How much VRAM do I need?

The number comes from the largest model you want to run. Weights at 4-bit are roughly half a byte per parameter, so an 8B model is around 5GB before you add context; a 70B is around 40GB. Training is a different budget again — gradients and optimiser states can multiply the weight footprint several times over, which is why people fine-tune with LoRA instead. Sizing from a named model and precision is far more reliable than sizing from a budget.

One big GPU or two smaller ones?

A single larger card is the right answer almost every time, unless there is a specific reason otherwise. It's simpler, it holds the whole model in one pool of memory, and nothing has to be synced. Two cards become the better choice when you're training data-parallel, running independent jobs at once, or serving more than one model. Two cards do not make a single model that already fits run any faster.

Can I fine-tune on this, or do I need cloud?

You can fine-tune. LoRA and QLoRA on one strong card is realistic and it's where most of the value sits for a small team — you're training a small set of adapter weights rather than the whole model, so the memory requirement is a fraction of a full fine-tune. Pre-training a large model from scratch is still a cluster job, and that capacity is better rented than owned.

Do the GPUs need NVLink?

Usually not. Data-parallel training syncs gradients once per step and PCIe handles that load at workstation scale. The interconnect only becomes the bottleneck when one model is sharded across cards so they exchange activations constantly — a cluster pattern more than a workstation one.

What about MATLAB, CFD and other simulation work?

Simulation has a different profile. That work depends on core count, memory bandwidth and RAM capacity, and many solvers get nothing from a GPU at all. It's worth checking whether your specific solver has GPU support before paying for one, and reading the licence terms — per-core licensing can make a very high core count an expensive way to run a job that stops scaling at sixteen.

Can the team share it, or does it have to sit on my desk?

A team can share it. These machines run headless: teams SSH in, serve models over the network, run notebooks remotely and keep jobs in containers. A single workstation under one desk regularly serves a whole team's inference. When it needs to be rack-mounted and run around the clock instead, that's SLATE, and our corporate solutions team can spec it with you.

AI workstations, built for training, inference and simulation

The size of the graphics card’s memory decides which models you can run at all. We build machines with enough of it to serve models, fine-tune on your own data and run simulations — no job queue, no hourly billing, and data that never leaves your building. FOCUS MAX for serving a single model; FOCUS X and HYPERFOCUS when training is the wait.

START FROM YOUR SOFTWARE SPEAK TO A SPECIALIST

4.9 STARS ON GOOGLEFrom more than 4,200 customer reviews
3-YEAR WARRANTYExtendable to 4 years
LIFETIME SUPPORTAustralia's largest PC service team
AWARD-WINNING BUILDERSBest Custom Desktop PC Brand 2020–2023
BUILT IN AUSTRALIAAssembled and tested to order
ON-SITE SERVICINGLess downtime — we come to you

Select your software

The machines that run them

Four machines on one ladder. FOCUS MAX is one strong card for models that fit its memory. FOCUS X adds a second when fine-tuning is the wait. HYPERFOCUS takes up to six cards when the machine is the cluster. CREATOR COMPACT is fourteen litres for the development seat. All four can be configured before we build them.

Focus Max

Single-GPU AI

One strong GPU — the cheapest VRAM per dollar
Serves local LLMs and runs LoRA fine-tunes
Cores and RAM sized to keep the GPU fully utilised
NVMe staging for datasets and checkpoints

from AUD$4,199$4,999

You save $800

CONFIGURE YOUR BUILD Enquire Now

Ultrafocus

Dual-GPU training

Two GPUs — split one job, or run two experiments
More total VRAM for larger models and batches
Power and cooling sized for both cards at full load
Data-parallel training scales close to linear

from AUD$10,399$10,900

You save $501

CONFIGURE YOUR BUILD Enquire Now

Hyperfocus

SUITE-GRADE PERFORMANCE. LOW NOISE

Up to four GPUs in one chassis
Sweeps and job queues that clear overnight
Built to hold clocks through runs measured in days
We check the job scales before quoting four cards

from AUD$35,999$37,000

You save $1,001

CONFIGURE YOUR BUILD Enquire Now

NEW

CREATOR COMPACT

Pro level · the development seat

Fourteen litres, one full-length graphics card
Writing, debugging and profiling the job, at the desk
Local inference and small fine-tunes on a single card
Submits the long runs to the machine built for them

from AUD$1,849$2,850

You save $1,001

CONFIGURE YOUR BUILD Enquire Now

Focus and Creator run the same performance at each level — the difference is the chassis and how much of it you want on the desk. The full range shows every model and level side by side.

What each framework needs

Training hits a wall at video memory. Simulation ignores it and asks for cores instead. Here is where each framework puts the load, what it looks like when a machine is short of it, and which model we would start you on.

