Scrolling moves from the assembled card through its PCB, power regulation, GA102 processor, memory, thermal interfaces, airflow, exploded mechanical order, and evidence.

GA102

Loading engineering model…

RTX 3090 / FE
RTX 3090 · SYSTEMS CUTAWAY
RTX 3090 FOUNDERS EDITION · BUYER'S CUTAWAY

NVIDIA
RTX 3090

A buyer-focused cutaway of GA102, 24 GB of GDDR6X, the 350 W power system, and the three-slot cooler that keeps them usable.

10,496CUDA CORES
24 GBGDDR6X
3 SLOTFLOW-THROUGH
TRACE THE HARDWARE
FIT + AIRFLOW

THE BOARD STOPS.
AIR KEEPS GOING.

The compact PCB leaves a full fin section beyond its edge. One fan moves air through that open section into the case; the bracket-side fan pushes a separate path through the I/O vents.

CASE FIT313 MM / 3 SLOT
AIR ROUTECASE + BRACKET
DISPLAY OUTPUTS3× DP 1.4a + HDMI 2.1
BEFORE YOU BUYConfirm GPU length, three-slot clearance, and a clear top or rear exhaust path behind the flow-through fan.
POWER + HOST LINK

350 W
STARTS HERE.

The 12-pin socket carries the auxiliary 12 V supply. The PCIe edge contributes slot power and the ×16 host link. The routes show those different jobs without inventing NVIDIA's unpublished rail map.

PSU12-PINPCB
12 V POWERPCIe DATA
TOTAL GRAPHICS POWER350 W
HOST LINKPCIe 4.0 ×16
18-PHASE FE POWER DELIVERY

18 PHASES
SHARE THE WORK.

Interleaved power stages switch in sequence. Their inductors and capacitors turn those pulses into steadier, low-voltage power for GA102 and GDDR6X; the exact rail allocation is not public.

SWITCHINDUCTORRAIL
ILLUSTRATED LOAD 0 W Scroll position controls this teaching state · not live telemetry
GA102 PROCESSOR

628 MM² OF
PARALLEL WORK.

The RTX 3090 enables 82 of GA102's 84 SMs: 10,496 CUDA cores, 82 second-generation RT cores, and 328 third-generation Tensor cores. The seven-group overlay explains hierarchy, not a proprietary die photograph.

CUDA / SCHEDULINGRTTENSOR
CUDA CORES10,496
2ND-GEN RT82
3RD-GEN TENSOR328
24 GB GDDR6X

24 GB
SURROUNDS GA102.

Twenty-four 1 GB packages—12 on each PCB face—serve twelve 32-bit controllers. Together they form a 384-bit interface capable of 936 GB/s, reducing how often large scenes or datasets must leave local memory.

PEAK BANDWIDTH936 GB/S
SIGNALPAM4 / 2 BITS
FOUR LEVELS, TWO BITSPAM4 sends one of four voltage levels each symbol, so every symbol represents two bits.
DIE-TO-AIR THERMAL STACK

FROM DIE
TO FINS.

Heat crosses the die interface into the cold plate and hybrid vapor chamber, spreads through the chamber and heat-pipe network, then enters the aluminum fins. Thermal pads give memory and power components their own path into the cooler.

DIETIMCHAMBERFINS
TWO-PHASE TRANSPORTWorking fluid evaporates near the heat source, condenses in cooler regions, and returns through a capillary wick.
TWO FANS · TWO PRESSURE PATHS

ONE EXHAUSTS.
ONE PASSES THROUGH.

The bracket-side fan drives a curved, higher-resistance path toward the I/O vents. The fan beyond the shortened PCB drives a straighter path through the open fin stack and into the chassis.

COOL INLETWARM EXHAUST
CASE DUTYTOP / REAR EXHAUST
SERVICE VIEW · RECONSTRUCTED

THE 3-SLOT
COOLING STACK.

The card opens in mechanical contact order: covers and fans, fin banks, two-phase spreader, interfaces, component-populated PCB, and rear thermal path. Unpublished fasteners and clearances remain reconstructed.

  1. Front cover + bracket-side fan
  2. Aluminum fin banks
  3. Hybrid chamber + spreaders
  4. Paste and elastomer pads
  5. GPU, memory, and VRM
  6. Multilayer PCB
  7. Rear memory + thermal pads
  8. Rear cover + flow-through fan
BUYER SUMMARY

24 GB NEEDS
ROOM TO BREATHE.

The RTX 3090 pairs 24 GB and mature CUDA/RTX support with a 350 W, three-slot thermal commitment. Check length, dedicated PSU cables, and a clear case exhaust path before buying.

ENGINEERING EVIDENCE

SPECS, PHOTOS,
AND BOUNDARIES.

CONFIRMED SPECIFICATION

GPU
GA102 / Samsung custom 8N
Die
628.4 mm² / 28.3B transistors
Memory
24 GB GDDR6X / 384-bit
Power
350 W / 18-phase FE design
Cooling
Hybrid chamber / aluminum fins
Envelope
313 × 138 mm / 3-slot
MODEL BOUNDARY Public sources define the architecture, dimensions, airflow paths, memory interface, and component families. Exact alloy grades, rail allocation, internal layer artwork, and assembly clearances are not public, so those details remain labeled reconstructions.