This card carries four ATOM chips, not one, and every figure on this row is the whole-card total. Rebellions states the count in its own SDK documentation, and the arithmetic confirms it from both directions: each stored value is exactly four times the single-chip ATOM card's (128 against 32 TFLOPS, 512 against 128 TOPS, 64 against 16 GB, 1,024 against 256 GB/s) and exactly one eighth of the eight-card ATOM-Max Server's (1,024 TFLOPS, 4,096 TOPS, 512 GB, 8 TB/s). Two power labels cover one number: the live product page calls 350 W the Max Power Consumption, the product brief calls it the TDP, and the white paper calls it an envelope, so both power fields carry it, exactly as they do on the Rebel100. INT8 is four times FP16 on this part rather than twice, and INT4 is twice INT8; that ratio is unchanged on the single-chip card, so it is a property of the ATOM datapath and not a transcription error, and neither figure should be halved to make the usual ladder hold. No Rebellions document for any ATOM part uses the word sparsity, none of the three figures is footnoted and no second column exists, so all three are dense as printed and sparsity_support is left empty rather than set to a denial. Memory bandwidth is 1,024 GB/s: the 2024 product brief prints 256 GB/s in a card-level table, which is one chip's share and is exactly what the single-chip ATOM card publishes. Rebellions publishes no process node, no engine count and no clock for this part, and the 4 MB of per-engine and 32 MB of per-chip SRAM its SDK documentation describes are chip figures with no card total behind them.
Rebellions logo

Rebellions ATOM-Max

Type: NPUArchitecture: ATOMForm factor: PCIe 4× acceleratorReleased: 2024Spec confidence: Official
Memory
64 GB
GDDR6
Bandwidth
1.0 TB/s
memory
TDP
350 W

Overview

The ATOM-Max is the largest card in Rebellions' first-generation line and the part the Korean designer actually ships in volume, sold as RBLN-CA25. It is a full-height full-length dual-slot PCIe card carrying four ATOM neural processors, rated together at 128 TFLOPS of FP16, 512 TOPS of INT8 and 1,024 TOPS of INT4, with 64 GB of GDDR6 at 1,024 GB/s inside a 350 W envelope. Choosing GDDR6 rather than HBM is the whole design argument: Rebellions sells the card on tokens per second per watt against an NVIDIA L40S rather than on peak arithmetic, and the product brief claims roughly four times the L40S' tokens per second per watt on Llama 3.1 70B. Each card carries two dedicated intercard connectors that let up to eight of them talk directly over PCIe Gen5 without going back through the host, in ring, torus or tree topologies, and the compiler splits a model across the chips through what Rebellions calls Rebellions Scalable Design. Eight of these cards make up the ATOM-Max Server at 1,024 TFLOPS and 512 GB. This is not the same silicon as the Rebel100: that part is the four-chiplet HBM3E design announced in 2025, while ATOM is the 2024 GDDR6 generation that preceded it.

Performance

Peak theoretical throughput by precision type

PrecisionPeak
FP64
No verified data available
FP32
No verified data available
TF32
No verified data available
BF16
No verified data available
FP16
16-bit floating point
128TFLOPS
FP8
No verified data available
FP6
No verified data available
FP4
No verified data available
INT8
8-bit integer
512TOPS
INT4
4-bit integer
1,024TOPS

Specifications

Architecture

ATOM

Form Factor

PCIe

Launch Year

2024

Process Node

No verified data available

Memory

64 GB GDDR6

Bandwidth

1,024 GB/s

TDP

350 W

Max power

350 W

Interconnect

MCIO

Spec Confidence

Official

Full Specifications

Compute Engine
GPUs per Module 4 (all figures are totals for the module)
Memory
Memory 64 GB
Memory Type GDDR6
Bandwidth 1.0 TB/s
Interface Width No verified data available
Interconnect & I/O
GPU-to-GPU MCIO
Interconnect Links 2
PCIe 5.0 x16
Power & Thermal
TDP 350 W
Max power 350 W
Power Connector CPU 8-pin
Cooling Passive
Enterprise Features
Compute APIs RBLN SDK, PyTorch, TensorFlow, Hugging Face, vLLM, Triton Inference Server, TorchServe
Physical & Media
Width 2-slot
General
Form Factor PCIe
Architecture ATOM
Process Node No verified data available
Launch Year 2024

Datasheet & Resources

Flopper Datasheet

Rebellions ATOM-Max specifications, generated from our database. Printable.

Vendor product page

Rebellions documentation for the ATOM-Max

View

ATOM-Max product page (Rebellions)

Rebellions · 2024-11-18

View

Data Provenance

Every figure traced to a source

Primary Source

Publisher
Rebellions
Published
2024-11-18

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

Frequently Asked Questions

How many TFLOPS does the Rebellions ATOM-Max have?

The Rebellions ATOM-Max delivers 128 TFLOPS FP16 at peak. Flopper does not currently have verified FP32 and FP8 throughput figures for it.

What is the power consumption of the Rebellions ATOM-Max?

The Rebellions ATOM-Max has a TDP (Thermal Design Power) rating of 350 watts.

How much memory does the Rebellions ATOM-Max have?

The Rebellions ATOM-Max is equipped with 64 GB of memory with 1,024 GB/s of memory bandwidth.

What architecture is the Rebellions ATOM-Max based on?

The Rebellions ATOM-Max is based on the ATOM architecture, launched in 2024.

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