Colossus MK2 GC200
The Colossus MK2 GC200 is Graphcore's second-generation Intelligence Processing Unit (IPU), a processor architected from the ground up for machine intelligence workloads rather than repurposed graphics rendering.
Publisher review
The Colossus MK2 GC200 is Graphcore's second-generation Intelligence Processing Unit (IPU), a processor architected from the ground up for machine intelligence workloads rather than repurposed graphics rendering. It targets organizations training large-scale AI models—such as natural language processing transformers, graph neural networks, and generative AI—that require massive parallelism and on-chip memory. The GC200 is the core of the IPU-Machine M2000, a 1U blade that delivers 1 petaFLOP of AI compute, and scales to IPU-POD64 systems and beyond, connecting up to 64,000 IPUs for data-center deployments.
Fabricated on TSMC's 7nm process with a die size of 823 mm², the GC200 packs 59.4 billion transistors—more than NVIDIA's A100 GPU (54 billion). It contains 1,472 independent processor cores, each capable of running six parallel threads, for a total of nearly 9,000 simultaneous program threads. Each IPU holds 900MB of In-Processor-Memory (on-die SRAM) with 47.5 TB/s bandwidth, eliminating the traditional GPU memory bottleneck. It delivers 250 teraFLOPS at FP16.16 and FP16.SR (stochastic rounding), supports FP32 compute, and connects via PCIe Gen 4 x16. The IPU-Fabric provides 8 TB/s bandwidth to all IPUs in a system, with 320 GB/s chip-to-chip bandwidth over 10 IPU links.
The GC200 competes directly with NVIDIA's A100 GPU. Per-chip, the IPU offers higher transistor density and on-chip memory (900MB vs. 40MB L2 cache on A100), but the A100 supports larger operand sizes (up to ~72,000×72,000 matrices vs. ~2,688×2,688 on the IPU) due to its 40GB HBM2 memory. In mixed-precision GEMM benchmarks, the GC200 achieves 58.9 TFLOPS (47.3% of theoretical peak) versus the V100's 90.0 TFLOPS (72.0%). The IPU-Machine M2000 (four GC200s) is priced at $32,450, while an IPU-POD64 (16 M2000s) scales to 64 IPUs. Graphcore has raised over $700M from investors including Microsoft, BMW i Ventures, and Dell.
Key trade-offs include limited device memory (900MB per chip vs. 40GB on A100), which restricts the size of models that fit on a single processor without model parallelism. The IPU's specialized architecture requires the Poplar SDK for programming, which has a steeper learning curve than CUDA. Employee reviews on Glassdoor cite limited career progression and concerns about leadership communication. Despite strong theoretical throughput, real-world mixed-precision performance lags behind NVIDIA's TensorCores in dense matrix operations. The GC200 is best suited for workloads that benefit from massive on-chip parallelism and memory bandwidth, such as sparse models and graph neural networks, rather than large dense transformers that require high off-chip memory capacity.
How it works
-
Massive transistor count
59.4 billion transistors on a 823 mm² die using TSMC 7nm process, exceeding NVIDIA A100's 54 billion transistors.
-
1,472 processor cores
Each core runs 6 parallel threads, yielding nearly 9,000 independent program threads per chip.
-
900MB on-chip memory
In-Processor-Memory provides 47.5 TB/s bandwidth, avoiding off-chip memory latency common in GPUs.
-
250 teraFLOPS AI compute
Achieved at FP16.16 and FP16.SR (stochastic rounding), with additional FP32 support.
-
IPU-Fabric interconnect
8 TB/s bandwidth to all IPUs in a system, with 320 GB/s chip-to-chip via 10 IPU links.
-
Scale-out to 64,000 IPUs
IPU-POD solutions connect up to 64,000 IPUs for data-center-scale AI workloads.
-
PCIe Gen 4 x16 support
Standard interface for integration into existing server infrastructure, used in Dell DSS8440 server.
Strengths and trade-offs
Strengths
- Designed exclusively for machine intelligence, not repurposed graphics hardware, enabling specialized parallelism for AI workloads.
- Higher transistor density (59.4 billion) than NVIDIA A100 (54 billion) on a similar die size (823 mm² vs. 826 mm²).
- 8x performance improvement over Graphcore's first-generation IPU (Mk1) for existing models, per company benchmarks.
- 900MB on-chip memory with 47.5 TB/s bandwidth eliminates GPU-style HBM memory bottleneck for many workloads.
Trade-offs
- Limited on-chip memory (900MB) restricts single-chip model size compared to NVIDIA A100's 40GB HBM2, requiring model parallelism for large transformers.
- Mixed-precision GEMM performance (58.9 TFLOPS) lags behind NVIDIA V100 (90.0 TFLOPS), achieving only 47.3% of theoretical peak vs. 72.0%.
- Smaller maximum matrix operand size (2,688×2,688) vs. GPU (72,000×72,000) limits dense linear algebra operations.
- Employee reviews on Glassdoor cite limited career progression and concerns about leadership communication and effectiveness.
