George Hotz's tiny corp Packs 14 AMD GPUs Into a $50,000 red whale
Tiny Corp's new $50,000 tinybox red whale packs 14 AMD Radeon AI Pro R9700 cards and 448 GB of GPU memory into a dual-PSU desk chassis.
- Tiny corp launched the tinybox red whale at $50,000 with 14x AMD R9700 GPUs.
- Total GPU memory hits 448 GB GDDR6 with roughly 8.96 TB/s aggregate bandwidth.
- Ships in the tall dual-PSU chassis also used for the upcoming 4x MI300X build.
- That case is now selectable in the configurator for custom builds.
- Lineup follows red flash (6 GPU, $27k) and red kimi ($30k).
- Tiny corp recently dropped all NVIDIA options from its store, going AMD-only.
- Made to order, 1-8 week lead time, bank transfer payment only.
George Hotz’s tiny corp has added the tinybox red whale to its online configurator. The made-to-order system combines 14 AMD Radeon AI Pro R9700 GPUs with 448 GB of aggregate GPU memory. It uses the tall, dual-power-supply chassis designed for tiny corp’s forthcoming four-GPU MI300X system.
The launch caps an AMD-only product range. Tiny corp recently removed its NVIDIA builds, including the tinybox pro, in a commit titled “rip nvidia.” The company has not published a detailed rationale for the change, but its current lineup centers on Radeon workstation cards and the tinygrad software stack.
Fourteen GPUs add up, with limits
Each Radeon AI Pro R9700 provides 32 GB of GDDR6 memory, 640 GB/s of memory bandwidth, and a 300 W total board power rating. Across 14 cards, the red whale’s headline specifications are:
- 448 GB of aggregate GDDR6 memory
- 8.96 TB/s of theoretical aggregate memory bandwidth
- 4.2 kW of GPU board power before accounting for the rest of the system
- Full-fabric PCIe connectivity, according to tiny corp’s specification
Each card exposes its own 32 GB memory allocation, so software must shard models and workloads across the GPUs. The 8.96 TB/s figure adds the local bandwidth of all 14 cards operating in parallel; it is not the bandwidth of a shared memory pool. Cross-GPU transfers still depend on PCIe, which offers less bandwidth than accelerator fabrics such as NVLink or Infinity Fabric.
A 400-billion-parameter checkpoint stored at one byte per parameter occupies about 400 GB, while a 200-billion-parameter bf16 checkpoint requires roughly the same capacity. Both nominally fit across the cards, leaving about 48 GB for KV cache, activations, runtime buffers, and other overhead. Practical limits will vary with quantization, context length, batch size, model architecture, and the quality of the sharding implementation.
The MI300X case makes room
Fitting 14 dual-slot cards drives the enclosure choice. Earlier red models use a shorter chassis built around conventional workstation layouts, while the whale adopts the tall case developed for the four-GPU MI300X configuration. Tiny corp has also exposed that enclosure as an option for custom builds.
The company’s hardware guide describes a reference platform with a water-cooled 32-core AMD EPYC processor, a server motherboard with baseboard management controller support, and Noctua fans. The R9700 configuration raises the published noise ceiling from below 50 dB to below 60 dB, with tiny corp describing the cards’ blower fans as “audible.”
Fourteen GPUs consume up to 4.2 kW before the processor, memory, storage, cooling, and conversion losses enter the calculation. Deployments therefore need electrical capacity and heat removal suited to a system drawing several kilowatts continuously.
A four-step Radeon lineup
Tiny corp now lists four R9700 configurations spanning 128 GB to 448 GB of aggregate GPU memory:
| Model | R9700 GPUs | GPU memory | Key configuration | Listed price |
|---|---|---|---|---|
| red | 4 | 128 GB | Base R9700 system | About $15,000 |
| red flash | 6 | 192 GB | Six-GPU configuration | $27,000 |
| red kimi | 4 | 128 GB | 16 TB NVMe array | $30,000 |
| red whale | 14 | 448 GB | Tall dual-PSU chassis | $50,000 |
The red kimi uses sixteen 1 TB NVMe drives to serve models that exceed its 128 GB of GPU memory. Tiny corp advertises 238.4 GB/s from the SSD array and uses a tinygrad driver to let the GPUs read model data directly. The company claims 40 tokens per second for Moonshot’s 2.8-trillion-parameter Kimi K3 by retaining frequently used weights in VRAM and fetching the remainder from storage. Comparisons with other systems also require the batch size, quantization, context length, and benchmark method.
Memory capacity drives the pitch
At its $50,000 list price, the red whale costs about $112 per gigabyte of GPU memory before shipping, taxes, duties, and infrastructure. That ratio compares favorably with H100-class servers built to reach similar memory capacity, though it says little about compute throughput, power efficiency, interconnect performance, support, or software compatibility.
Inference and batched evaluation for large dense models are the clearest use cases, particularly when the model can be sharded cleanly across PCIe-connected GPUs. Training workloads that depend on frequent collective communication face tighter constraints, and software built around CUDA, NCCL, or NVLink may require substantial changes for the R9700 and tinygrad environment.
Tiny corp sells the tinybox on a made-to-order basis with a stated lead time of one to eight weeks. Orders require bank transfer, and buyers remain responsible for applicable restocking fees, shipping costs, and import duties.
Comments
