Here, opening a can of worms........GeForce QC.

Nasgul

n00b
Joined
Jun 11, 2005
Messages
54
So, it's been what? 14 years or so since the Dual Core processor came out and 2 years later or so the Quad Core but it's been since I started building PCs in 2000 when I bought a Riva TNT2 and here we are today with 8, 12 core processors BUT NEVER EVER has a Dual Core or Quad Core GPU ever been released or even talked about it.

And by this, I mean Cores.....not two GPUs in the same mainboard or two GPUs slapped in two mainboards.

For the stinking price of $1,200 flagship video card, s__t! I'd be better if it had more than ONE core, DIE or whatever it'd be called. Nevertheless.......
 
depends what type of core we're talking about. a titan RTX has 4608 CUDA cores, 576 tensor cores, 72 RT cores...
Yes, and to add to the confusion cores from one brand differ in functionality to other brands, and even between models. By the way, it's not completely true that it's never been talked about, AMD/ati has released boards with two physical chips on board tied together (kind of like a built in cross fire) as has Nvidia.
https://graphicscardhub.com/dual-gpu-graphics-cards/
 
depends what type of core we're talking about. a titan RTX has 4608 CUDA cores, 576 tensor cores, 72 RT cores...
"CUDA core" is a misnomer. Those "cores" inside SM block are just execution units that execute the same instructions on multiple data kinda like SIMD units on CPU. This is what you typically want in graphics oriented workloads, just process bunch of pixels in exactly the same way. Of course these "cores" are a little bit more advanced with various degree of freedom but the general idea is the same and such core does not contain everything that is needed to execute code.

If we are talking about multiple GPU cores that are equivalent to CPU cores then the term is SM blocks for Nvidia and Compute Units for AMD.
In theory you could have whole Nvidia GPU with just one SM block and it should work just fine.

Besides, in modern processor you also have multiple execution units per core but we do not call those as "cores". In CPU's what we usually want is higher IPC so those execution units are trying to help to get as high IPS as possible but we can also utilize those execution units using SIMD extensions in a way very similar to how CUDA "cores" are used withing SM block. It is not exactly the same thing and capabilities are somewhat different but general idea is the same.
 
A $1,200 CPU is going to give you around 200 GFLOP/s single precision arithmetic performance. A 2080 Ti gets around 13 TFLOP/s single precision. I don't know what we're arguing about here. GPU and CPU architecture are completely different and made for different tasks. Besides, as mentioned above a 2080 Ti would have 68 "cores" compared to the 16 you would find in a similarly priced CPU.
 
"CUDA core" is a misnomer. Those "cores" inside SM block are just execution units that execute the same instructions on multiple data kinda like SIMD units on CPU. This is what you typically want in graphics oriented workloads, just process bunch of pixels in exactly the same way. Of course these "cores" are a little bit more advanced with various degree of freedom but the general idea is the same and such core does not contain everything that is needed to execute code.

If we are talking about multiple GPU cores that are equivalent to CPU cores then the term is SM blocks for Nvidia and Compute Units for AMD.
In theory you could have whole Nvidia GPU with just one SM block and it should work just fine.

Besides, in modern processor you also have multiple execution units per core but we do not call those as "cores". In CPU's what we usually want is higher IPC so those execution units are trying to help to get as high IPS as possible but we can also utilize those execution units using SIMD extensions in a way very similar to how CUDA "cores" are used withing SM block. It is not exactly the same thing and capabilities are somewhat different but general idea is the same.
regardless, the OP's point is completely moot.
 
So, it's been what? 14 years or so since the Dual Core processor came out and 2 years later or so the Quad Core but it's been since I started building PCs in 2000 when I bought a Riva TNT2 and here we are today with 8, 12 core processors BUT NEVER EVER has a Dual Core or Quad Core GPU ever been released or even talked about it.

And by this, I mean Cores.....not two GPUs in the same mainboard or two GPUs slapped in two mainboards.

For the stinking price of $1,200 flagship video card, s__t! I'd be better if it had more than ONE core, DIE or whatever it'd be called. Nevertheless.......

I think you need to study up on GPUs. Lots of hints in the replies you’ve received already.
 
How do you make dual core processor to avoid any confusion?...

