AMD Instinct MI350P · AMD EPYC 9005

Liqid UltraStack with AMD Instinct MI350P
Up to 30 GPUs. One server. One model per GPU.

A complete AI inference system from Liqid, built on AMD Instinct GPUs and AMD EPYC. One dual-socket AMD EPYC 9005 server takes 4 to 30 AMD Instinct MI350P GPUs over Liqid's PCIe Gen5 fabric, so each team's model gets its own card. Capacity grows by adding GPUs, not servers. It is air-cooled throughout and runs AMD ROCm with vLLM and SGLang.

Quote UltraStack MI350P →Check your servers Size a pool in the calculator →

4–30^[src]

MI350P GPUS ON ONE HOST

144 GB^[src]

HBM3E PER GPU

4.32 TB^[src]

HBM3E AT 30 GPUS

69 PF^[src]

FP8 AT 30 GPUS

0^[src]

NETWORK HOPS BETWEEN GPUS

~22 kW^[src]

WHOLE SYSTEM AT 30 GPUS, AIR-COOLED

What is in it

A fixed, tested system, not a kit

Liqid sells the MI350P system as fixed, tested configurations from 4 to 30 GPUs, headlined in four tiers^[src]. Every size uses the same building blocks.

Component	Specification
Host server	Dual-socket AMD EPYC 9005 Series (Dell or Supermicro)
GPUs	4 to 30 × AMD Instinct MI350P PCIe: 144 GB HBM3E, 600 W (configurable to 450 W), air-cooled
GPU memory	576 GB to 4.32 TB HBM3E, about 4 TB/s per GPU
AI performance	9.2 to 69 PFLOPS FP8; up to 138 PFLOPS MXFP4/MXFP6 peak (AMD published)
GPU pooling fabric	Liqid PCIe Gen5 fabric and host bus adapter; peer-to-peer GPU communication
Expansion chassis	Liqid EX-Series 10-slot GPU chassis: 1, 2 or 3 (one per 10 GPUs)
Management	1 × Liqid Director 1U management appliance
Software	AMD ROCm with vLLM and SGLang through AMD Inference Microservices
Storage and networking	Liqid NVMe SSD; Ethernet or InfiniBand NICs (recommended, configurable)
Cooling	Air-cooled; no liquid cooling or facility redesign

Source: Liqid UltraStack with AMD MI350P datasheet, dated 13 Oct 2026. Storage and networking can be customised or omitted.

The 30-GPU configuration: three EX-Series chassis composed to one AMD EPYC 9005 host, with no network hop between GPUs. Datasheet configuration; the Liqid Director appliance carries the Matrix management plane.

Configurations

Four sizes, from one job to a whole enterprise

Liqid names the tiers by who they serve. Any count from 4 to 30 GPUs is available; Liqid sales leads with 10, 20 and 30, and partners start at 4^[src].

	4 GPUs	10 GPUs	20 GPUs	30 GPUs
Tier	One job	Team	Department	Enterprise
Models in parallel	4	10	20	30
Aggregate HBM3E	576 GB	1.44 TB	2.88 TB	4.32 TB
FP8 AI performance	9.2 PF	23 PF	46 PF	69 PF
GPU power (600 W each)	2.4 kW	6 kW	12 kW	18 kW
Liqid 10-slot chassis	1	1	2	3
7B-class tokens/s	18,526 measured	44,383 measured	89,000 measured	~133,000 projected

7B-class throughput: FP4, one model per GPU, concurrency 128, tokens/s, not latency-gated. The 20-GPU figure is given by Liqid as approximately 89,000. Source: MI350P datasheet.

HBM3E, FP8 compute and GPU power by configuration, each bar scaled to its own maximum. Power is the GPUs only; Liqid puts the whole 30-GPU system at about 22 kW. Source: MI350P datasheet.

Scaling

How throughput holds up as GPUs are added

Liqid tested the system from 1 to 20 GPUs on one host and projected the 30-GPU result from that scaling^[src]. Its headline is 9.5× throughput from 1 to 10 GPUs. Each GPU runs its own model, so the cards never split one job between them; that is the setup Liqid measured.

Measured 7B-class throughput (FP4, one model per GPU, concurrency 128) against perfect linear scaling from the single-GPU result. The 30-GPU point is Liqid's projection and is drawn as such. Source: MI350P datasheet, Liqid Phase-3 testing.

Read the chart for what it is: many independent models in parallel, not one large model split across GPUs. If your plan is to shard one very large model or train across all GPUs at once, an NVLink system is the comparison to make; see NVIDIA DGX and our AI servers data study.

Sizing

What fits on each GPU

Liqid's sizing rule is one model per GPU, up to about a 120B-class model on each 144 GB card: for example Llama 3.3 70B at FP8 or gpt-oss-120b at MXFP4^[src].

10 GPUs · Team

Every job one team runs

Liqid's example layout: 4 × Llama 3.1 8B for triage, 4 × Llama 3.3 70B for drafting and 2 × gpt-oss-120b for QA, on 1.44 TB of HBM3E^[src].

20 GPUs · Department

A department’s model suite

Liqid's example: five agents on four GPUs each, such as prior authorisation, appeals, eligibility, claims triage and denials, on 2.88 TB^[src].

30 GPUs · Enterprise

A private model hub

Liqid's example: 10 GPUs of team jobs, 14 for a department suite and 6 for shared reasoning and Q&A, on 4.32 TB^[src].

