Apple publishes no throughput figure of any kind for this part. Its newsroom copy rates the M5 Ultra only as "up to 4.3x the peak AI compute performance when compared to M3 Ultra and a staggering 9.8x more than M1 Ultra", and its technical specifications pages give core counts, Neural Engine cores and memory bandwidth and nothing else. The FP32 value here is therefore derived on this catalog's standing Apple formula, GPU cores x 128 ALUs x 2 FMA operations x clock, and it comes out at exactly twice the M5 Max, which is what a quad-die part built from two M5 Max dies should do. It is a vector rate, so it is flagged is_tensor_core = 0. NO FP16, BF16 OR INT8 FIGURE IS STORED: M5 puts a Neural Accelerator inside each GPU core and that is genuine matrix hardware, but Apple rates it only against a previous generation, and the separate 32-core Neural Engine is not a GPU metric either. The clock is the M5 family figure used across the other M5 rows here and is itself an estimate, which is the main reason this row is vendor_claimed rather than official. Two GPU bins exist, 64-core and 80-core; this row is the 80-core configuration.

Apple M5 Ultra NEW

Apple M5 SoC 2026 3nm Spec confidence: Vendor claimed
FP32
33
TFLOPS
VRAM
512 GB
LPDDR5X
Bandwidth
1.2 TB/s
memory
Download datasheet

Overview

The M5 Ultra is the largest chip Apple has ever built and the first with four dies, made by joining two dual-die M5 Max chips over next-generation UltraFusion. That gives up to 80 GPU cores, each carrying its own Neural Accelerator, a 36-core CPU and a 32-core Neural Engine. For local model work the number that matters is not compute but memory: 512 GB of unified memory at 1.2 TB/s, which is more capacity than any single datacenter GPU on this site and roughly a quarter of the bandwidth of a top-end HBM part. That trade is the whole argument for a Mac Studio as an inference machine, since a model that will not fit on an 80 GB accelerator fits here comfortably and simply runs slower. Apple skipped the Ultra tier of the M4 generation altogether, so this is the first new Ultra since the M3 Ultra of March 2025, and Apple's own comparison is up to 4.3x its peak AI compute.

Performance

Peak theoretical throughput by precision type

PrecisionPeak
FP64
No verified data available
FP32
32-bit floating point
33TFLOPS
TF32
No verified data available
BF16
No verified data available
FP16
No verified data available
FP8
No verified data available
FP6
No verified data available
FP4
No verified data available
INT8
No verified data available

Every figure here is dense. The vendor documents no structured (2:4) sparsity mode for this part, so there is no second number to quote.

Specifications

Architecture

Apple M5

Form Factor

SoC

Launch Year

2026

Process Node

3nm

Memory

512 GB LPDDR5X

Bandwidth

1,200 GB/s

TDP

No verified data available

Max Power

No verified data available

Interconnect

UltraFusion

GPU cores

80

Spec Confidence

Vendor claimed

Full Specifications

Compute Engine
GPU cores 80
Chip Design
Process Node 3nm
Chiplets 4 (Two dual-die M5 Max chips joined by next-generation UltraFusion)
Memory
VRAM 512 GB
Memory Type LPDDR5X
Bandwidth 1.2 TB/s
Interface Width No verified data available
Interconnect & I/O
GPU-to-GPU UltraFusion
Power & Thermal
TDP No verified data available
Max Board Power The vendor publishes no maximum board power; the Max Power tile above estimates it from TDP
Enterprise Features
Sparsity No
General
Form Factor SoC
Architecture Apple M5
Process Node 3nm
Launch Year 2026

Datasheet & Resources

Data Provenance

Every figure traced to a source

Primary Source

Publisher
Apple
Published
2026-08-25

Data Quality

Spec confidence
Vendor claimed
Clock basis
Boost
Core precisions with figures
1 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

Apple M5 Max

SoC · 2026
FP32: 17 TFLOPS
Compare vs M5

Apple M5 Pro

SoC · 2026
FP32: 8.3 TFLOPS
Compare vs M5

Apple M5

SoC · 2025
FP32: 4.2 TFLOPS
Compare vs M5

Frequently Asked Questions

How many TFLOPS does the Apple M5 have?

The Apple M5 delivers 33 TFLOPS FP32 at peak. Flopper does not currently have verified FP16 and FP8 throughput figures for it.

What is the power consumption of the Apple M5?

Flopper does not currently have a verified TDP (Thermal Design Power) figure for the Apple M5.

How much memory does the Apple M5 have?

The Apple M5 is equipped with 512 GB of VRAM with 1,200 GB/s of memory bandwidth.

What architecture is the Apple M5 based on?

The Apple M5 is based on the Apple M5 architecture, launched in 2026.

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