Titan RTX Review

Titan RTX Review


Turing Makes its Titan Appearance

Product Description

Historically speaking, we knew this was going to happen. Nvidia releases a new range of chips (GM102, GP102, TU102), they get released as Quadros as full chips, and we get a cut down chip for a flagship gaming card. There are a few variations to this over time, such as the Titan X (Pascal) being a cut chip, which was later replaced at the same price, but the Titan Xp based on the full chip.

Enter Q3 2018. Nvidia announces the Quadro RTX series, with the big daddy Quadro featuring the full fat TU102 chip. Boasting 4608 cores, 96 rops, and a staggering 24GB of memory. Just a little while later, Nvidia announces the gaming cards, based on the same chips, albeit missing a few cores. The 2080ti enters the ring, priced at 1200 but only missing a slim fraction of the hardware (enter ‘I can’t afford that’ sound bit). We knew it was coming simply because of the Quadro release, given Nvidias history with how they position their pro vs consumer cards based on the same chips. Finally, after a little PR stunt involving Tech Tuber ‘enthusiasts’, Nvidia shows the Titan RTX.

Unlike with the Pascal series, Nvidia didn’t opt in with a cut chip for consumer cards. This time, for the ‘low’ price of $2500, you yourself can have a blinged out gold Titan RTX.

“$2500! But I thought it was a gaming card!” Yeah, Nvidia treated Titans as such for a long time before this. Starting with the OG Titan, launching on a cut down GK110 chip, there was almost nothing about that card that made it a ‘prosumer’ card. It didn’t get any dedicated drivers that made it better over the later launched 780ti, in fact it was worse. Fast forward a few generations and just when Nvidia got comfortable selling Titans as premiere gaming cards, AMD launched Vega FE. While Vega didn’t beat the Titan X (Pascal) in gaming, what it did do is rub its nose in the lack of pro driver support. The illusion that anyone that thought they were getting a semi-pro GPU with the Titan quickly evaporated with the Vega beating the Titan in many pro workloads.

Nvidia took notice of this, and on this rare occasion they backed up their customers. They released true semi-pro drivers and in a very, how do I say this… ‘Nvidia’ like fashion, boasted about how new drivers increased performance up to 3x. Yes, Nvidia did a press release of sorts letting you know they were no longer kneecapping your performance on your $1200 purchase. To this day, you still only get partial support (this is true from both AMD and Nvidia), but the cards now make sense for home users that need that performance for pro workload like Maya or Solidworks.

Today, I think its safe to say, that we finally have semi pro graphics cards from Nvidia that make some sense. Since the ‘Titan’ driver released, Nvidia has since forked the drivers between Titan and Geforce giving them their own categories, and has since launched two true compute oriented graphics cards, Titan V and Titan RTX.

From a technical standpoint, the Turing architecture and manufactured size are marvels in their own right. With TU102 coming in at an astounding 754mm² and packing 18.6 billion transistors, Nvidia held nothing back in pushing TSMCs node to its limit. This chip is the second largest mass produced die ever built, slightly behind the GV100 GPU at 815mm².

So, what do you get with your $2500? Full fat TU102 packing 4608 cores, 576 Tensor cores, 72 Ray Tracing cores, 96 ROPs, 288 TMUs and that 24GB of GDDR6 we talked about earlier. This graphics card is no slouch in the specs department, but how exactly does it perform? Continue reading and find out.

Specs
Graphics CardQuadro RTX 6000Titan RTXRTX 2080 TiTitan V
CUDA Cores4608460843525120
Tensor Cores576576544640
Integer Cores4608460843525120
ROPs96968896
DP Cores0002560
RT Cores7272680
Core Clock1455135013501200
Boost Clock1770177015451455
Memory Clock700070007000850
Bus Width384b384b352b3072
Bandwidth672GB/s672GB/s616GB/s652.8GB/s
VRAM24GB24GB11GB12GB
L2 Cache6MB6MB5.5MB4.5MB
FP16 TFlops32.632.626.927.6
FP32 TFlops16.316.313.413.8
FP64 TFlops0.50.50.446.9
INT8 TIPs16.316.313.413.8
FP16 Tensor130.5130.5107.6110
FP16/32 Tensor130.5130.553.8N/A
Int8 Tensor Ops261261215.255.2
Int4 Tensor Ops52252243055.2