4× V100 16GB SXM2 server: what it runs, what it costs
Four Tesla V100 16GB SXM2 modules on an NVLink board give 64GB of HBM2, enough for a 27B model at FP8 with a 256K-token context. Dell's PowerEdge C4140 carries the four in 1U, and 1Cat-vLLM, a vLLM fork for V100s, runs current models on them.
As of October 3, 2026, one in ten of the 11 complete 4× V100 16GB SXM2 server listings on eBay asks $6,500 or less shipped. The median ask is $6,990. 1Cat-vLLM's developer reports Qwen3.8-27B-FP8 decoding at 65.05 tok/s (batch 1, all four GPUs).
What it runs: published benchmarks
| Model and engine | Setup and workload | Result | Source |
|---|---|---|---|
| Qwen3.8-27B-FP8 1Cat-vLLM 1.3.0 | All 4 GPUs, tensor parallel (TP4) Batch 1; 1K-token prompt; the model's recommended sampling (temperature 1.0); no speculative decoding; FP16 KV cache; CUDA Graph | 65.05 tok/s decode 46.95 tok/s at a 128K-token context, 36.96 at 256K. Four requests at once: 192.39 tok/s in total. The record first named four V100 32GB cards; PR #229 corrected it to 16GB SXM2 before the release. A later 1Cat design note says no physical 16GB GPUs are available for its tests. | 1Cat-vLLM 1.3.0 acceptance record Reported released Aug 17, 2026 |
| Qwen3.8-27B-FP8 1Cat-vLLM 1.2.3 dev image (d76126608) | All 4 GPUs, tensor parallel (TP4), NVLink up 4 requests at once; 1,000 tokens in, 1,000 out; temperature 0; Volta DeltaNet decode kernel off | 38.3–41.9 tok/s output in total, about 10 per request Three repeats; median time per output token 94–103 ms. 1Cat's own record above reports 16.88 ms at four requests; CloudRift's older image can't build the Volta DeltaNet decode kernel, so it ran with that kernel off. A cap of 16 requests gave 133.8 tok/s in total. With NVLink off, time to first token was 2.5–3.5× longer. | CloudRift's Qwen3.8-27B-FP8 recipe results Community re-qualified Sep 15, 2026 |
| Qwen3.8-27B-QUASAR-NVFP4, with DFlash2 speculative decoding (7 draft tokens) 1Cat-vLLM 1.5.0 with two patches from the post | All 4 GPUs, tensor parallel (TP4), every pair NVLinked; capped at 175 W with clocks locked at 1260 MHz Batch 1; decoding at a 200K-token context; FP8 KV cache; DFlash2 speculative decoding | 100–142 tok/s decode at a 200K-token context Speed follows how many drafted tokens are accepted: 3.5–4.3 per step. Cold prefill 1,212 tok/s; peak memory 15.73 of 16GB per card. The poster, a self-described enthusiast working with Gemini, filed it as a bug report with two patches. Their host is an older Xeon E5-2680 v4 server, not a C4140. | 1Cat-vLLM issue #467 Community posted Sep 3, 2026 |
Reported: published by the engine's developer; not reproduced by rackbites. Community: posted by a user; not reproduced by rackbites. rackbites hasn't measured this configuration yet. Every figure is the source's own, on its stated setup; a different engine version, driver or workload gives a different number.
Before you buy
- Buy on VRAM. This configuration is 16GB per card, 64GB in all. Qwen3.8-27B-FP8 needs all four: CloudRift puts it near 40GB to serve, more than two cards hold. A listing that doesn't state the card size is a near miss below, not a match.
- Buy the right C4140. Dell builds it five ways, and only Configurations K and M carry the SXM2 NVLink board; B, C and G are PCIe-switch builds for PCIe cards. A listing that says PCIe, or doesn't say, is a near miss below.
- Power: four 300 W GPUs on 2,000 W or 2,400 W supplies, and Dell's guide says the C4140 requires 220 V. A 120 V household outlet won't do.
- Dell requires both CPU sockets filled and four identical GPUs, and doesn't support unbuffered (UDIMM) memory.
- NVLink: run nvidia-smi topo -m when it arrives. CloudRift's host had a partial mesh, with two GPU pairs talking over PCIe; the issue #467 poster's had every pair linked. With NVLink off, CloudRift measured time to first token 2.5–3.5× longer.
- Noise: Dell designed the C4140 for an "unattended data center" and says hearing protection may be expected where it runs. It belongs in a rack, a garage or a basement, not an office.
- Drivers: NVIDIA's R580 branch is the last to support Volta, and CUDA 13 dropped it. Stay on CUDA 12.x.
- Software: all three figures come from 1Cat-vLLM, a vLLM fork for V100s (SM70) whose recent work comes from one developer, and on three different versions (1.2.3, 1.3.0 and 1.5.0). Pin the version you test.
Complete 4× V100 16GB SXM2 server listings as of October 3, 2026
eBay affiliate disclosure: The listing links below lead to eBay. rackbites may earn a commission if you make a purchase.
Listings with the same configuration and price share one row, and the numbered links are the other copies. Each copy still counts toward the listing total, the low end and the median.
The second figure, “with 128GB RAM”, adds the missing memory at the low-end price of DDR4-2666 RDIMMs ($5.62/GB, from DIMMsum), so servers that ship with different amounts compare fairly. A dash means the listing doesn't say how much RAM ships.
Near misses: 4× Tesla V100 servers that differ on one detail
14 near misses, showing 12 rows: listings that leave a detail out come first, since they may still be the right machine, then listings that differ. Cheapest first within each. The reason comes from what the title states.
Or build it from parts
| Part | Qty | Low-end each | Subtotal | |
|---|---|---|---|---|
| Dell PowerEdge C4140 barebone, SXM2 (Configuration K or M) | 1× | $3,878 | $3,878 | |
| Tesla V100 16GB SXM2 module | 4× | $155 | $620 | |
| Xeon Gold 6138 (20-core) | 2× | — | — | |
| DDR4-2666 RDIMM, 128GB | 128GB | $5.62/GB | $720 | |
| Parts total | — | |||
The barebone line counts only C4140s listed as SXM2. Dell's layout puts a heatsink on each GPU, but used barebones don't always include them, and loose SXM2 modules are often sold without them. Confirm both before you count on this total.
The C4140 boots from an internal BOSS M.2 card; check the barebone has one. The total leaves out drives and rails. Each line is that part's low-end ask with shipping, which one in ten of its listings match or beat.
A line needs 3 live listings before it gets a price. See all proven builds.