BLUF: Choose the chassis from the model tier, not from the model name: 2U 2288H V7 for DeepSeek-R1-70B class inference, 1U 1288H for edge and light serving, 4U G5500 when one node must hold more accelerators than a 2U can carry. Verified specs, sizing arithmetic and the spares plan follow.
Figures come from xFusion’s English product pages and technical specifications, checked 2026-10-30. Where xFusion publishes no number we say so; the one projection is labelled as arithmetic. No pricing lists, no part numbers, and no claim to vendor-channel status.
1. Size the workload before you size the server
DeepSeek deployments fail on sizing more often than on hardware: teams pick a chassis they like, then find the model does not fit or the rack cannot power the node. Four numbers, in this order, prevent that.
| Order | Number to fix first | What it decides | The mistake we see most |
|---|---|---|---|
| 1 | Weight footprint per model instance (GB) | Whether the model fits on one node at all | Counting one instance per node and forgetting the failover replica |
| 2 | Memory bandwidth per accelerator | Tokens per second, especially at batch size 1 | Buying capacity and never checking bandwidth |
| 3 | Interconnect between accelerators | Whether tensor parallelism is usable at all | Assuming a PCIe fan-out behaves like a dedicated GPU interconnect |
| 4 | Rack power and cooling per node | Chassis height, PSU rating, room capacity | Sizing the node first and asking the facility afterwards |
Three definitions get mixed up constantly. VRAM is the memory on an accelerator, capping model size and resident context. KV cache is per-request state that grows with context length and concurrency, competing with the weights for the same VRAM. Tensor parallelism splits one model across accelerators and works well only when the link between them is fast.
Do the arithmetic before the model conversation. A 70B parameter model needs about 140 GB for BF16 weights, about 70 GB at FP8 and about 35 GB at INT4. Add roughly a quarter for KV cache and activation headroom and the BF16 case lands near 175 GB: three 80 GB accelerators as a floor, four with headroom at load.
A 671B mixture-of-experts model is a different problem. At BF16 the weights alone are about 1.3 TB, so it spans several nodes and the question stops being chassis height and becomes interconnect and parallel strategy. Plan that deployment from the cluster down, not from a single node up.
2. The 1U line: 1288H V5, V6 and V7 for edge and light serving
The 1288H is the 1U dual-socket branch: the right machine for an inference gateway, an embedding or reranking service, a small model behind an internal API, or the control-plane node of a cluster whose accelerators sit in 2U and 4U chassis. It is the wrong machine for heavy serving, and xFusion’s own documentation says so.
| Model | CPU | Memory slots / speed | Maximum memory | Drive bays | PCIe | GPU support listed | PSU |
|---|---|---|---|---|---|---|---|
| 1288H V5 | 1 or 2x 1st/2nd Gen Xeon Scalable, up to 205 W | 24x DDR4 up to 2933, plus 12x PMem100 | Not published | 4x 3.5″ or 8/10x 2.5″; NVMe 4/8; 2x M.2 | Not published | Not stated | 900 W Platinum AC in official V5 configurations; full line-up not published |
| 1288H V6 | 1 or 2x 3rd Gen Xeon Scalable, up to 270 W | 32x DDR4 up to 3200, plus 16x PMem200 | Not published | 8-12x 2.5″ or 4x 3.5″; mixed NVMe up to 10; 2x M.2 | 6 (3x Gen4, 1 RAID, 2 FlexIO/OCP 3.0) | No | 2x hot-swap 1+1, 900 W and up |
| 1288H V7 | 1 or 2x 4th/5th Gen Xeon Scalable, up to 385 W | 32x DDR5 up to 5600 | Not published | 4x 3.5″ or 8-12x 2.5″; 2x M.2 hardware RAID, hot-swap | 5 (2x FlexIO OCP 3.0 hot-swap + 3x PCIe, one PCIe 5.0) | No | 900/1200/1500/2000 W, 1+1 hot-swap |
Two cells in that table need their provenance stated, because the V5 and V6 pages are not equally available. xFusion has retired the V5 product pages — the 1288H V5 and 2288H V5 URLs now return 404 — so every V5 figure here comes from the FusionServer Pro V5 datasheet of September 2020 and from xFusion’s own published configuration listings for its EU value program, which is where the 900 W Platinum AC PSU and the onboard 2 x GE + 2 x 10GE LOM are stated. Where those documents are silent we write “Not published” and leave the cell open rather than fill it in; the V5 GPU cell reads “Not stated” for the same reason, not because the platform cannot take a card. The 1288H V6 page is live, and its banner advertises “12 TB memory capacity” — but the technical specifications give 32 DIMM slots and no maximum, and 12 TB does not fit 32 slots, so we quote the slots and leave the ceiling open.
