The GTX 1080 Ti posed a fun opportunity to roll-out our new GPU test bench, something we’ve been working on since end of last year. The updated bench puts a new emphasis on thermal testing, borrowing methodology from our EVGA ICX review, and now analyzes cooler efficacy as it pertains to non-GPU components (read: MOSFETs, backplate, VRAM).
In addition to this, of course, we’ll be conducting a new suite of game FPS benchmarks, running synthetics, and preparing for overclocking and noise. The last two items won’t make it into today’s content given PAX being hours away, but they’re coming. We will be starting our Hybrid series today, for fans of that. Check here shortly for that.
If it’s not obvious, we’re reviewing nVidia’s GTX 1080 Ti Founders Edition card today, follow-up to the GTX 1080 and gen-old 980 Ti. Included on the benches are the 1080, 1080 Ti, 1070, 980 Ti, and in some, an RX 480 to represent the $250 market. We’re still adding cards to this brand new bench, but that’s where we’re starting. Please exercise patience as we continue to iterate on this platform and build a new dataset. Last year’s was built up over an entire launch cycle.
NVIDIA GTX 1080 Ti Specs vs. 1080, 980 Ti
| NVIDIA Pascal Specs Comparison | ||||
| Tesla P100 | GTX 1080 Ti | GTX 1080 | GTX 1070 | |
| GPU | GP100 Cut-Down Pascal | GP102 Pascal | GP104-400 Pascal | GP104-200 Pascal |
| Transistor Count | 15.3B | 12B | 7.2B | 7.2B |
| Fab Process | 16nm FinFET | 16nm FinFET | 16nm FinFET | 16nm FinFET |
| CUDA Cores | 3584 | 3584 | 2560 | 1920 |
| GPCs | 6 | 6 | 4 | 3 |
| SMs | 56 | 28 | 20 | 15 |
| TPCs | 28 TPCs | 20 TPCs | 15 | |
| TMUs | 224 | 224 | 160 | 120 |
| ROPs | 96 (?) | 88 | 64 | 64 |
| Core Clock | 1328MHz | - | 1607MHz | 1506MHz |
| Boost Clock | 1480MHz | 1600MHz | 1733MHz | 1683MHz |
| FP32 TFLOPs | 10.6TFLOPs | ~11.4TFLOPs | 9TFLOPs | 6.5TFLOPs |
| Memory Type | HBM2 | GDDR5X | GDDR5X | GDDR5 |
| Memory Capacity | 16GB | 11GB | 8GB | 8GB |
| Memory Clock | ? | 11Gbps | 10Gbps GDDR5X | 4006MHz |
| Memory Interface | 4096-bit | 352-bit | 256-bit | 256-bit |
| Memory Bandwidth | ? | ~484GBs | 320.32GB/s | 256GB/s |
| Total Power Budget ("TDP") | 300W | 250W | 180W | 150W |
| Power Connectors | ? | 1x 8-pin 1x 6-pin |
1x 8-pin | 1x 8-pin |
| Release Date | 4Q16-1Q17 | TBD | 5/27/2016 | 6/10/2016 |
| Release Price | - | $700 | Reference: $700 MSRP: $600 Now: $500 |
Reference: $450 MSRP: $380 |
NVidia’s GTX 1080 Ti launches with an MSRP of $700 for the reference and AIB partner designs, with AIB partners flexible to offer higher end SKUs over MSRP. This price announcement kicked the GTX 1080 down to an MSRP of $500, as instituted by some partners already.
All these 1080 Ti cards are built on the new GP102 die; if you wanted the full identifier, it’s GP102-350-K1for the die underneath. We’ve already dismantled our unit for thermal analysis and future projects, including plans for a VRM analysis by Buildzoid later this weekend.
This block diagram of the GTX 1080 Ti shows its layout, broken into 6 GPCs in a similar fashion to how GP100 is broken up, though with a different SM layout. The focus is removed from anything but FP32 with the GTX 1080 Ti, where we have 128 FP32 CUDA cores per SM. Considering a total of 28SMs, we’re at 3584 CUDA cores total; the GTX 1080 GP104 chip hosts 20SMs, totaling 2560 cores, leaving us about 1000 cores over the GTX 1080. This seems to align with nVidia’s marketing claims that the GTX 1080 Ti increases performance by 35% over the GTX 1080 that launched in May of last year, something we’ll validate later.
The rest of the block diagram is familiar: We’re faced with 224 TMUs and a base clock of 1480MHz, or boost of 1582MHz, resulting in a texture filter rate of 354.4GT/s boosted. A total of 88 ROPs are present on the card, compared to 64 on the GTX 1080. If you start totaling these things up, GP102 looks more familiar to GP100 in some ways – it’s got the same TMU count and the same GPC count, but ultimately, this is not the biggest Pascal die that exists. They’re still different in significant ways. Most obviously, GP100 is built for acceleration, not for consumer graphics. The precision on GP100 splits itself between FP32, FP16, and FP64 capabilities, whereas GP102 focuses most heavily on FP32 capabilities. The more proper comparison would be to GP104, where we’ve gained 8 SMs on GP102 and have widened the memory interface to support a new memory configuration.
Memory moves to 11GB of GDDR5X (“G5X”), up from 8GB of GDDR5X on the GTX 1080, and down from 12GB on the Titan XP (review here) – because we couldn’t have them be equal, clearly. Micron upgraded its GDDR5X from what debuted on the GTX 1080, shifting to 11Gbps capabilities from 10Gbps on launch day. Singing a similar memory tune, new SKUs of the lower end Pascal cards will now offer memory OC variants. That’s a different story, though, and is explained here.
GTX 1080 Ti Founders Edition Teardown
This is a bigger die than the 314mm2 GP104, as the GTX 1080 Ti measures in at 471mm2 to accommodate the additional SMs.
Tearing down the cooler will be fully detailed in our standalone video piece, but the basics are the same as the GTX 1080 FE card: We’re looking at a thin, metal backplate secured by ~16 PH00 screws, under which rest another 16 4.0mm Allen screws. These secure the baseplate to the PCB, with another 4 spring-tensioned screws for the heatsink.
Manage to pull all those off, and another 5 for the I/O shield, and we can remove the entire heatsink + shroud unit from the PCB. This reveals a dual-FET layout, which really just means that nVidia is doubling its phases (2x FETs per phase, all Fairchild). The memory VRM leverages E6930 FETs, something that Buildzoid will talk about in our VRM analysis coming up this weekend. Subscribe to the YT channel for that.
The heatsink is still a vapor chamber finstack, using a radial fan for cooling. NVidia removed the DVI port from this card, freeing up some of the air channels (see: page 3), and has included a DP->DVI adapter for those in need.