Liqid notes these layouts are sizing arithmetic, not benchmark results. To check the VRAM a model actually needs at your context length and concurrency, use the AI GPU calculator or read how much VRAM an LLM needs.

Facility

Power and cooling, by size

Each MI350P draws 600 W and can be capped at 450 W. Liqid's figures for planning^[src]: about 2.4 kW of GPU load at 4 GPUs, about 5 kW at 8 (an estimate) and about 22 kW for the whole 30-GPU system. It is air-cooled only, with no liquid-cooling requirement.

At 22 kW one system can fill a rack's power budget on its own. Check the circuit and inlet temperature before ordering. Liqid's 10-slot Gen5 chassis, the EX-5410P, is specified for 0–27 °C operation^[src], lower than most servers; we confirm the exact chassis and its envelope on the quote. Plan the rack in our rack power density planner, then convert the load into cooling with the server room cooling calculator.

Supply & quote

Quote an UltraStack MI350P system

Send three numbers, as Liqid sizes it: the jobs, the model for each job, and peak concurrency.
We map models to GPUs and pick the 4, 10, 20 or 30-GPU configuration, with room to grow in pairs.
We check rack power, circuits and inlet temperature against the configuration.
You get one quote covering the system and the networking around it.

Get a Liqid quote

Tell us what you want to pool and which servers you run. We size it against Liqid's hardware compatibility list and quote the complete build.

Business enquiries only — please use your company email. Gmail, Hotmail and other personal addresses are not accepted.

Delivery country (optional)

We only ship hardware to the countries in the dropdown. Software and maintenance services are global.

We use these details to reply to your enquiry and prepare any quote. No marketing emails without your consent. Privacy Policy

Interested in

Main workload

GPUs to pool

Host servers

Work email *

Name

Company

Phone

Anything else?

No obligation · response within 4 business hours

FAQs

UltraStack with AMD MI350P: common questions

What is Liqid UltraStack with AMD Instinct MI350P?

A fixed, tested AI inference system: one dual-socket AMD EPYC 9005 server (Dell or Supermicro) with 4 to 30 AMD Instinct MI350P PCIe GPUs composed to it over Liqid’s PCIe Gen5 fabric, in one to three Liqid 10-slot GPU chassis. It runs AMD ROCm with vLLM and SGLang, and Liqid designs it to run one model per GPU.

How much GPU memory does it have?

Each MI350P carries 144 GB of HBM3E, so the system ranges from 576 GB at 4 GPUs to 4.32 TB at 30 GPUs. Liqid says one MI350P holds a 70B model at FP8, or gpt-oss-120b at MXFP4: roughly a 120B-class model per GPU. Check your own models with our AI GPU calculator.

Does it need liquid cooling?

No. Liqid specifies it as air-cooled with no facility redesign. Each GPU draws 600 W and can be set to 450 W. GPU power runs from 2.4 kW at 4 GPUs to 18 kW at 30, and Liqid puts the whole 30-GPU system at about 22 kW.

Can I start small and add GPUs later?

Yes. Liqid says you can start at 4 GPUs and grow in pairs up to 30 on the same server, adding GPUs rather than servers: no new network, no rebuild and nothing to migrate. Liqid leads with 10, 20 and 30-GPU configurations, and its partners start at 4.

How fast is it?

Liqid’s Phase-3 testing measured 4,653 tokens/s on one GPU, 18,526 on 4, 44,383 on 10 and about 89,000 on 20 GPUs, for a 7B-class model at FP4 with one model per GPU and 128 concurrent requests. The 30-GPU figure of about 133,000 is Liqid’s projection, not a measurement. Liqid quotes FP8 compute of 9.2 PFLOPS at 4 GPUs up to 69 PFLOPS at 30.

Why one server instead of several GPU servers?

Liqid’s case is fewer operating systems, NICs, switch ports and licences: one host and one OS image, and no network hops between GPUs. Its efficiency figures (up to 75% fewer servers, up to 2.1× tokens per dollar, up to 1.8× tokens per watt and up to 50% lower power, versus multi-server deployments of the same GPUs) are labelled on the datasheet as internal projections, “targets, not measurements”. We model the comparison with your numbers before you commit.

Can Servnet quote the MI350P system?

Yes. Liqid sizes the system from three numbers: the jobs you want on one server, the model for each job, and peak concurrency. Send us the same and we’ll come back with a configuration and a quote, including rack power and cooling.

The Liqid range

More Liqid products we supply and quote

Liqid composable infrastructure →

How GPU, memory and NVMe pooling works, and where it pays.

Liqid GPU pooling: SmartStack & UltraStack →

Up to 30 GPUs to one server, or shared across up to 16.

Liqid CXL memory pooling →

Up to 20 TB per chassis; up to 100 TB composed to one server.

LIQID Matrix software →

One UI and API for composing GPUs, memory and storage.

Liqid PCIe & CXL fabric hardware →

EX-5410P/C chassis, FX switches and Gen5 HBAs.

Liqid LQD4500 IO Accelerator →

Up to 32 TB NVMe on one PCIe Gen4 x16 card.

Liqid server & GPU compatibility →

Which servers and GPUs Liqid has tested, by vendor.

Keep exploring

Related Servnet guides, tools and hardware

Tool

AI GPU calculator →

Guide

Self-hosting LLMs vs cloud GPUs →

Guide

How much VRAM does an LLM need? →

Research

AI servers: 2026 data study →

Compare

NVIDIA DGX systems →