AI Hardware
Reliable. Customised. Optimised.
Aliencode Tech builds and supplies deep learning workstations for machine learning research and AI development, including VRAM-modded GPUs that are difficult to source and validate locally.
Available now
Which GPUs do we sell?
Our own VRAM-modded cards are the RTX 4090 48GB and the RTX 4080 and 4080 SUPER 32GB, all of them factory NVIDIA boards reworked with additional memory rather than NVIDIA product lines. We also source and supply the standard NVIDIA range, consumer GeForce through professional RTX and datacentre accelerators, either as bare cards or installed in a complete workstation we specify and build. Availability and pricing move with the market, so everything is quoted rather than listed.
Flagship
RTX 4090 48GB
A factory NVIDIA GeForce RTX 4090 board reworked with double the VRAM. Identical silicon to the 24GB card (16,384 CUDA cores, 2520 MHz boost, 384-bit bus), with 48GB of GDDR6X and 1.01 TB/s of bandwidth. Two fit in a standard tower.
- Stock NVIDIA drivers, stock CUDA, no patched runtime
- Windows and Linux both fully supported
- Also available in quiet triple-fan and AIO versions
Also available
RTX 4080 / SUPER 32GB
Reworked RTX 4080 and 4080 SUPER boards with double the VRAM, for workloads that need more memory than a standard 16GB card allows but do not need a 4090’s power envelope. A practical fit for mid-size model fine-tuning and inference.
- 32GB VRAM in a standard consumer power budget
- Supplied bare or installed in a complete build
Everything else
Every other part
Not every workload wants a modded card. We supply standard NVIDIA GPUs across the range, consumer GeForce through professional RTX workstation cards and datacentre accelerators, alongside every other component that goes into the machine. Take a complete build, a bare card, or just the parts you are short of.
- GPUs, consumer to datacentre
- CPUs and server-grade motherboards
- ECC memory and NVMe storage
- Air cooling, case fans and CPU coolers
- Power supplies and chassis
- OEM rack server chassis
- Assembly, burn-in and OS setup
Comparison
RTX 4090 48GB vs RTX 6000 Ada: which is faster?
In sustained deep learning workloads, the RTX 4090 48GB is faster. The RTX 6000 Ada has more CUDA and Tensor cores on paper, but its 300 W power limit stops it reaching its advertised throughput under real load, a gap that short synthetic benchmarks hide.
| Specification | RTX 4090 48GB | RTX 6000 Ada |
|---|---|---|
| Base clock | 2235 MHz | 915 MHz |
| Boost clock | 2520 MHz | 2505 MHz |
| Memory size | 48 GB | 48 GB |
| Memory type | GDDR6X | GDDR6 |
| Memory bus | 384-bit | 384-bit |
| Memory bandwidth | 1.01 TB/s | 960 GB/s |
| CUDA cores | 16,384 | 18,176 |
| Tensor cores | 512 | 568 |
| CUDA compute capability | 8.9 | 8.9 |
| TDP | 450 W | 300 W |
Note the RTX 6000 Ada leads on core counts. That is exactly the point: core count without the power budget to feed it does not turn into throughput.
Real-life benchmarks
How much faster is it in practice?
Between 11.6% and 24.6% faster, depending on the workload. These are application-level results on real models rather than synthetic FLOPS, measured by Aliencode Tech on identical host configurations.
Benchmarked on pytorch-labs/gpt-fast. Token generation is bottlenecked by memory bandwidth.
Benchmarked on aredden/flux-fp8-api.
Benchmarked on aredden/flux-fp8-api.
| Benchmark | Metric | RTX 4090 48GB | RTX 6000 Ada | Advantage |
|---|---|---|---|---|
| Llama 2 13B (gpt-fast) | Tokens / second | 36.22 | 32.45 | +11.6% |
| FLUX.1 dev, 1024×720 | Iterations / second | 3.36 | 2.80 | +20.0% |
| FLUX.1 dev, 1024×1024 | Iterations / second | 4.71 | 3.78 | +24.6% |
Benchmarks conducted by Aliencode Tech Solutions using pytorch-labs/gpt-fast and aredden/flux-fp8-api, 2025. Separately, MLCommons MLPerf Inference (Datacenter) v5.0 records 2× RTX 4090 at 8.29 samples/s on 3D-UNET against 2× NVIDIA L40S at 7.69 samples/s, and the L40S carries a 50 W higher power limit than the RTX 6000 Ada.
