The Titan X Hybrid mod we hand-crafted for a viewer allowed the card to stretch its boost an additional ~200MHz beyond the spec. This was done for Sam, the owner who loaned us the Titan XP, and was completed back in August. We also ran benchmarks before tearing the card down, albeit on drivers from mid-August, and never did publish a review of the card.
This content revisits the Titan XP for a review from a gaming standpoint. We'd generally recommend such a device for production workloads or CUDA-accelerated render/3D work, but that doesn't stop that the card is marketed as a top-of-the-line gaming device with GeForce branding. From that perspective, we're reviewing the GTX Titan X (Pascal) for its gaming performance versus the GTX 1080, hopefully providing a better understanding of value at each price-point. The Titan X (Pascal) card is priced at $1200 from nVidia directly.
Review content will focus on thermal, FPS, and overclocking performance of the GTX Titan X (Pascal) GP102 GPU. If you're curious to learn more about the card, our previous Titan XP Hybrid coverage can be found here:
NVIDIA Titan X (Pascal) Specs
|NVIDIA Pascal vs. Maxwell Specs Comparison|
|Titan X||GTX 1080||GTX 1070||GTX 1060||GTX 980 Ti||GTX 980||GTX 960|
|GPU||GP102-? Pascal||GP104-400 Pascal||GP104-200 Pascal||GP106 Pascal||GM200 Maxwell||GM204 Maxwell||GM204|
|Fab Process||16nm FinFET||16nm FinFET||16nm FinFET||16nm FinFET||28nm||28nm||28nm|
|Memory Capacity||12GB||8GB||8GB||6GB||6GB||4GB||2GB, 4GB|
|Memory Clock||10Gbps||10Gbps GDDR5X||8Gbps||8Gbps||7Gbps GDDR5||7Gbps GDDR5||7Gbps|
|Power Connectors||1x 8-pin
|1x 8-pin||1x 8-pin||1x 6-pin||1x 8-pin
|2x 6-pin||1x 6-pin|
|Release Price||$1200||Reference: $700
NVIDIA Titan X (Pascal) GP102 Architecture
The Titan XP uses a GP102 GPU under Pascal architecture, the largest Pascal chip presently available on a GeForce-branded card. The only current Pascal chip that's equipped in a flashier fashion than the GP102 is GP100, used for the Tesla P100 Accelerator and not meant for gaming. We detailed GP100 a few months ago, if interested in learning more about Pascal's intricacies.
GP102 hosts a total of 6 GPCs, for 28 SMs and 3584 CUDA cores. This isn't the biggest Pascal chip out, in terms of total SMs, but it's the biggest in the GeForce line. GP102 follows GP104's architecture and splits GPCs into sets of 5SMs, as opposed to the GP100 GPC-SM alignment of 10 simultaneous multiprocessors per GPC. This is why we often remind folks that cores and SMs can't be compared cross-generation, or sometimes even intra-generation. The GP100 chip used for the Tesla P100 Accelerator is built for simulation and deep learning, which means it's got a completely different implementation of cores. GP100 makes use of FP64 cores in a way that the Titan XP does not, and so is more suitable for double-precision tasks than the FP32-focused Titan X card of this generation.
Also like GP104-400, the Titan XP runs 8 TMUs per SM, totaling 224 texture map units. Clock-rate natively operates at 1531MHz with the stock cooler and hits 1700MHz or higher during load. Note also that we pushed to nearly 2000MHz in our liquid content without needing to overclock – that's GPU Boost 3.0 in action. The Boost behavior has detected thermal headroom to push the clock higher, and so it increases the operating frequency until a point at which a new limiter is encountered (normally power or voltage).
Cache size also plays a big role in the Titan XP. The TiXP expands cache to 3072KB over the 2048KB on the GTX 1080. VRAM on-board is 12GB for the Titan XP, versus 8GB for the GTX 1080. Both use GDDR5X at 10Gbps native.
Titan XP and GP102 host 12 billion total transistors and, despite a focus on FP32, works to introduce a new INT8 deep-learning instruction set. We're not benchmarking that today, though.
NVIDIA Titan X (Pascal) Tear-Down, PCB, & Cooler
Below are a few photos from our tear-down content:
The disassembly process for the Titan XP is almost exactly the same as what is used to disassemble the GTX 1080. The card uses a vapor chamber cooler for its thermal solution, coupled with a blower fan and usual PWM control. An aluminum baseplate rests between the PCB and the faceplate, which uses thermal pads to conduct heat into the plate. Air from the blower fan then pushes air through fins atop the baseplate, dissipating heat and pulling it away from VRAM and the VRM. This air is eventually pushed through the vapor chamber cooler, then out the back of the card.
The PCB hosts two power connectors and a third set of solder points for an additional 8-pin header, located at the right side of the board. The core VRM uses a 7-phase power design, with the memory VRM using a 2-phase design.
Continue to page 2 for test methodology.