NVIDIA Presents the TITAN RTX 24GB Graphics Card at $2,499

[Vya Domus]

I am not even stressing about this anymore, it is what it is...AMD nowhere near the performance so no competition so these numbers are not that shocking anymore.

May I point out how the 1000$ Titan brand was introduced 5 years ago ? And how the 8800 Ultra who was also nearly 1000$ ,5 years prior ? Surely within that time span AMD was at least a couple of times competitive, so what gives ? Competition does not drive prices down by it self, this has been proven time and time again, just look at the history of every industry out there. Take smartphones, the competition there is insane yet what do you know, we also got 1000$ phones.

This 1k$ figure shows up a lot it seems, why do you think that is ? It's because of the competition ? Nah, it's because of your wallet. Matter of the fact is AMD could come out with a card 10x faster , if they price it the same as Nvdia, it will sell. Because currently, that's the buying power of those types of customers.

Companies do not compete for who has the best product or who sells most products or who can sell them for cheapest, they ultimately compete for who can have the biggest profits.

Who said anything about CPU? Would you buy four TITAN RTX cards for $10,000 (4x 144 = 576 FP64 cores), or one Tesla V100 for the same price (2560 FP64 cores)?

That's not the point of what I said. ~500 GFLOPS of DP is still a lot extra compute performance should you need it, it's not as useless as you may think it is. Of course there are cards out here with more DP performance.

 

[unikin]

I was looking forward to release of Titan RTX hoping to get some cheaper 64-bit double precision math performance out of it and now this: Titan RTX = 0.51 TFLOPS Double Precision FP64, my old Titan V is giving me 6.9 TFLOPS FP64! This is a huge downgrade for anyone doing math modeling. Yet another NVidias' product segmentation. They deliberately crippled double precision performance on Titan RTX. The only current choice remaining for us pro users is Tesla V100 with $10K price tag. I paid $8.5K for triple Titan V GPUs giving me nearly 21 TFLOPS of FP64 and I'd have to pay $30K for 21 FP64 performance now. Where is the logic in that? NVidia has gone completely loco with pricing. When my Titans V die on me I'll go for AMD VEGA Instinct MI60 if price is more normal. If not, I'm f...ed.

 

[EarthDog]

snip

AMD has always been competitive in the price to performance market, but not so much in that high-end space. I think their $1500 flagship (r9 295x2) was the last thing they had in the high end space... and that was a dual GPU solution using 500W of power. The 8800 Ultra debuted with an MSRP of $829. I suppose 18% off of $1000 is 'nearly' (as in horseshoes, hand grenades, and nuclear war... close enough). The 290x out of the box was 13% slower than a 780 Ti without using its 'uber' mode and 8% with it (according to TPU review of 780Ti). That, IIRC, was about the closest they came in those 10 years.

Competition is in fact one variable which can (but not always) drives prices down (the absolute nature of you words make that passage incorrect). Another variable about pricing is what the market supports (as you said). If there is zero competition at that performance segment and the market supports it.. guess what.... we have pricey cards. But it isn't a case of the chicken or the egg.

I was looking forward to release of Titan RTX hoping to get some cheaper 64-bit double precision math performance out of it and now this: Titan RTX = 0.51 TFLOPS Double Precision FP64, my old Titan V is giving me 6.9 TFLOPS FP64! This is a huge downgrade for anyone doing math modeling. Yet another NVidias' product segmentation. They deliberately crippled double precision performance on Titan RTX. The only current choice remaining for us pro users is Tesla V100 with $10K price tag. I paid $8.5K for triple Titan V GPUs giving me nearly 21 TFLOPS of FP64 and I'd have to pay $30K for 21 FP64 performance now. Where is the logic in that? NVidia has gone completely loco with pricing. When my Titans V die on me I'll go for AMD VEGA Instinct MI60 if price is more normal. If not, I'm f...ed.

What about the DP performance on the RTX Quadros? Are those any different?

 

[Konceptz]

NVIDIA's greatest leap since the invention of the CUDA GPU in 2006 and the result of more than 10,000 engineering-years of effort,

Wait, what? 10,000 years!?!?!

 

[unikin]

What about the DP performance on the RTX Quadros? Are those any different?

I don't know because they are not out yet. RTX 8000 will have 8 TFLOPs with $10K price tag, maybe RTX 6000 will be an option with the same TFLOPs perf selling for $6.3K on preorders. But we're still talking about 107 % price increase for 12 % more performance. Titan V has been poor mans' Quadro and Nvidia probably doesn't like it, that's why they decided to cripple Titan line.

 

[EarthDog]

Wait, what? 10,000 years!?!?!

If there were 10 engineers who put in 1000 hours on the card. Yes. That passage is written properly.

I don't know because they are not out yet. RTX 8000 will have 8 TFLOPs with $10K price tag, maybe RTX 6000 will be an option with the same TFLOPs perf selling for $6.3K on preorders. But we're still talking about 107 % price increase for 12 % more performance. Titan V has been poor mans' Quadro and Nvidia probably doesn't like it, that's why they decided to cripple Titan line.

I'm just wondering about the RTX 5000 for $2300. If that is, say 2 TFLOPS, perhaps it can be a viable option? I don't know.

 

[Konceptz]

If there were 10 engineers who put in 1000 hours on the card. Yes. That passage is written properly.

LOL I was like huh? I know i'm sleep deprived but....

 

[unikin]

I'm just wondering about the RTX 2000 or w/e for $2300. If that is, say 2 TFLOPS, perhaps it can be a viable option? I don't know.