Pricing context
IPU-Machine M2000 (4x GC200) priced at $32,450; IPU-POD64 (16 M2000s) available as a rack-scale system; per-chip pricing not disclosed.
Getting started with Colossus MK2 GC200
-
Order an IPU-Machine M2000
Contact Graphcore sales or an authorized reseller to purchase the IPU-Machine M2000, which contains four GC200 IPUs. Pricing starts at $32,450 per unit. Ensure your data center has available rack space and power for the 1U blade.
-
Install the Poplar SDK
Download and install the Poplar SDK from Graphcore's developer portal. The SDK includes the Poplar graph compiler, PopLibs libraries, and tools for programming the GC200. Follow the installation guide for your Linux distribution and verify the installation with the provided test scripts.
-
Connect the IPU-Machine to your server
Mount the IPU-Machine M2000 in your server chassis and connect it via PCIe Gen 4 x16. Ensure the IPU-Fabric cables are properly attached for chip-to-chip communication. Power on the system and confirm the IPUs are detected by running the `gc-monitor` tool.
-
Load and run a sample model
Navigate to the Poplar SDK examples directory and select a sample model, such as a simple neural network or graph neural network. Use the `poprun` command to compile and execute the model on the GC200. Monitor performance with `gc-monitor` to verify correct operation.
-
Scale to an IPU-POD64 system
For larger workloads, deploy an IPU-POD64 by connecting 16 IPU-Machine M2000 units. Configure the IPU-Fabric to link all 64 IPUs. Use the Poplar SDK's distributed training tools to partition your model across the system and run training jobs at scale.
Frequently Asked Questions
What is the Graphcore Colossus MK2 GC200 IPU?
The Colossus MK2 GC200 is Graphcore's second-generation Intelligence Processing Unit (IPU), designed specifically for machine intelligence workloads. It features 59.4 billion transistors on a 7nm TSMC process, 1,472 cores, and 900MB of on-chip memory for massive parallelism in AI tasks.
How does the GC200 compare to NVIDIA's A100 GPU?
The GC200 has more transistors (59.4 billion vs. 54 billion) and 900MB on-chip memory, but the A100 offers 40GB HBM2 memory. The GC200 supports smaller matrix operands (2,688×2,688 vs. 72,000×72,000) and lags in mixed-precision GEMM performance compared to NVIDIA's TensorCores.
What are the key specifications of the Graphcore GC200?
The GC200 has 1,472 cores running 6 threads each, 900MB on-chip SRAM with 47.5 TB/s bandwidth, and delivers 250 teraFLOPS at FP16.16. It connects via PCIe Gen 4 x16 and uses IPU-Fabric for 8 TB/s system bandwidth, scaling to 64,000 IPUs in data centers.
What is the price of the IPU-Machine M2000 with GC200?
The IPU-Machine M2000, containing four GC200 IPUs, is priced at $32,450. The IPU-POD64 system scales to 64 IPUs with 16 M2000s, but per-chip pricing for the GC200 alone is not publicly disclosed by Graphcore.
What are the strengths and weaknesses of the GC200 for AI?
Strengths include specialized AI architecture, high on-chip memory bandwidth, and 8x performance over the Mk1. Weaknesses are limited 900MB memory per chip, lower mixed-precision GEMM efficiency than NVIDIA GPUs, and a steeper learning curve with the Poplar SDK compared to CUDA.
What workloads is the Graphcore GC200 best suited for?
The GC200 excels at workloads benefiting from massive on-chip parallelism and memory bandwidth, such as sparse models, graph neural networks, and generative AI. It is less ideal for large dense transformers requiring high off-chip memory capacity, which may need model parallelism across multiple chips.
Alternatives
How Colossus MK2 GC200 compares
Direct head-to-head against 2 competitors. Picked by 7wData.
Colossus MK2 GC200
- Pricing
- IPU-Machine M2000 (4x GC200) priced at $32,450; IPU-POD64 (16 M2000s) available as a rack-scale system; per-chip pricing not disclosed.
- Target
- The Colossus MK2 GC200 is Graphcore's second-generation Intelligence Processing Unit (IPU), a processor architected from the ground up for machine intelligence workloads rather than repurposed
- Strength
- Designed exclusively for machine intelligence, not repurposed graphics hardware, enabling specialized parallelism for AI workloads.
- Watch for
- Limited on-chip memory (900MB) restricts single-chip model size compared to NVIDIA A100's 40GB HBM2, requiring model parallelism for large transformers.
NVIDIA A100 GPU
- Pricing
- Custom/Contact sales; A100 80GB PCIe ~$15,000+ per unit
- Target
- AI training and inference in data centers
- Deployment
- On-prem, cloud
- Strength
- Mature CUDA ecosystem and 54B transistor GPU with Tensor Cores
- Watch for
- Higher per-chip memory capacity but lower on-chip memory than IPU
AMD Instinct MI250X
- Pricing
- Custom/Contact sales; ~$15,000-$20,000 per accelerator
- Target
- HPC and AI workloads
- Deployment
- On-prem, cloud
- Strength
- High FP64 performance for scientific computing
- Watch for
- Smaller software ecosystem compared to NVIDIA CUDA
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.