This part has no DRAM. The 900 MB of In-Processor Memory is the whole memory system, and the 65.25 TB/s is on-chip SRAM bandwidth, not comparable with an HBM or GDDR figure, which is why it sits in the effective bandwidth column. Up to 256 GB of Streaming Memory is available per Bow-2000, but that is DDR4 on the 1U machine reached through the IPU-Gateway chip and shared by all four IPUs, with no per-IPU share ever stated, so it is recorded here rather than in a column. The two throughput rows come from two different documents and their ratio is 4.01 rather than exactly 4 for that reason: FP16 350 is Graphcore's own per-processor figure from the launch post, while FP32 87.25 is the datasheet's 0.349 PFLOPS divided by the four processors printed beside it, because Graphcore publishes no per-processor FP32 rate for this part. Graphcore rounds its machine totals down slightly, so 4 x 350 is 1,400 TFLOPS against a printed 1.394 PFLOPS. Graphcore has no structured sparsity feature, so there is no sparse figure on any row.
Graphcore logo

Graphcore Bow IPU

Type: ASICArchitecture: Colossus MK2Released: 2022Spec confidence: Official
FP32
87
TFLOPS

Overview

The Bow IPU, announced in March 2022, was the first processor in volume production built with wafer-on-wafer 3D stacking. Graphcore bonded a second wafer under the compute die whose only job is power delivery, putting deep trench capacitors directly beneath the logic they feed, and the cleaner supply let the same Colossus MK2 design clock at 1.85 GHz instead of 1.325. Nothing else changed: 1,472 IPU-Cores, 8,832 threads and 900 MB of In-Processor Memory, architecturally compatible with the GC200 so that existing code runs unmodified. The result is 350 TFLOPS of FP16.16 against the GC200's 250, on-chip bandwidth up from 45 to 65 TB/s, and Graphcore's claim of up to 40 percent more real-world performance and 16 percent better efficiency. Like the GC200 it has no DRAM and no card SKU: it ships four to a 1U Bow-2000 IPU-Machine, and those stack into Bow Pod systems of up to 256.

Performance

Peak theoretical throughput by precision type

PrecisionPeak
FP64
No verified data available
FP32
32-bit floating point
87TFLOPS
TF32
No verified data available
BF16
No verified data available
FP16
16-bit floating point
350TFLOPS
FP8
No verified data available
FP6
No verified data available
FP4
No verified data available
INT8
No verified data available

Specifications

Architecture

Colossus MK2

Form Factor

No verified data available

Launch Year

2022

Process Node

No verified data available

Memory

No verified data available On-die SRAM

Bandwidth

No verified data available

TDP

No verified data available

Max power

No verified data available

Interconnect

IPU-Link

IPU-Cores

1,472

Spec Confidence

Official

Full Specifications

Compute Engine
IPU-Cores 1,472
Base Clock 1.85 GHz
Chip Design
Memory
Memory No verified data available
Memory Type On-die SRAM
Bandwidth No verified data available
Interface Width No verified data available
Effective Bandwidth 65.3 TB/s
On-Die SRAM 900 MB
Interconnect & I/O
GPU-to-GPU IPU-Link
Power & Thermal
TDP No verified data available
Max power No verified data available
General
Form Factor No verified data available
Architecture Colossus MK2
Process Node No verified data available
Launch Year 2022

Datasheet & Resources

Flopper Datasheet

Graphcore Bow IPU specifications, generated from our database. Printable.

Bow-2000 IPU-Machine Datasheet, Product description (Graphcore)

Graphcore · Latest version

View

Data Provenance

Every figure traced to a source

Primary Source

Publisher
Graphcore
Published
No verified data available

Data Quality

Spec confidence
Official
Clock basis
Boost
Core precisions with figures
2 of 9
Normalization
All values in TFLOPS

Compare cloud providers offering on-demand GPU instances for AI training, inference, and HPC workloads.

Browse GPU Cloud Providers

Similar GPUs

Graphcore C600

PCIe · 2022
FP32: No verified data available
Compare vs C600

Graphcore GC200

· 2020
FP32: 63 TFLOPS
Compare vs GC200

Frequently Asked Questions

How many TFLOPS does the Graphcore Bow IPU have?

The Graphcore Bow IPU delivers 87 TFLOPS FP32 and 350 TFLOPS FP16 at peak. Flopper does not currently have a verified FP8 throughput figure for it.

What is the power consumption of the Graphcore Bow IPU?

Flopper does not currently have a verified TDP (Thermal Design Power) figure for the Graphcore Bow IPU.

How much memory does the Graphcore Bow IPU have?

Flopper does not currently have verified memory capacity or bandwidth figures for the Graphcore Bow IPU.

What architecture is the Graphcore Bow IPU based on?

The Graphcore Bow IPU is based on the Colossus MK2 architecture, launched in 2022.

Stay Updated on GPU Releases

Get notified when new GPUs are added or specifications are updated.

Loading verification...

No spam, unsubscribe anytime.

Back to GPUs

© 2026 Flopper.io - Compare the GPUs Powering AI