NVIDIA Announces GeForce RTX 2060: Starting At $349, Available January 15th
See the original posting on Anandtech
Kicking off CES 2019 with a surprisingly announcement-packed keynote session, NVIDIA this evening has announced the next member of the GeForce RTX video card family: the GeForce RTX 2060. The newest and now cheapest member of the RTX 20 series continues the cascade of Turing-architecture product releases towards cheaper and higher volume market segments. Designed to offer performance around the outgoing GeForce GTX 1070 Ti, the new card will hit the streets next week on January 15th, with prices starting at $349.
We don’t yet have top-to-bottom specifications for the card, but based on the information NVIDIA has released thus far, it looks like the GeForce RTX 2060 is based on a cut-down version of the TU106 GPU that’s already being used in the GeForce RTX 2070. This is notable because until now, NVIDIA has used a different GPU for each RTX card – TU102/2080TI, TU104/2080, TU106/2070 – making this the first such card in the family. It’s also a bit of a shift from the status quo for GeForce xx60 parts in general, which have traditionally always featured their own GPU, with NVIDIA going smaller to reduce costs.
|NVIDIA GeForce Specification Comparison|
|RTX 2060 Founders Edition||GTX 1060 6GB||GTX 1070||RTX 2070|
|Memory Clock||14Gbps GDDR6||8Gbps GDDR5||8Gbps GDDR5||14Gbps GDDR6|
|Memory Bus Width||192-bit||192-bit||256-bit||256-bit|
|Single Precision Perf.||6.5 TFLOPS||4.4 TFLOPs||6.5 TFLOPS||7.5 TFLOPs
FE: 7.9 TFLOPS
|Manufacturing Process||TSMC 12nm "FFN"||TSMC 16nm||TSMC 16nm||TSMC 12nm "FFN"|
|Launch Price||$349||MSRP: $249
In any case, let’s dive into the numbers. The GeForce RTX 2060 sports 1920 CUDA cores, meaning we’re looking at a 30 SM configuration, versus RTX 2070’s 36 SMs. As the core architecture of Turing is designed to scale with the number of SMs, this means that all of the core compute features are being scaled down similarly, so the 17% drop in SMs means a 17% drop in the RT Core count, a 17% drop in the tensor core count, a 17% drop in the texture unit count, a 17% drop in L0/L1 caches, etc.
Unsurprisingly, clockspeeds are going to be very close to NVIDIA’s other TU106 card, RTX 2070. The base clockspeed is down a bit to 1365MHz, but the boost clock is up a bit to 1680MHz. So on the whole, RTX 2060 is poised to deliver around 87% of the RTX 2070’s compute/RT/texture performance, which is an uncharacteristically small gap between a xx70 card and an xx60 card. In other words, the RTX 2060 is in a good position to punch above its weight in compute/shading performance.
However TU106 has taken a bigger trim on the backend, and in workloads that aren’t pure compute, the drop will be a bit harder. The card is shipping with just 6GB of GDDR6 VRAM, as opposed to 8GB on its bigger brother. The result of this is that NVIDIA is not populating 2 of TU106’s 8 memory controllers, resulting in a 192-bit memory bus and meaning that with the use of 14Gbps GDDR6, RTX 2060 only offers 75% of the memory bandwidth of the RTX 2070. Or to put this in numbers, the RTX 2060 will offer 336GB/sec of bandwidth to the RTX 2070’s 448GB/sec.
And since the memory controllers, ROPs, and L2 cache are all tied together very closely in NVIDIA’s architecture, this means that ROP throughput and the amount of L2 cache are also being shaved by 25%. So for graphics workloads the practical performance drop is going to be greater than the 13% mark for compute throughput, but also generally less than the 25% mark for ROP/memory throughput.
I also have some specific concerns here about the inclusion of just 6GB of VRAM – especially in an era where game consoles are shipping with 8 to 12GB of VRAM – but this is something we can look at later with the eventual review.
Moving on, NVIDIA is rating the RTX 2060 for a TDP of 160W. This is down from the RTX 2070, but only slightly, as those cards are rated for 175W. Cut-down GPUs have limited options for reducing their power consumption, so