Solution Brief
Structured data. No per-token fees.

Documents in.Data out.

On-premises AI extraction on AMD Instinct™ MI350P.Loan, KYC and contract files become validated JSON, on one LIQID UltraStack server inside your data center.

Download PDFTalk to an Expert
4
GPUs
MI350P on one host, room for 30
576
GB
HBM3E on one server
2.2
×
throughput, 2 to 4 GPUs, measured
0
hops
network hops between GPUs

The Re-Keying Dilemma

Keep analysts re-keying documents by hand, or send regulated files to a public AI API.

Manual Pressure

Analysts re-key every field

From PDFs and scans into core systems, by hand.

Slow and error-prone

Every manual field is a delay and a chance for a mistake.

Cost Pressure

Per-token fees at volume

Public API costs grow with every document you process.

Spikes become budget spikes

Month-end and deal volume drive token bills up.

Data Pressure

Regulated files can’t leave

Loan, KYC and contract files can’t go to a public AI API.

Output feeds core systems

Every record has to arrive as clean, validated fields.

Read It. Extract It. Validate It.

Four MI350P GPUs on one LIQID UltraStack server run a 70B model that reads each document and returns validated JSON (parties, amounts, dates, covenants, identity fields) into your workflow.

  • Data never leaves your environment
  • Fixed capacity replaces per-token API fees
  • Standard PCIe GPUs, open-source software stack
GPU allocation for document extraction: Llama 3.3 70B (FP8) with validated JSON output. 2 GPUs for extraction, 1 for triage and classification, 1 for QA check, on a LIQID GPU chassis with 4 AMD Instinct MI350P and 576 GB HBM3E, connected over PCIe Gen 5 to one host server with one OS image that grows to 30 GPUs.

Extraction · 2 GPUs

Llama 3.3 70B replicas read each document and return validated JSON.

~7,800
docs/hr on 4 GPUs*
~2.2
docs/sec on 4 GPUs*

Triage + Classify · 1 GPU

Sorts incoming loan, KYC and contract files before extraction.

1 GPU
per 70B model at FP8
144 GB
HBM3E per MI350P

QA Check · 1 GPU

Llama 3.3 70B checks each record against the source before it posts.

~18
interactive users*
552
tok/s on 4 GPUs
Measured: LIQID Phase-3 testing, 72B-class model on MI350P. *Illustrative; assumptions on page 2.
LIQID UltraStack with AMD MI350P for Financial Services

Proof: Measured on MI350P

No re-keying, no public API: one server turns regulated documents into structured data.

Result
What was measured
Why it matters here
552 tok/s
on 4 GPUs
72B-class model, FP8, under interactive limits: P99 TTFT ≤ 2 s, ≥ 30 tok/s per user
Interactive extraction on the 4-GPU configuration
Near-linear
2 → 4 → 8 GPUs
254 tok/s on 2 GPUs, 552 on 4, 1,256 on 8, one replica per GPU
Add GPUs and extraction throughput grows with them
1 GPU
per 70B model
Llama 3.3 70B is ~70 GB at FP8 and fits on one 144 GB MI350P with room for cache
No multi-GPU model splitting, so every GPU is its own replica
What it means
Every document becomes data
Analysts check exceptions instead of re-keying fields
*Illustrative. Assumes 1,024 input and 256 output tokens per document; ~18 = concurrent interactive users. 72B-class: FP8, one replica per GPU. Source: LIQID Phase-3 testing. Llama 3.3 70B performance is projected from same-class 72B dense-model results and has not been measured directly. Results vary with document length and output size.

Grow on the Same Server: 4 to 30 GPUs

Add GPUs and a new job, not a new cluster.

Start here
4
GPUs
576 GB HBM3E

Regulated document extraction

Loan, KYC and contract files into validated JSON.

Model layout

2 × Llama 3.3 70B extraction · 1 × triage and classification · 1 × QA check

Built for

Lending, KYC and contract operations. Re-keying backlogs; documents that can’t go to a public API.

10
GPUs
1.44 TB HBM3E

Financial crimes investigation agent

Triage, evidence, narrative and QA for every alert.

Model layout

4 × Llama 3.1 8B triage · 4 × Llama 3.3 70B evidence and narrative · 2 × gpt-oss-120b typology and QA

Who buys, why now

Head of Financial Crimes / BSA Officer. Alert backlog, exam findings, investigator attrition.

20
GPUs
2.88 TB HBM3E

Operations and compliance suite

A department’s jobs on one server: triage, loan and KYC extraction, complaints, credit memos and QA.

Model layout

4 × Llama 3.1 8B triage · 6 × Llama 3.3 70B extraction · 2 × Mistral Small 3.2 complaints · 4 × 70B credit memos · 4 × gpt-oss-120b review and QA

Who buys, why now

COO / Head of Operations. Several pilots competing for GPUs; API bills growing.

At 30 GPUs, a bank-wide private model hub runs every line of business on the same host. See the Financial Crimes Investigation brief for the full 10, 20 and 30-GPU path.

What UltraStack Adds

  • One model per GPU: 144 GB HBM3E holds a full 70B model at FP8, or gpt-oss-120b.
  • No network between GPUs: peer-to-peer over PCIe Gen 5 to one EPYC host.
  • Air-cooled: about 22 kW for the full 30-GPU system.

Start Fixed, Grow on One Host

  • One host: one OS image, one license footprint, one thing to patch.
  • Room to grow: 4 to 10 to 20 to 30 GPUs on the same host.

What Is Proven Today

  • P2P bandwidth saturates the link; RCCL collectives pass, 1 to 10 GPUs.
  • ROCm integration complete, developed jointly with AMD.
  • 200+ production deployments, 85+ patents.
The Efficiency Case · Targets, not measurements*
75%
Up to 75% fewer servers
2.1×
Up to 2.1× more tokens per dollar
1.8×
Up to 1.8× better tokens per watt
50%
Up to 50% lower power

Bring Us Three Numbers

Two questions settle the size of the opportunity: how many documents do your analysts re-key each day, and what does each one cost?

Size Your VolumeContact UsDownload PDF
1
Volume:

documents per day and the peak backlog.

2
Document size:

pages and fields per file.

3
Concurrency:

users and systems calling extraction at peak.

Built for

Loan files · KYC files · Contract files · Validated JSON output

Proven on

AMD Instinct™ MI350P · AMD EPYC™ 9005 · ROCm™ · 200+ LIQID deployments

*Efficiency targets compare against multi-server deployments of the same GPUs; they are LIQID internal projections using AMD published specifications, and results vary by configuration and workload.

Performance figures for the AMD Instinct MI350P are based on AMD published specifications and LIQID Phase-3 testing from 1 to 20 GPUs, and are subject to change. 30-GPU figures are projected from measured 20-GPU scaling. Every layout runs one model per GPU; 20- and 30-GPU layouts are sizing arithmetic on published model sizes, not benchmark results. Models shown are US and EU open-weight models. Docs/sec, docs/hr and user figures are illustrative and vary with document length and output size.

Speak With An Expert