It depends on whether you need one big model or several separate machines. Two NVIDIA RTX Spark PCs such as the ASUS ProArt GR1X give you two independent 128 GB computers: useful for two users, two workloads or a spare if one fails, but they can't pool memory to run a single larger model, because RTX Spark has no supported clustering method. One Apple Mac Studio M5 Ultra gives you up to 512 GB in a single box, more memory bandwidth and one machine to manage, and up to four Macs can later be linked over Thunderbolt 5. If what you really want is two NVIDIA machines that work as one, a pair of GB10 systems with ConnectX-7 is the supported route.
How do the options compare on cost and capability?
| Option | Price | Memory | Can run one model across both? |
|---|---|---|---|
| Two ASUS ProArt GR1X, 128 GB/2 TB | S$18,000–21,600 (SourceIT estimate, see below) | 2 x 128 GB, separate | No supported method |
| Two ASUS ProArt GR1X, 64 GB/1 TB | S$12,400–14,400 (SourceIT estimate) | 2 x 64 GB, separate | No supported method |
| One Mac Studio M5 Ultra, 256 GB/2 TB | About S$16,699 configured (calculated from Apple Singapore's upgrade pricing; confirm on apple.com/sg) | 256 GB in one machine | Not needed; add Macs later over Thunderbolt 5 |
| One Mac Studio M5 Max, 128 GB/1 TB | About S$7,849 configured (calculated as above) | 128 GB | Not needed |
| Two ASUS Ascent GX10, 128 GB/1 TB | S$20,700 (2 x S$10,350, listed 8 Nov 2026) | 2 x 128 GB, linkable | Yes, over ConnectX-7 200GbE |
Source: Apple Singapore list prices (M5 Ultra from S$7,999, M5 Max from S$3,499) and upgrade pricing; SourceIT listings (8 Nov 2026); NVIDIA Cluster Assistant documentation; ASUS GR1X specifications. GR1X figures are the SourceIT estimate based on comparable RTX Spark devices, not an official ASUS price. Totals are our arithmetic.
On these numbers, two 128 GB RTX Spark desktops cost about the same as one 256 GB M5 Ultra, somewhere between S$1,300 less and S$4,900 more. The question is what you get for the money.
Can two RTX Spark PCs work as one machine?
Not in any supported way, as of November 2026. RTX Spark drops the ConnectX-7 200GbE networking that DGX Spark and other GB10 machines use to link units. The GR1X has a 10GbE port, about a twentieth of ConnectX-7's bandwidth, and we found no NVIDIA- or OEM-supported clustering method and no public results for splitting a model across two RTX Spark machines over 10GbE or USB.
So "two 128 GB PCs" is not "256 GB". Each machine can only run models that fit in its own memory, and on Windows on Arm the GPU's share of that memory is fixed. Our 128 GB post looks at which models fit.
The GB10 route is different. NVIDIA's Cluster Assistant supports two to four units, with two connected directly by a QSFP cable, and NVIDIA says two linked units can run models of about 400B parameters. Its own two-unit test ran Qwen3 235B in NVFP4 at 11.73 tokens per second (NVIDIA, Oct 2025). Even so, one developer found the 200G link "fast for a network, but slow compared to the memory bandwidth within a single DGX Spark" (Dataiku, Sep 2026).
Two RTX Spark PCs double your machines, not your model size. If you need a bigger model, buy more memory in one box or buy hardware designed to link.
Which gives more throughput for a small team?
No RTX Spark desktop has been benchmarked publicly, so the nearest data comes from laptops with the same N1X chip and memory bandwidth (about 300 GB/s).
- Two RTX Spark machines: Techtesters measured gpt-oss-120b at 63.3 tokens per second on an MSI Prestige N16 128 GB laptop (Oct 2026; tool and quantisation not published). Two machines serving one user each would give about 126.6 tokens per second combined, or 63.3 each.
- One M5 Ultra: a community oMLX run measured gpt-oss-120b at about 110 tokens per second for one request and 180.6 total for two (Sep 2026, unverified), so about 90 tokens per second each for two users.
These use different software and hardware classes, so treat them as indications. The pattern fits the hardware: the M5 Ultra's 1.2 TB/s of memory bandwidth is about four times RTX Spark's. The two-PC setup wins on isolation, since one user's long prompt can't slow the other machine, and on prompt-heavy, image or CUDA work, where NVIDIA's tensor cores help.
What does running two machines involve?
- Twice the administration. Two Windows installs, two sets of driver and CUDA updates (CUDA on Windows on Arm was still a developer preview in November 2026), two backups and two security policies. See our Windows on Arm compatibility guide.
- Routing requests. To share two machines between users or agents, you need either fixed assignments or a gateway that sends each request to whichever box is free.
- Redundancy. This is the real advantage. If one machine fails or is being updated, the other keeps working. A single Mac is a single point of failure.
- Power. No measured GR1X power figures exist. ASUS rates it at a 140 W sustained TDP with a 240 W adapter; even two adapters at full rating all year would cost S$1,310.22 at 31.16¢/kWh, against S$1,050.90 for an M5 Ultra at its 385 W maximum. Both are upper limits; real use will be far lower.
How does each option scale later?
Mac Studio: RDMA over Thunderbolt 5 (macOS 26.2+, MLX 0.30.1+) links up to four Macs in a mesh with direct cables. Enabling it needs a reboot into Recovery mode on each Mac. Apple's four-Mac M5 Ultra 512 GB cluster ran an 816B model at 36.73 tokens per second (Apple's own figure). Adding a second Mac later is the closest thing to "doubling up" that works today.
RTX Spark: scaling means adding more independent machines. That suits a team where each person or agent needs their own computer, but not one shared large model.
GB10: two to four linked units with NVIDIA's tools, on Linux with full CUDA. See RTX Spark vs GB10 and three ways to build an AI setup.
Which should you choose?
- Two RTX Spark PCs if you have two people or workloads that each fit in 128 GB, want Windows and CUDA, and value having a spare.
- One Mac Studio M5 Ultra if you need models larger than 128 GB, want the fastest single-stream speed in this group, and prefer one machine to manage.
- Two GB10 machines if you want NVIDIA hardware that can genuinely combine into one larger system.
- One M5 Max 128 GB if your models fit in about 96 GB of GPU memory and budget matters more than headroom. See whether you're overbuying.
Buying in Singapore
The ASUS ProArt GR1X is available on a price-on-request basis, and MSI's RTX Spark machines are in our MSI RTX Spark collection. The Mac Studio M5 Ultra 256 GB, M5 Ultra 512 GB and M5 Max 128 GB are also price on request. For linked NVIDIA systems we list the ASUS Ascent GX10 128 GB at S$10,350 and the NVIDIA DGX Spark 4 TB at S$9,300 (listed 8 Nov 2026); ask us about the QSFP cable for pairing. Prices change; ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Can I combine two RTX Spark PCs into 256 GB? No. RTX Spark has no ConnectX networking and no supported clustering method, so each machine is limited to its own 128 GB.
Can two Mac Studios be linked? Yes. RDMA over Thunderbolt 5 on macOS 26.2 or later links up to four Macs with direct cables, using MLX.
Which NVIDIA desktops can be linked? GB10 machines such as DGX Spark and ASUS Ascent GX10, over ConnectX-7. NVIDIA's tools support two to four units.
Are two cheaper machines more reliable than one? In one sense, yes: if one fails, the other keeps working. But you have twice the updates, backups and hardware to look after.
Is the GR1X price official? No. The figures here are a SourceIT estimate based on comparable RTX Spark devices, not an official ASUS price. Ask us for a quotation.