Built by us
What a finished machine looks like.
All workstations are made with utmost craftsmanship and rigorous burn-in testing. We pick the highest quality components and leave ample headroom for future upgrades. Select any image to enlarge it.
Clients
Delivered for
Singapore Eye Research InstituteSERI
Supplied a dual-GPU deep learning workstation with the RTX 4090 48GB for machine learning research, achieving great cost-savings.
FAQ
AI hardware: FAQ
- What is the RTX 4090 48GB?
- The RTX 4090 48GB is a factory NVIDIA RTX 4090 board reworked with double the standard VRAM. It is not an NVIDIA product line: the base card is NVIDIA’s, the memory rework is not. Its specification is otherwise identical to a 24GB RTX 4090 (same 16,384 CUDA cores, same 2520 MHz boost clock, same 384-bit bus), but with 48GB of GDDR6X and 1.01 TB/s of memory bandwidth.
- Why buy a VRAM-modded GPU instead of a server-grade card?
- Cost and speed, without giving up reliability. A modded card is a factory NVIDIA board reworked with higher-capacity memory of the same grade, then burned in under sustained load before it ships, so it behaves like the card it started as. It runs on stock NVIDIA drivers and stock CUDA on both Windows and Linux, with no patched runtime, so nothing in your stack has to change to use it. It presents as a 4090 with 48GB of VRAM, which is exactly what you want your framework’s allocator to see. What changes is the price and the throughput: 48GB of VRAM costs a fraction of a datacentre card carrying the same memory, and in sustained deep learning work the RTX 4090 48GB outruns an RTX 6000 Ada, whose 300 W power limit stops it reaching its advertised numbers.
- Does the RTX 4090 48GB need special drivers or a patched CUDA?
- No. It runs on stock NVIDIA drivers with stock CUDA, on both Windows and Linux, and works with any NVIDIA software out of the box exactly as a standard RTX 4090 does. There are no additional installation steps and no restricted operating system support.
- Is the RTX 4090 48GB faster than the RTX 6000 Ada?
- In real deep learning workloads, yes. On paper the RTX 6000 Ada has more CUDA and Tensor cores, but its 300 W power limit prevents it from reaching its advertised throughput in sustained work. In our benchmarking the RTX 4090 48GB was 11.6% faster on Llama 2 13B token generation and 20 to 25% faster on FLUX.1 image generation.
- I also see an RTX 4090D 48GB, what is the difference?
- The RTX 4090D 48GB is a rework of the China-only RTX 4090D board, which carries fewer Tensor cores for AI FLOPS and downgraded VRAM, lowering memory bandwidth and with it performance in bandwidth-bottlenecked workloads such as LLM inference. It also uses GDDR6 rather than GDDR6X, which is significantly slower again. That is why it looks cheaper. We supply both, and we recommend the full-powered version.
- How much does an AI workstation cost?
- Price depends on the GPU configuration, CPU platform, memory capacity and storage, and moves with GPU market pricing, so we quote per build rather than publishing a list price. Tell us your workload and budget and we will specify against it. We accept multiple payment currencies and methods.
- Why air cooling rather than a liquid AIO cooler?
- Liquid all-in-one coolers carry a leak risk directly above the most expensive components in the machine, and deep learning does not need CPU overclocking. A well-chosen air cooler with high-airflow case fans handles sustained load quietly and has nothing to fail wet.
- What do you preinstall on the workstation?
- Ubuntu with the CUDA Toolkit installed and validated against the GPUs, so the machine trains on arrival rather than after a week of driver debugging. We can match a specific Ubuntu and CUDA version to your existing environment on request.
- Can you specify a build for our specific workload?
- Yes, and we prefer to. Image-based training pipelines are usually CPU-bottlenecked in preprocessing and benefit from single-core speed and memory bandwidth; large-model inference is bound by VRAM and memory bandwidth. Tell us what you actually run and the specification follows from that.
- Do you supply hardware for on-premise LLM deployment?
- Yes. Where a corpus cannot leave your network, we build the workstation or server and deploy open-weight models onto it, so inference runs entirely inside your own infrastructure. See our LLM and agentic AI solutions for the software side.
Get in touch
Tell us what you train.
Send us the model sizes and batch shapes you actually run and we will specify against them.