The V7 is the generation worth buying new: 32 DDR5 DIMMs at up to 5600 MT/s to feed the accelerators; five expansion positions, two of them hot-swap OCP 3.0 for networking; and PCIe 5.0 with 400G NIC interconnect, so a 1U node can front a GPU pool. Cooling is eight hot-swap counter-rotating fans in N+1, depth is 798 mm in both drive configurations (43 x 447 x 798 mm), and the operating range is 5-45 degrees C.
One point we will not soften: xFusion’s 1288H V7 product page does not list GPU support. The technical specifications run from form factor to dimensions with no GPU card row. The 2288H V7 and 5885H V7 pages do carry one, so this is a statement about the model, not a gap across the generation. Accelerators in a 1U are a per-project question with the vendor documentation open.
The V5 generation is narrower: 24 DDR4 slots at up to 2933 with 12 PMem100 modules, 4x 3.5″ or 8/10x 2.5″ bays, NVMe in 4 or 8 drive configurations. It still works as a gateway or light inference node, but it is PCIe 3.0 and the published documentation is thinner. More on the platform: 1288H V7 rack server overview, 1288H V7 1U configuration and 1288H V5 rack server.
3. The 2U line: 2288H V5, V6 and V7 as the main inference workhorse
This is where most DeepSeek inference lands: four double-width accelerator positions, enough DIMM slots to stop counting, and a power envelope a normal rack can feed.
| Model | CPU | Memory slots / speed | Maximum memory | Drive bays | PCIe | GPU support | PSU |
|---|---|---|---|---|---|---|---|
| 2288H V5 | 1 or 2x 1st/2nd Gen Xeon Scalable, up to 205 W | 24x DDR4 up to 2933, plus 12x DCPMM | 3 TB with 128 GB modules; 7.5 TB with DCPMM | 20x 3.5″ or 31x 2.5″; NVMe 4/8/12/24/28; 2x M.2 | 10x PCIe 3.0 | 2x double-width full-height full-length | 900 W Platinum AC in official V5 configurations; full line-up not published |
| 2288H V6 | 1 or 2x 3rd Gen Xeon Scalable, up to 270 W | 16/32x DDR4 up to 3200, plus 16x PMem200 | 12 TB (xFusion figure, 32 slots) | 8-31x 2.5″ or 12-20x 3.5″; NVMe 4/8/16/24; up to 45x 2.5″ or 34x all-NVMe | 14x PCIe 4.0 | 4x 300 W double-width, 11x half-height, or 8x single-width | 2x hot-swap 1+1, 900 W and up |
| 2288H V7 | 1 or 2x 4th/5th Gen Xeon Scalable, up to 385 W | 32x DDR5 up to 5600, plus 16x via CXL, up to 48 slots | Not published by xFusion; 32 x 256 GB is about 8 TB, our arithmetic projection, not an official figure | 8-33x 2.5″ (up to 45) or 12-18x 3.5″; NVMe up to 28; 2x M.2 | Up to 19 (2x FlexIO OCP 3.0 + 17 standard, 8 of them PCIe 5.0) | 4x double-width or 14x single-width | 900/1200/1500/2000/3000 W, 1+1 |
The V7 is the default for a new inference cluster, and three numbers explain it: eight PCIe 5.0 slots inside nineteen positions, so four double-width accelerators are fed without starving the network cards; 48 memory slots counting the 16 CXL-attached positions, enough host memory for several replicas per node; and a 3000 W option with 1+1 redundancy, the difference between four accelerators at full draw and four derated. Cooling is four hot-swap counter-rotating fans in N+1, depth 798 mm, range 5-50 degrees C.
The V6 still makes sense when the budget is fixed and the workload known: twelve terabytes of DDR4 at 3200 MT/s, fourteen PCIe 4.0 slots for four 300 W double-width accelerators, and drive configurations up to 45x 2.5″ or 34x all-NVMe. What it lacks is Gen5 bandwidth.
The V5 is the value option and the honest limitation is PCIe 3.0: ten Gen3 slots and two double-width positions will serve a 70B model at modest concurrency, and 3 TB of DDR4 with DCPMM is no bottleneck for inference, but every accelerator added competes for Gen3 lanes. For that trade-off see pros and cons of the 2288H V5; for the current generation see our 2288H V7 performance and configuration notes.
