Nvidia hard-launches its GeForce RTX 2070 graphics card


Nvidia's GeForce RTX 2070 graphics card launches this morning. First, the bad news: we don't have an RTX 2070 in the TR labs for testing. We're working to obtain one of those cards as soon as we can to review in depth, but my test bench (and attention) is presently occupied by Intel's Core i9-9900K and a raft of other CPUs. Stay tuned for more details on those chips soon.


A block diagram of the TU106 GPU. Source: Nvidia

With the bad news out of the way, here's what we know so far about the RTX 2070. The smallest Turing sibling introduces a third chip powered by Nvidia's latest architecture called TU106. From the top, this chip uses three Turing graphics processing clusters (or GPCs), each with 12 streaming multiprocessors (SMs) for a total of 36 graphics-processing "cores." Each of those SMs has 64 shader ALUs, or CUDA "cores," for a total of 2304 such resources. TU106 has 144 texture units, 64 ROPs, and a 256-bit bus to 8 GB of GDDR6 memory running at 14 Gbps. Here's how that stacks up versus some of today's most popular graphics cards:

Boost
clock
(MHz)
ROP pixels/
clock
INT8/FP16
textures/clock
Shader
processors
Memory
path (bits)
Memory
bandwidth
Memory
size
RX Vega 56 1471 64 224/112 3584 2048 410 GB/s 8 GB
GTX 1070 1683 64 108/108 1920 256 259 GB/s 8 GB
RTX 2070 FE 1710 64 120/120 2304 256 448 GB/s 8 GB
GTX 1080 1733 64 160/160 2560 256 320 GB/s 8 GB
RX Vega 64 1546 64 256/128 4096 2048 484 GB/s 8 GB
RTX 2080 FE 1800 64 184/184 2944 256 448 GB/s 8 GB
GTX 1080 Ti 1582 88 224/224? 3584 352 484 GB/s 11 GB
RTX 2080 Ti FE 1635 88 272/272 4352 352 616 GB/s 11 GB
Titan Xp 1582 96 240/240 3840 384 547 GB/s 12 GB
Titan V 1455 96 320/320 5120 3072 653 GB/s 12 GB
Peak
pixel
fill
rate
(Gpixels/s)
Peak
bilinear
filtering
INT8/FP16
(Gtexels/s)
Peak
rasterization
rate
(Gtris/s)
Peak
FP32
shader
arithmetic
rate
(TFLOPS)
RX Vega 56 94 330/165 5.9 10.5
GTX 1070 108 202/202 5.0 7.0
RTX 2070 FE 109 246/246 5.1 7.9
GTX 1080 111 277/277 6.9 8.9
RX Vega 64 99 396/198 6.2 12.7
RTX 2080 115 331/331 10.8 10.6
GTX 1080 Ti 139 354/354 9.5 11.3
RTX 2080 Ti 144 473/473 9.8 14.2
Titan Xp 152 380/380 9.5 12.1
Titan V 140 466/466 8.7 16.0

Other Turing architectural changes might not show up in our tables, but they're still important. The RTX 2070 has twice as much L2 cache as the GTX 1070, at 4 MB versus 2 MB, and the total size of its register files (distributed among each SM) has ballooned to 9.216 MB, versus 3.840 MB on the GTX 1070. Keeping more data close to the execution units that can digest it is one reliable way to improve performance, and the on-chip cache of TU106 can certainly claim a much greater endowment in that regard compared to its Pascal predecessor.

Like its bigger Turing siblings, the RTX 2070 also features some resources devoted to the acceleration of certain ray-tracing operations. The RTX 2070 has 288 Turing tensor cores (eight per SM) and 36 Turing RT cores (one per SM). Compare that to the 368 tensor cores and 46 RT cores of the slightly-cut-down TU104 chip that powers the RTX 2080, or the 544 tensor cores and 68 RT cores of the modestly-gelded TU102 chip that runs the RTX 2080 Ti.

Since we don't have any real-world applications that use these processing resources yet, it's hard to say how the RTX 2070 stacks up in the ray-tracing department beyond the obvious point that it will be less proficient at those tasks than its larger siblings. For what it's worth, Nvidia says the RTX 2070 Founders Edition can perform 45 RTX tera-OPS (a measurement of the performance potential of hybrid rendering with Turing GPUs), versus 60 for the RTX 2080 FE and 78 for the RTX 2080 Ti FE. Until we have applications in hand that can take advantage of those resources, we'll hold off judgment on just how useful the ray-tracing features in the RTX 2070 will be.


The RTX 2070 Founders Edition

Since the RTX 2070 is fabricated on TSMC's 12-nm FFN process, its 10.8 billion transistors don't benefit from much, if any, of an areal shrink versus the 16-nm FinFET process used to make Pascal. Consequently, the TU106 die is 445 mm², compared to 314 mm² for the GP104 chip that powered the GTX 1070 and GTX 1080. The RTX 2070's board power is up versus the GTX 1070, as well, at 175 W for the "reference" spec and 185 W for the "factory-overclocked" Founders Edition. The GTX 1070 needed 150 W to do its thing.

From a pool of 25 games and synthetic benchmarks, Nvidia's testing labs expect that the RTX 2070 will deliver a median 33% better performance for 2560x1440 SDR gaming over the GTX 1070 and a median 35% improvement for HDR gaming. If the RTX 2070 delivered that improvement for the same price as the GTX 1070 before it, we would be shouting from the rooftops about it. As with all Turing cards, though, Nvidia is charging more for that performance increase. The company suggests partner cards will start at $499, and the RTX 2070 Founders Edition will land for $599. That's versus $379 for partner GTX 1070 cards or $449 for the Founders Edition when those pixel-pushers launched. Ultimately, we'll need to put an RTX 2070 through its paces in the TR labs to see how it stacks up.

Tip: You can use the A/Z keys to walk threads.
View options