Nvidia's GeForce GTX 1070 graphics card getsundressed and dissected. Its core the PascalGP104 GPU exposed and beautifully photographed. These are the first ever public die shots of the 314mm². If you're a hardware enthusiast grab a drink, sit back, relax and prepare to drool.

Nvidia GP104 GPU Dissected- Cutting ThroughTheHeart Of The GTX 1080 & 1070

Firs things first, all credit goes to Fritzchens Fritzwho has gone to astounding lengths to get these amazing die shots. He has a wonderful collection of pristine quality die shots of a wide range of GPUs, the latest of which happens to be GP104. Nvidia's mid-sized Pascal GPU powering the GeForce GTX 1080 and its little brother the GTX 1070.

2 of 9

The couple ofphotos you see above are of a GTX 1070that has been torn down. The GP104 ASIC was first removed from its socketand the die separated fromits mini circuit board.

The building block of theMaxwellarchitecture has been stripped apart and redesigned to create Pascal. This building block which Nvidia dubs the Streaming Multiprocessor or SM for short is the engine that drives the graphics and compute horsepowerof every Pascal chip.

2 of 9

With Maxwell, Pascal's predecessor powering the GTX 900 series, Nvidia introduced theStreaming Maxwell Multiprocessor. The SMM built on the strengths ofNvidia's Kepler SM - introduced with the GTX 600 and 700 series - which Nvidia dubs the SMX. It also done away with many unnecessary complexities which enabled the engine to delivermore throughput and higher clock speeds.The Pascal SM in its own right is an evolution of the Maxwell SM, a smarter, more streamlined engine.

Nvidia GTX 1080 - GP104 Block DIagram

Inside of a fully unlocked GP104 there are four Graphics Processing Clusters or GPCs for short. Each GPC consists of five Streaming Multiprocessors or SMs for short - each SM contains 128 CUDA cores - and sixteen Texture Mapping Units , AKA TMUs. Each GPCincludes two render back-ends made up of eightRender Output Units, ROPs, each. In total this adds up to2560 CUDA cores, 160 TMUs and 64 ROPs inside GP104. Finally the engine is connected via eight 32-bit GDDR5X memorysegments - 256bit memory controller - to 8GB of GDDR5X memory.

Each GP104streaming multiprocessor includes 128 FP32 CUDA cores, the same asMaxwell. Within each GP104streaming multiprocessor there are four 32 CUDA core partitions, fourdispatch units, two warpschedulers and a fairly large instructionbuffer.Twice as large compared to Maxwell.

GP104,Bare

There it is, GP104 in all its glory. It looks nothing like the block diagram and for good reason. The GP104 diagram published by Nvidia is no more than a simplistic visual representation of the architecture. The physical implementation of that architecture however is far more complex. It's no surprise that designing a modern high performance graphics chip can costwell into the hundreds of millions.

The architecture itself, putting pen to paper, isnot the hardest or most expensive part of developing a chip like GP104. The physical implementation itself, putting light to sand so to speak, is where things can get reallychallenging. The bigger the chip the bigger the challenge.

The physical layout of the GP104 GPU is grouped into four main GPC engine divisions. Each GPC taking roughly one quarter of the chipand housing5 SMs. Feeding the beast are the eight 32bit GDDR5X memory segments that make-up ring around the periphery of the die.

Nvidia GP104 GPU die shot via Fritzchens Fritz with added annotations by Wccftech.com

The SM arrangementis almost identical to what we've seen with the much larger 3840 CUDA core GP100 GPU that powers the Tesla P100 accelerator.Only inside GP100 each SM contains exactly half the number of CUDA cores, dispatch units and warp schedulers vs GP104. But in turn there are twice as many SMs per GPC.

So the primary layout differencebetweenGP104 and GP102 is that Nvidia is grouping 64 CUDA core SMs inpairsmade up of 128 CUDA cores each and in turnnaming the larger128 unit an SM instead. This is all whilemaintaining the exact same ratio of dispatch units, warp schedulers and instruction buffers per CUDA core that we've seen with GP100.

The GP104 Pascal Streaming Multiprocessor

So think of it as Nvidia just pairing 64 CUDA core groupstogether in a single SM.This decision is likely influenced by the significant reduction of FP64, double precision, CUDA cores per SM inside GP104 vs GP100. GP104 only contains 1 FP64 CUDA core for every 32 FP32 CUDA cores. While GP100 has one FP64 CUDA core for every two FP32 CUDA cores, 16 timesmore than GP104.

Additionally, each GP104SM has twicethe number of registersasMaxwell. This in turn means that not only can Pascal accommodate more threads compared to Maxwell but each thread has access to more registers and thus a lot more throughput.Finally, each warp scheduler can dispatch two instructions per clock.

Nvidia's Senior Architect, Lars Nyland admitted that the 16nm FinFET process played an important role in realizing the team's power efficiency goals for Pascal, but maintains that numerous architectural improvements aided in further reducing the energy footprint of the architecture. Including the employment of new on-chip voltage signaling techniques.

The GTX 1080's bigger brother the GTX 1080 Ti is expected to launch some time next year withwell over three thousand CUDA cores and GTX Titan X Pascal'esque performance.

Official Geforce GTX 1080 and Geforce GTX 1070 Specifications