AI & Computing
These AI Workstations Look Like PCs, but Pack a Stronger Punch
Mar 24, 2026, 2:00 PM
出典: IEEE Spectrum AI
<img src="https://spectrum.ieee.org/media-library/a-sleek-illuminated-box-resembling-a-personal-computer-workstation.jpg?id=65329122&width=1245&height=700&coordinates=0%2C469%2C0%2C469"/><br/><br/><p>The rise of generative AI has spurred demand for AI workstations that can run or train models on local hardware. Yet modern PCs <a href="https://spectrum.ieee.org/ai-models-locally" target="_self">have proven inadequate for this task</a>. A typical laptop only has enough memory to load a large language model (LLM) with 8 to 13 billion parameters—many times smaller, and much less intelligent, than frontier models that are presumed to have over a trillion parameters. Even the most capable workstation PCs struggle to serve LLMs with more than 70 billion parameters.</p><p><a href="https://tenstorrent.com/waitlist/tt-quietbox" rel="noopener noreferrer" target="_blank">Tenstorrent’s QuietBox 2</a> is an attempt to fill that gap. Though it looks like a PC workstation, the QuietBox 2 contains four of the company’s custom Blackhole AI accelerators, 128 gigabytes of GDDR6 memory—specialized memory used in GPUs—and 256 GB of DDR5 system memory (for a total of 384 GB). This configuration provides enough memory to load OpenAI’s GPT-OSS-120B and can run <a href="https://spectrum.ieee.org/meta-llama-3" target="_blank">mid-sized models</a> like Meta’s Llama 3.1 70B at speeds of nearly 500 tokens per second. For reference, that’s several times quicker than an average response from OpenAI’s GPT-5.2 or Anthropic’s Claude 4.6. The QuietBox 2 carries an expected retail price of US $9,999 and is slated to launch in the second quarter of 2026. </p><p>“The 128 gigabytes of GDDR that we have with our AI accelerators really defines how big of a model you can run at a reasonable speed,” says <a href="https://www.linkedin.com/in/milosstrajkovic/?originalSubdomain=ca" rel="noopener noreferrer" target="_blank">Milos Trajkovic</a>, co-founder and systems engineer at Tenstorrent. “Our 128 gigabytes of GDDR6 RAM would require four Nvidia RTX 5090 graphics cards. That couldn’t fit in today’s 1600-watt form factor, and the cost for four RTX 5090 GPUs is huge.”</p><h2>An AI workstation built at the home office</h2><p>Wattage, it turns out, is critical. Nvidia recommends a system power of 1,000 watts for a single RTX 5090, so even a dual-GPU setup exceeds the continuous power draw for a typical 15-amp, 120-volt power circuit. A system with four RTX 5090s could require 4,000 watts or more at load.</p><p><span>The QuietBox 2, on the other hand, draws only 1,400 watts at full load. It won’t trip the breaker, so it can be used anywhere a typical desktop PC might be plugged in, including a home office.</span></p><p>That’s not the only way the QuietBox 2 poses as a run-of-the-mill PC. The machine’s custom case is built to support the micro-ATX motherboard form factor, and the motherboard itself is an AMD chipset hosting an AMD CPU. The hardware is kept cool by closed-loop liquid cooling similar to that used by PC workstations and gaming computers. It even has customizable RGB LED lighting and a large semi-transparent window that shows off the hardware. </p><p>“A lot of even our internal developers have requested a QuietBox because they’re just so easy to deploy,” says <a href="https://www.linkedin.com/in/chris-goulet-9ab67372/?originalSubdomain=ca" target="_blank">Chris Goulet</a>, a thermal-mechanical engineer and team lead at Tenstorrent. “You just ship them the unit, they slap it on their desk, power it up, and they’re going.” </p><p>Where the QuietBox 2 differs from desktop PCs, though, is its AI accelerators. It’s equipped with four of Tenstorrent’s Blackhole <strong></strong>application-specific ICs, a RISC-V chip designed specifically for AI workloads. Blackhole is packaged on an add-in card; each card has 120 Tensix AI accelerators and 32 GB of GDDR6 memory, for a total of 480 Tensix AI accelerators and 128 GB of GDDR6. Blackhole also has a large amount of on-chip SRAM at 180 MB per accelerator.