I need at least 20 TFLOPS per comp for modeling unfortunately and even this is on the slow side when you have client's deadlines to catch. 3 Titans V per computer are OK. We have 17 such PCs in our offices, so I'm looking at $320K instead of $150K investment when Titans start dying.

 

[Vayra86]

Wait, what? 10,000 years!?!?!

Yeah that reads like engineering years are dog's years or something

 

[EarthDog]

I need at least 20 TFLOPS per comp for modeling unfortunately and even this is on the slow side when you have client's deadlines to catch. 3 Titans V per computer are OK. We have 17 such PCs in our offices, so I'm looking at $320K instead of $150K investment when Titans start dying.

Further market segmentation FTL.

EDIT: That FP64 though... literally is the only place where its slower than the Titan V (right?)...

 

[jabbadap]

Wait, what? 10,000 years!?!?!

Well i.e. thousand engineers tackling with the task for ten years makes it 10 000 engineering-years.

 

[diatribe]

Yeah that reads like engineering years are dog's years or something

Work Year = ~ 2,080 work hours

10,000 * 2,080
= 10,000 work years
OR
20,800,000 hours.

 

[unikin]

Further market segmentation FTL.

EDIT: That FP64 though... literally is the only place where its slower than the Titan V (right?)...

Yet very important one. Unfortunately the accuracy of double precision compute is a must in precise physics modelling. If you can get away with FP16, FP32 compute precision Titan RTX will be slightly better than Titan V, but as I said most physics simulation&modelling or high accuracy financial computation require capable FP64 GPUs nowdays.

Btw: FP64 : FP32 ratio is very, very low on Titan RTX. Usual ratio lies between 1:2 and 1:3, Titan RTX has abysmal 1:32 ratio almost exactly as 3yo Quadro M6000. One more proof of deliberate market segmentation.

 

[Vya Domus]

high accuracy financial computation require capable FP64 GPUs nowdays.

You'd never ever want to use floating point be it even of double precision for financial applications. Even if you would, CPUs would better serve these types of applications anyway due to the sequential nature of operations that need to be performed.

Unfortunately the accuracy of double precision compute is a must in precise physics modelling.

For some, sure, but most of the time you can get away with FP32 if you are careful enough to craft numerically stable algorithms. DP simply hasn't been of great focus, after all it was only with the development of Kepler and GCN that this features was even added.

 

[unikin]

You'd never ever want to use floating point be it even of double precision for financial applications. Even if you would, CPUs would better serve these types of applications anyway due to the sequential nature of operations that need to be performed.

For some, sure, but most of the time you can get away with FP32 if you are careful enough to craft numerically stable algorithms. DP simply hasn't been of great focus, after all it was only with the development of Kepler and GCN that this features was even added.

We mostly work on turbulent flow models (fluid dynamics) and we always use double precision when modeling except when data sets won’t fit in memory with double precision, algorithm does so little arithmetic that it is bandwidth limited or where we are able to use the vector hardware and single precision (16 singles in parallel instead of 8 doubles). I never worked as financial modeler, but I do read double precision is used there too.

 

[TheOne]

With the announcement and release of the Titan RTX before the 2060, and with the 2060 most likely using the price point of the xx70's $350-450, this series is meant for the high end market and not the majority of gamers, which will make RT adoption slow, but at least now PhysX is open source.

 

[Vya Domus]

but I do read double precision is used there too.

Sparingly, at best. Too many issues with rounding, questionable arbitrary conditioning and loss of money after certain repeated calculations. Fixed point (64 bit integer) reigns king.

 

[efikkan]

People need to stop complaining about the price of Titan cards, and if you don't get the point of the card, it's not meant for you!

This is not a sign of lacking competition; even AMD have their counterpart "Vega Frontier Edition", priced at double the price of Vega 64. These are cards meant for developers and researchers, not for gamers. Also remember, these cards have limited availability in the market. Partners don't necessarily pay this price when getting these cards.

 

[Gasaraki]

Well that's what happens when you have a monopoly. I say the government needs to step in a chop nVidia into multiple companies like they have in the past with other monopolies.

WTF are you on about? What is nVidia a monopoly of? The gpu market? Cause they are not the market leader. Just because they release something you can't afford, doesn't make them a monopoly. OMG I can't afford that new 8K 70" TV. What a monopoly, break them up. /s

 

[Sandbo]

how much price of tesla v100? do they sacrifice tesla card?
576 multi-precision Turing Tensor Cores, providing up to 130 teraflops of deep learning performance.

View attachment 111756

I just feel pity to people/university/corporate who own tesla based HPC.

Tesla is still worth it for DP, just not HP anymore.

 

[GoldenX]

So, Quadro price for GeForce performance.

 

[R-T-B]

Well that's what happens when you have a monopoly. I say the government needs to step in a chop nVidia into multiple companies like they have in the past with other monopolies.

I would not say we're there yet. We're certainly treading dangerous turf though.

 

[xkm1948]

Would be nice for our collaboration lab to have one of these in addition to the (now dated) 1950X

 

[XXL_AI]

2500$, HAHAHAHA.



Knowing how faster RTX cards are than Pascals, maybe a 30% increase in workloads for 2500$ is ....

its not about %30 increase on general computing performance, it's about what we can do with other cores while training the dataset.

 

[R-T-B]

You do not need governments regulating luxuries.

You do if they truly are a monopoly. Even luxuries can suffer from predatory pricing and computers aren't really a "luxery" anymore (though a Titan certainly is).

AMD still makes mainstreamish grade GPUs though. Saying this is like the AT&T monopoly is to massively overstate the situation.

why bother complaining to the govt when they do not put regulations on tech?

Because they do? Case in point, Microsoft has been bitch slapped by the USA several times in the past. Even moreso in the EU...