Two naming traps. The 2288H V7 and 2488H V7 are not the same machine: both are 2U, but the 2488H V7 is quad-socket with ten PCIe slots (one FlexIO plus nine standard) and no GPU row, while the nineteen slots, eight Gen5 positions and four double-width GPUs belong to the 2288H V7. The 2288 and 2288H differ too: 16 DDR4 slots and 2 TB on the 2288 V5, 24 slots and 3 TB on the 2288H V5. Variant compatibility is collected in our RH2288H V7 compatibility list.
4. The 4U line: G5500 for DeepSeek-R1-70B class deployments
The G5500 is the 4U multi-accelerator platform, and the reason it exists is physical: more card positions, more power headroom, larger fans, and enough space to service an accelerator without pulling the node out of the rack. For a 70B class deployment at production concurrency, that space is the difference between one node and two.
We publish no slot counts, accelerator counts, maximum VRAM or bay numbers for the G5500: those figures are configuration-dependent and we will not quote them from memory. The selection rule is more useful: stay on 2U while the accelerator count fits within four double-width positions; move to 4U when the model, the concurrency target or the replica count pushes past what a 2U can carry. Supported accelerator brands are listed on the G5500 V6 4U AI server page; the card mix and power budget are confirmed per project.
5. Memory, storage and power: what decides whether the accelerators are fed
Under-performing inference nodes are rarely short of compute: they are short of host memory bandwidth, of local NVMe read speed at model load, or of power headroom.
| Line | Memory rule | Storage rule | Power rule | Thermal envelope |
|---|---|---|---|---|
| 1U 1288H | Fill for bandwidth, not capacity: 32 DDR5 slots at 5600 MT/s on V7, matched speed bins | 2x M.2 hardware RAID for the OS, 2.5″ NVMe for the model cache | 900-2000 W 1+1 hot-swap; do not over-buy wattage on a light node | 8 fans N+1, 5-45 degrees C, 798 mm |
| 2U 2288H | 32 DDR5 slots plus 16 CXL positions on V7; 48 slots allows several replicas per node | NVMe up to 28 drives on V7, up to 34 all-NVMe on V6; weights on local NVMe | 900-3000 W 1+1; check the rack feed before specifying 2000 W and above | 4 fans N+1, 5-50 degrees C, 798 mm |
| 4U G5500 | Sized per project with the accelerator plan, not before it | Local NVMe for weights and checkpoints | Power budget is the binding constraint | Larger chassis, more fan redundancy; confirm per project |
Think in bandwidth per accelerator rather than terabytes: a node with four accelerators and one DIMM per channel will idle them waiting on host-side work. The V7 helps twice, with DDR5 at 5600 MT/s raising per-channel bandwidth and the 16 CXL-attached positions adding capacity without breaking the population rules on the main 32 slots. On V5 the equivalent was DCPMM, which is why a 2288H V5 is quoted at 3 TB of DRAM or 7.5 TB counting persistent memory.
Keep the OS on the M.2 pair with hardware RAID and weights on local NVMe: a 70B model that reloads from network storage on every restart turns a reboot into a 20-minute outage. The 2288H V7 takes up to 28 NVMe drives, the V6 up to 34 all-NVMe.
On power, 1+1 hot-swap is the floor, not an upgrade. Above 2000 W per node the constraint moves from the server to the rack: confirm feed, PDU rating and UPS headroom before the purchase order. The full line-up, including four-socket models, is indexed on our xFusion FusionServer model overview.
6. Spare parts, warranty and lead time
An inference node is a production asset and failures are predictable: power supplies, fan modules, DIMMs, drive caddies and backplanes, and the M.2 or RAID cache device holding the OS. On a 1U 1288H V7 that is eight fans and hot-swap PSUs, so a swap is a two-minute job in service; on a 2U 2288H V7, four fans and one or two PSUs.
Our practice for an inference fleet is one spare PSU and one spare fan pair per four to six chassis, plus matched DIMMs in the exact speed bin of the installed population, because mixed-speed DIMMs down-clock the whole population. Drive caddies and backplane generation get scarce first on older platforms, so they belong in the kit for any V5 or V6 node still carrying production traffic. We quote parts against the label on your unit: send a photo and we confirm item, condition and revision before quoting. No part numbers in articles, because a wrong number in a blog post becomes a wrong purchase order.
Warranty follows condition and is stated on the order confirmation. New and factory-sealed units carry a three-year term backed by us, from our delivery date and not from the factory build date. Professionally refurbished units carry one year, tested and graded by Beijing Xinchuan Technology Co., Ltd., which is not a manufacturer recertification. Enterprise storage and separately ordered spare parts carry one year unless the line item says otherwise, so a mixed order carries two clocks. We never present refurbished hardware as new. Lead time is confirmed per project, because it depends on configuration, condition and quantity. For availability and a written quotation, contact admin@sell-server.com or WhatsApp +86 180 0106 0290.