</p><h2>Two visions of desktop AI</h2><p>Tenstorrent is not alone in its approach. Nvidia’s DGX Spark, released last year, packed Nvidia’s GB10 chip into a machine the size of a lunch box. Orders for the Spark’s big brother, the <a href="https://www.nvidia.com/en-us/products/workstations/dgx-station/" target="_blank">DGX Station</a> with Nvidia’s GB300, were opened on 16 March 2026. </p><p>The DGX Station looks like a desktop PC workstation, and variants will be built by well-known PC brands like Dell and Asus. Nvidia’s offering has more memory than QuietBox 2, at up to 748 GB, but system power is quoted at 1,600 watts—rather close to the maximum a 15-amp, 120-volt breaker will handle. This reflects differing visions for how their machines will be used. And, of course, the Nvidia DGX Station’s extra memory doesn’t come cheap. While most DGX Station system builders have not yet announced pricing, <a href="https://www.centralcomputer.com/msi-ct60-s8060-nvidia-dgx-station-cpu-memory-up-to-496gb-lpddr5x-nvidia-blackwell-ultra-gpu-1x-10-gbe-2x-400-gbe.html">one retailer has listed a DGX Station from PC maker MSI for $85,000</a>. </p><p>When I spoke to <a href="https://www.linkedin.com/in/allenbourgoyne/" target="_blank">Allyn Bourgoyne</a>, director of product marketing at Nvidia, after the announcement of DGX Spark and Station in 2025, he said the company expects most DGX owners will use the devices as remotely accessed workstations. “A common thing you might see is that I’ve got my Windows laptop, and I’m going to use my DGX Spark over the network. I’m going to send jobs over to it.” He added that companies could deploy DGX Spark and Station systems to serve multiple people at once. <strong></strong></p><p>The Tenstorrent QuietBox 2 can be used in this way, but the company also wants to target a good experience for people going one-on-one with the computer. “You don’t have to remote SSH into the box. You connect your monitor through HDMI, and it’s just like your PC at home. It has the Ubuntu desktop and utilities,” says Trajkovic.</p><p>Nvidia’s DGX systems also run a variant of Ubuntu (DGX OS) and include a desktop environment, <a href="https://www.jeffgeerling.com/blog/2025/dells-version-dgx-spark-fixes-pain-points/" target="_blank">but the devil is in the details</a>. DGX systems use Nvidia CPUs based on ARM architectures and custom chipsets. The QuietBox 2 uses an AMD x86 CPU and compatible chipset, and is configured more like a traditional PC. That should be a boon for the QuietBox 2’s software compatibility. </p><p>Tenstorrent leans into that with a focus on open source software. The QuietBox 2’s entire software stack, from TT-Forge (the company’s AI compiler) to TT-Metalium (a low-level software development kit that provides kernel-level hardware control), is open source and available on GitHub. Tenstorrent has also published the instruction set architecture for its Tensix cores, so developers can see exactly how their workloads execute on the hardware. Nvidia, by contrast, is focused on its proprietary CUDA ecosystem, and DGX OS is not open source. </p><p>“A lot of our software stack is completely open, and we felt that from a hardware perspective, we kind of wanted to take a similar path,” says Goulet. </p>
Details
The rise of generative AI has spurred demand for AI workstations that can run or train models on local hardware. Yet modern PCs have proven inadequate for this task. A typical laptop only has enough memory to load a large language model (LLM) with 8 to 13 billion parameters—many times smaller, and much less intelligent, than frontier models that are presumed to have over a trillion parameters. Even the most capable workstation PCs struggle to serve LLMs with more than 70 billion parameters.
Tenstorrent’s QuietBox 2 is an attempt to fill that gap. Though it looks like a PC workstation, the QuietBox 2 contains four of the company’s custom Blackhole AI accelerators, 128 gigabytes of GDDR6 memory—specialized memory used in GPUs—and 256 GB of DDR5 system memory (for a total of 384 GB). This configuration provides enough memory to load OpenAI’s GPT-OSS-120B and can run mid-sized models like Meta’s Llama 3.1 70B at speeds of nearly 500 tokens per second. For reference, that’s several times quicker than an average response from OpenAI’s GPT-5.2 or Anthropic’s Claude 4.6. The QuietBox 2 carries an expected retail price of US $9,999 and is slated to launch in the second quarter of 2026.
“The 128 gigabytes of GDDR that we have with our AI accelerators really defines how big of a model you can run at a reasonable speed,” says Milos Trajkovic, co-founder and systems engineer at Tenstorrent. “Our 128 gigabytes of GDDR6 RAM would require four Nvidia RTX 5090 graphics cards. That couldn’t fit in today’s 1600-watt form factor, and the cost for four RTX 5090 GPUs is huge.”
