6000 cards have about 3x the M5 Ultra's ceiling at about 3x its price before the host PC derived. The limits: no CUDA; Apple publishes no FLOPS figure, so prefill and fine-tuning speed have to be measured; and whole-machine cost per GB is higher than the AMD machines. Memory upgrade steps cost about the same per GB: Apple charges $2,000 to go from 48 to 128 GB on the M5 Max 40-core ($25 per GB) cited [S10], and Framework's 64 to 128 GB step on the Max+ 395 is $1,490 ($23 per GB) spec [S27].
-
Mac Studio, M5 Ultra (30-core CPU / 64-core GPU, or 36 / 80)
96 / 256 / 512 GB · 1,200 GB/s spec [S1]$5,499: 30/64-core, 96 GB / 1 TB (2026-08-25) spec [S2]$10,799: 36/80-core, 256 GB / 1 TB (Apple configurator: 512 GB/16 TB $29,799 minus the 256 GB and 1 TB steps) (2026-10-30) spec [S62]Apple Mac Studio (maker page)
-
Mac Studio, M5 Max 18-core CPU / 40-core GPU
48 / 64 / 128 GB · 614 GB/s spec [S1]$3,099: 48 GB / 512 GB (2026-08-25) cited [S10]$5,099: 128 GB / 512 GB ($3,099 + $2,000) (2026-08-25) derived [S10]Apple Mac Studio (maker page)
-
MacBook Pro 14/16, M5 Max 40-core GPU
48 / 64 / 128 GB · 614 GB/s spec [S4]not verified for 128 GB configurations; check apple.comApple MacBook Pro (maker page)
-
Mac Studio, M5 Max 18-core CPU / 32-core GPU
36 GB · 460 GB/s spec [S1]Apple Mac Studio (maker page)
-
Mac mini, M5 Pro
24 / 48 / 64 GB · 307 GB/s spec [S5]$1,699: 24 GB / 512 GB (2026-08-25) cited [S13]$2,699: 64 GB / 512 GB ($1,699 + $1,000) (2026-08-25) derived [S13]Also with this chip: MacBook Pro 16, M5 Pro (18-core CPU / 20-core GPU), 24 GB / 1 TB, at Walmart We earn a commission if you buy through this link.Apple Mac mini (maker page)
Sources
- [S1] Apple, Mac Studio tech specs
- [S2] Apple newsroom, new Mac Studio with M5 Max and M5 Ultra (2026-08-25)
- [S3] Apple, Mac Studio buy page
- [S4] Apple, MacBook Pro tech specs
- [S5] Apple, Mac mini tech specs
- [S7] Apple Support, Mac Studio power consumption and thermal output
- [S9] Daring Fireball, M5 Max Mac Studio base prices (2026-08-25)
- [S10] EveryMac, Mac Studio M5 Max 18-core CPU / 40-core GPU
- [S12] MacRumors, 512 GB M5 Ultra in October (2026-08-25)
- [S13] Macworld, M6 and M5 Pro Mac mini prices (2026-08-25)
- [S14] NVIDIA, DGX Spark product page
- [S15] NVIDIA docs, DGX Spark hardware overview
- [S16] The Register, GB10 workstations arrive (2025-10-13)
- [S17] The Register, $4,999 64 GB DGX Spark; 128 GB to $6,950 (2026-10-02)
- [S18] NVIDIA blog, DGX Spark 64 GB (2026-10-02)
- [S19] StorageReview, DGX Spark review (2025-10-14)
- [S21] Tom's Hardware, DGX Spark $3,999 to $4,699 (Feb 2026; headline read)
- [S22] AMD, Ryzen AI Max+ 395 specifications
- [S23] AMD blog, Variable Graphics Memory FAQ (2025-07-29)
- [S24] AMD, Ryzen AI Halo with Ryzen AI Max+ PRO 495
- [S25] AMD blog, Ryzen AI Halo and Ryzen AI Max PRO 400 series (2026-05-20)
- [S26] AMD ROCm on Radeon and Ryzen, Linux support matrices (7.2.1)
- [S27] Framework, Desktop DIY Edition (prices, stock)
- [S28] Framework, Desktop overview
- [S29] Framework blog, 192 GB Framework Desktop pre-order (2026-09-30)
- [S31] Tom's Hardware, $3,999 Ryzen AI Halo (headline read)
- [S32] NVIDIA, RTX Blackwell GPU architecture whitepaper (Table 3)
- [S33] NVIDIA, GeForce RTX 5090 product page
- [S34] NVIDIA, GeForce RTX 4090 product page
- [S35] NVIDIA, GeForce compare page
- [S36] NVIDIA newsroom, RTX 50 series prices (2025-01-06)
- [S37] GPU Poet, RTX 5090 prices September 2026 (eBay listings)
- [S38] Wikipedia, GeForce RTX 40 series
- [S39] NVIDIA, RTX PRO 4000 Blackwell
- [S40] NVIDIA, RTX PRO 6000 Blackwell
- [S42] Llama 3 paper, Table 3
- [S43] Qwen3-30B-A3B model card
- [S44] Qwen3-30B-A3B config.json
- [S45] Qwen3-235B-A22B model card
- [S46] MLX docs, mlx.core.quantize
- [S47] mlx-lm README
- [S48] Rajbhandari et al., ZeRO
- [S49] Let's Data Science, RTX PRO 6000 price (2026-08-12)
- [S50] PCMag via Yahoo, $3,999 DGX Spark (2025-10-13)
- [S51] NVIDIA Developer Forums, 2/23/2026 price change (search result)
- [S52] Phoronix, Framework Desktop PRO 495 from $6,799 (2026-09-30)
- [S53] Let's Data Science, Strix Halo mini PC for $3,999 (2026-07-06)
- [S54] gpuprix, RTX PRO 4000 Blackwell US prices
- [S55] SHI, PNY RTX PRO 4000 Blackwell (search result)
- [S56] Intel, Thunderbolt 5 technology brief
- [S57] EveryMac, Mac Studio M5 Ultra 30-core CPU / 64-core GPU
- [S58] Daring Fireball, configurations and pricing for new Mac minis and Mac Studios (2026-08)
- [S59] Framework, Desktop DIY (Max+ PRO 495) configurator: SSD extra or bring your own; Batch 2 ships November (read 2026-10-20)
- [S60] Borncity, Ryzen AI Halo 192 GB at Micro Center for $7,299 (2026-10-14)
- [S61] PCWorld, the RTX Spark laptops available at launch (2026-10-08)
- [S20] NVIDIA docs, RTX Spark porting guide, system overview
- [S62] Apple, Mac Studio configurator (36/80-core M5 Ultra, 512 GB, 16 TB: $29,799; option steps shown; ships 10-12 weeks; read 2026-10-30)
Accessed 2026-10-15; volatile prices re-checked 2026-10-20.