Tested beyond installation: the environment builds, the model loads and a run completes. Tell us the models and frameworks you use and we will size the card to them.
Program	What it depends on	How it shows up	Start with
PyTorch	VRAM before anything else. The card's memory decides which models fit; compute decides how long they take.	An out-of-memory error rather than a slow run — the job does not start at all.	UltrafocusFine-tuning and parallel sweeps
TensorFlow	VRAM, then the input pipeline. Cores and NVMe keep the card fed.	A GPU sitting idle between batches because data loading cannot keep up.	UltrafocusDistributed training in one box
MATLAB	Cores and memory, not VRAM. Only some toolboxes touch the graphics card at all.	A solve that scales with core count, and large matrices that exhaust the workspace.	Focus MaxCores, not video memory
Ansys	Cores, memory and solver scratch. Licensing is often per core, so the spec and the licence have to be decided together.	FEA and CFD models that will not fit in memory, and solves that run overnight either way.	HyperfocusCheck your licensing first
Running models locally	The card's memory is the ceiling. Quantised weights lower it; a longer context window raises it again.	A model that loads and runs on one machine and refuses to load on another with a faster card.	Size VRAM to the modelNot to the benchmark

Where your work sits

Where your work sits matters as much as which framework you open. Running published models locally is bounded by the card's memory alone, and a single strong GPU covers it.

The professional load

Fine-tuning on your own data is where a second card stops the waiting. Multi-GPU training and sweeps is the level at which owning the machine starts costing less than renting it by the hour.

What makes an edit faster

Five things decide how fast an edit feels. Here they are in the order that matters for post — what each one does, and what you notice when there is not enough of it.

Graphics
VRAM first, speed second. Memory decides whether the job runs at all; the core count only decides how long it takes once it does.
Processor
Loading, decoding, tokenising and augmenting all happen on the CPU. With too few cores, the GPU sits partly idle waiting for the next batch. For solvers, the CPU is the whole job.
Memory
System RAM comfortably above total VRAM is the safe rule. It holds the dataset between epochs, and it's where layers land the moment anything is offloaded off the card.
Storage
Training reads millions of small files in random order, then writes a checkpoint over and over. That pattern is what NVMe is for — the working dataset stages locally, the rest archives.
Power & cooling
A run is hours or days at 100%, not a benchmark burst. Throttling doesn't produce an error — it simply adds hours to every epoch.

Focus Max · Focus X · Hyperfocus · Creator Compact

VRAM is the first hard limit

Everything a run needs lives in the graphics card’s own memory — the model weights, the activations, and if you are training, the gradients and optimiser states on top. If it does not all fit, the job does not start. VRAM is the first number to settle and the last one to compromise on, and it is why FOCUS MAX, FOCUS X and HYPERFOCUS are separated by card memory before anything else.

When a model doesn't fit, there are three workarounds, each with a cost. Quantising to 8-bit or 4-bit gives up a little quality. Offloading layers to system RAM cuts throughput sharply, because data must cross the PCIe bus for every token. Reducing batch size and context length until the job fits slows training and shortens how much the model can see at once.

Every machine on this page is built to meet what TensorFlow, PyTorch and CUDA actually ask for, with components chosen to hold performance through a long run rather than throttle after a short burst. The most useful sizing detail is the largest model you plan to run, at the precision you plan to run it — with that, our corporate solutions team can size the card to the job rather than to a price bracket.

Batch size and context length are VRAM questions. Longer context and bigger batches both cost memory before they cost time.
Training costs far more memory than inference. A model you can serve comfortably can still be out of reach to fine-tune at full precision.
Two cards is more total VRAM, not one big pool. A single model only spans both if your framework shards it — otherwise each card holds its own copy.

Why choose Aftershock PC?

Anyone can order the same parts. What you are really choosing is who builds the machine, who tests it before it ships, and who answers the phone when something goes wrong two years from now. That is where we are different.

4.9 on Google from over 4,200 customer reviews

Best Custom Desktop PC Brand 2020–2023

8 years of workstation builds — universities, studios and practices

Award-winning build quality

Every system is checked as it is built, then stress-tested and fully updated before it leaves us — so problems are found here, not at your desk. It arrives ready to work, hand-foamed for safe transit anywhere in Australia. Four years of national awards say we get that right.

3-year warranty, lifetime support

Full-system cover for three years, extendable to four — the whole machine, not a chain of separate component warranties you have to chase. If something happens in transit or years later, a dedicated team gets you running again. Support carries on for as long as you own it.

On-site servicing available

Add on-site servicing and we come to you. No packing it up, no courier, no waiting while it travels both ways — a technician diagnoses and repairs it where it sits. It is the difference between days of downtime and hours of it.

Expert workstation team

A dedicated workstation team — not a call centre — with eight years of these builds behind it, trusted by universities, studios and practices. Call or email and an Australian expert helps you match the right tool to the job — before you buy, and long after.

The longer answer

What we actually check before a machine is packed, how we specify one in the first place, what happens in year three, and an honest comparison against buying off the shelf or building it yourself.

WHY AFTERSHOCK

Where our machines already work

Liquid particle simulation at the University of Melbourne, a museum installation running on brain-cell computers, and a security system pulling multiple 8K streams at once. Compute that runs unattended and cannot fall over.

Research & simulation

Melbourne University

A Hyperfocus workstation used for advanced liquid particle simulation — modelling how dirt dams respond to real-world conditions, to help predict and prevent catastrophic structural failures before they occur.

Hyperfocus

Real-time art installation

National Communications Museum

Real-time graphics processing alongside Cortical Labs’ brain-cell computers, for an installation that used biological neural cells to interpret and generate visual data. Our hardware ran the live Unreal Engine graphics.

REAL-TIME GRAPHICS

Security infrastructure

PICA

A Hyperfocus workstation running PICA’s security infrastructure — a four-GPU system handling multiple simultaneous 8K streams, consistently and stably, around the clock.

HYPERFOCUS -4 GPU

All eleven case studies