An AI workstation built at the home office
Wattage, it turns out, is critical. Nvidia recommends a system power of 1,000 watts for a single RTX 5090, so even a dual-GPU setup exceeds the continuous power draw for a typical 15-amp, 120-volt power circuit. A system with four RTX 5090s could require 4,000 watts or more at load.
The QuietBox 2, on the other hand, draws only 1,400 watts at full load. It won’t trip the breaker, so it can be used anywhere a typical desktop PC might be plugged in, including a home office.
That’s not the only way the QuietBox 2 poses as a run-of-the-mill PC. The machine’s custom case is built to support the micro-ATX motherboard form factor, and the motherboard itself is an AMD chipset hosting an AMD CPU. The hardware is kept cool by closed-loop liquid cooling similar to that used by PC workstations and gaming computers. It even has customizable RGB LED lighting and a large semi-transparent window that shows off the hardware.
“A lot of even our internal developers have requested a QuietBox because they’re just so easy to deploy,” says Chris Goulet, a thermal-mechanical engineer and team lead at Tenstorrent. “You just ship them the unit, they slap it on their desk, power it up, and they’re going.”
Where the QuietBox 2 differs from desktop PCs, though, is its AI accelerators. It’s equipped with four of Tenstorrent’s Blackhole application-specific ICs, a RISC-V chip designed specifically for AI workloads. Blackhole is packaged on an add-in card; each card has 120 Tensix AI accelerators and 32 GB of GDDR6 memory, for a total of 480 Tensix AI accelerators and 128 GB of GDDR6. Blackhole also has a large amount of on-chip SRAM at 180 MB per accelerator.
Two visions of desktop AI
Tenstorrent is not alone in its approach. Nvidia’s DGX Spark, released last year, packed Nvidia’s GB10 chip into a machine the size of a lunch box. Orders for the Spark’s big brother, the DGX Station with Nvidia’s GB300, were opened on 16 March 2026.
The DGX Station looks like a desktop PC workstation, and variants will be built by well-known PC brands like Dell and Asus. Nvidia’s offering has more memory than QuietBox 2, at up to 748 GB, but system power is quoted at 1,600 watts—rather close to the maximum a 15-amp, 120-volt breaker will handle. This reflects differing visions for how their machines will be used. And, of course, the Nvidia DGX Station’s extra memory doesn’t come cheap. While most DGX Station system builders have not yet announced pricing, one retailer has listed a DGX Station from PC maker MSI for $85,000.
When I spoke to Allyn Bourgoyne, director of product marketing at Nvidia, after the announcement of DGX Spark and Station in 2025, he said the company expects most DGX owners will use the devices as remotely accessed workstations. “A common thing you might see is that I’ve got my Windows laptop, and I’m going to use my DGX Spark over the network. I’m going to send jobs over to it.” He added that companies could deploy DGX Spark and Station systems to serve multiple people at once.
The Tenstorrent QuietBox 2 can be used in this way, but the company also wants to target a good experience for people going one-on-one with the computer. “You don’t have to remote SSH into the box. You connect your monitor through HDMI, and it’s just like your PC at home. It has the Ubuntu desktop and utilities,” says Trajkovic.
Nvidia’s DGX systems also run a variant of Ubuntu (DGX OS) and include a desktop environment, but the devil is in the details. DGX systems use Nvidia CPUs based on ARM architectures and custom chipsets. The QuietBox 2 uses an AMD x86 CPU and compatible chipset, and is configured more like a traditional PC. That should be a boon for the QuietBox 2’s software compatibility.
Tenstorrent leans into that with a focus on open source software. The QuietBox 2’s entire software stack, from TT-Forge (the company’s AI compiler) to TT-Metalium (a low-level software development kit that provides kernel-level hardware control), is open source and available on GitHub. Tenstorrent has also published the instruction set architecture for its Tensix cores, so developers can see exactly how their workloads execute on the hardware. Nvidia, by contrast, is focused on its proprietary CUDA ecosystem, and DGX OS is not open source.
“A lot of our software stack is completely open, and we felt that from a hardware perspective, we kind of wanted to take a similar path,” says Goulet.
Related Knowledge
mentions
Large Language Models (LLMs)
LLMs are a class of AI models that are trained on vast amounts of text data to understand and generate human-like language. They leverage deep learning techniques to predict the next word in a sentence, enabling them to perform various language-related tasks such as translation, summarization, and conversation.
mentions
AI Accelerators
AI accelerators are specialized hardware designed to efficiently perform the computations required for artificial intelligence tasks, particularly deep learning. These devices, which include GPUs, TPUs, and custom ASICs, are optimized for parallel processing and can significantly speed up the training and inference of machine learning models.