7. Running the selection in one afternoon
- Fix the model tier and concurrency target: weights footprint in GB at the precision you will serve, plus the KV cache budget.
- Divide by per-accelerator memory, add 25 percent headroom, and note the count per node. That number decides 1U, 2U or 4U.
- Check the interconnect. If the plan needs tensor parallelism, confirm the link between accelerators before the chassis.
- Check the rack: power per node, PDU rating, cooling and depth. 798 mm covers both the 1288H V7 and the 2288H V7; the 1U runs 5-45 degrees C, the 2U 5-50 degrees C.
- Fix memory population and storage layout: matched DIMMs, OS on hardware-RAID M.2, weights on local NVMe.
- Put the spares list and the warranty basis in the same document as the bill of materials.
FAQ
Which xFusion model should I choose for DeepSeek-R1-70B inference?
Start with the 2U 2288H V7: four double-width accelerators, eight PCIe 5.0 slots and up to 48 memory slots cover a 70B model at production concurrency in most designs. Move to the 4U G5500 when one node must hold more accelerators than a 2U can, and use the 1U 1288H for gateways, embedding services and light models.
Can the 1288H V7 be configured with GPUs?
xFusion’s 1288H V7 product page does not list GPU support; the technical specifications run from form factor to dimensions without a GPU card row. The 2288H V7 and 5885H V7 pages do list it, so this is specific to the model. Treat accelerators in 1U as a per-project question.
How much memory can the 2288H V7 take?
It has 32 DDR5 slots at up to 5600 MT/s plus 16 DIMM positions attached through CXL, for up to 48 slots. xFusion does not publish a maximum capacity in terabytes. About 8 TB is arithmetic on 32 slots of 256 GB and is a projection, not an official specification; confirm the supported capacity for a given build against the current memory compatibility list.
What is the difference between the 2288H V7 and the 2488H V7?
The 2288H V7 is dual-socket with up to 19 PCIe slots, 8 of them PCIe 5.0, and supports four double-width or 14 single-width accelerators. The 2488H V7 is quad-socket with 10 PCIe slots, one FlexIO plus nine standard, and no GPU row on its product page. Both are 2U, which is why specs get copied across by mistake.
Do you supply spare parts, and what warranty applies?
Yes. New and factory-sealed units carry a three-year term backed by us from our delivery date; professionally refurbished units carry one year, tested and graded by us; separately ordered storage and spare parts carry one year unless the line item says otherwise. Parts are quoted against the label on your unit. Contact admin@sell-server.com or WhatsApp +86 180 0106 0290.
Next step
Send the model tier, concurrency target, accelerator plan, rack power available and quantity. We will come back with a configuration per line, the memory and storage population, the spares kit and a written quotation. Use the quote request form.
Sources
- xFusion 1288H V7 product page, Technical Specifications — https://www.xfusion.com/en/product/rack-server/fusionserver-1288h-v7
- xFusion 2288H V7 product page, Technical Specifications — https://www.xfusion.com/en/product/rack-server/fusionserver-2288h-v7
- xFusion 1288H V6 product page, Technical Specifications (six PCIe slots, seven fans, 5-45 degrees C, no GPU row) — https://www.xfusion.com/en/product/rack-server/fusionserver-1288h-v6
- xFusion 2288H V6 product page, Technical Specifications (14 PCIe 4.0, four 300 W dual-width GPUs, 45 x 2.5″ or 34 all-NVMe, 5-45 degrees C) — https://www.xfusion.com/en/product/rack-server/fusionserver-2288h-v6
- xFusion EU value program configuration listings, the surviving official source for V5 builds (900 W Platinum AC PSU, 2 x GE + 2 x 10GE LOM, 12 x 3.5″ chassis) — https://www.xfusion.com/minisite/en/2022eu-value-program
- FusionServer Pro V5 datasheet, September 2020, and Huawei press release on 1st/2nd Gen Xeon Scalable V5 platforms — https://www.huawei.com/en/news/2020/3/intel-huawei-processor
- xFusion support portal, product documentation and lifecycle bulletins — https://www.xfusion.com/support/
- Internal spec baseline, xFusion rack server V5-V7 reference, checked 2026-10-30 and re-checked 2026-11-04
Xinchuan Server | Enterprise Server Hardware Supplier