This is how:
main-qimg-5b266a1bf23f826f2016bd9d940f9afb.jpg
 
I don't get this topic. You're comparing two different components, as video cards and CPUs behave differently and serve different purposes.

If video cards had multiple GPU dies on chips, it wouldn't matter since things have to be coded for that to work (look at all the past dual-gpu single cards that have released in the past - they are limited by developer multi-gpu support). Also, other things like technology being immensely improved since the 2000's, and inflation is a thing, as well as supply/demand and let's not forget these are luxury items - people pay those prices for these things. So that's why?
 
Let me rephrase the OP:

"I want more performance for my stinkin' twelve hundred dollars!!!"
Except you're already getting the absolute fastest card on the market with that $1,200. By a mile compared to the competitor's current offerings.
 
Except you're already getting the absolute fastest card on the market with that $1,200. By a mile compared to the competitor's current offerings.
Maybe your mile is longer than my mile ;). Don't get me wrong, 20% is a good bump in performance, but I wouldn't call it a mile. Here's to hoping AMD can close that gap, and NVIDIA can bring some of that speed down to us lowly folk who can't send >$1k for a video card.
Edit:. Disclosure, Radeon VII is typically around 15-25% slower, seems to vary a lot depending on game. 5700 xt is a bit slower than that.
 
Maybe your mile is longer than my mile ;). Don't get me wrong, 20% is a good bump in performance, but I wouldn't call it a mile. Here's to hoping AMD can close that gap, and NVIDIA can bring some of that speed down to us lowly folk who can't send >$1k for a video card.
Edit:. Disclosure, Radeon VII is typically around 15-25% slower, seems to vary a lot depending on game. 5700 xt is a bit slower than that.
1589299349864.png
 
I dunno, I looked at about 4 different sites and they all had them closer than what your showing in most games, some where lopsided of course. Anyways, I didn't sum the totals.and compare, just from browsing and seeing the differences in a couple games. Maybe I glanced to quickly and missed some games that had large differences. Or possibly i was focusing more on1440p since that is more realistic resolution to play with the vii. Or I'm just slow and was reading the charts wrong, lol. At 40% faster I would consider that pretty substantial, so I guess the Mile metric would stand.
 
  • Like
Reactions: N4CR
like this
If someone thinks performance/price of 2080Ti is not good then there are better options:
performance-per-dollar_3840-2160.png
 
Chiplets would have to split-up the load into multiple pieces

There are scaling issues with Alternate-frame Rendering in DX11. And Explicit Multi-GPU is barely supported in DX12.
Until you have support for applications that can scale perfectly, you have no market for GPU chiplets.


Nvidia/AMD tried to hack their way into multil-GPU, but they eventually gave-up. By the time you setup local memory for each chip, you might as well make a single die. There is no cost savings splitting up GPU into multiple dies, unless you can figure out how to share the same memory (think console APUs). Since high-performance GPUs are VERY LIMITED on memory bandwidth, we're not going that route for multiple GPUs anytime soon!

Nvidia DGX Supercomputer platforms are the closest you're going to get to that (and they STILL use separate GPUS/ separate local memory on that!)
 
Last edited:
If someone thinks performance/price of 2080Ti is not good then there are better options:
View attachment 244993
And I would think most of us here understand that you're getting diminishing returns for your dollar the higher in the performance bracket you go. It's been like that for as long as I can remember. You just can't expect to target 60 FPS at 4K on a GTX 1660 without making significant sacrifices in image quality.
 
3dfx nailed multi GPU with their SLI, "It just works"
Compatibility with 99% games, no game profiles, near ~100% scaling in all games and most importantly it is only SFR so no input lag increase, no frame pacing issues and no micro stuttering bullshit.
If Nvidia SLI also just worked then I bet we would see quite a lot of cards using more than one chip and a lot more people would use multiple cards.
Each chip still had to have its own memory and keep copy of all textures and own framebuffers, albeit smaller because each card rendered at half (or quarter in case of Voodoo 6 6000) resolution.

GeForce 256, the card that Voodoo 5 5500 competed with had four pixel pipelines so in fact it was already a "quad core" chip.
It quite exactly opposite to what OP was claiming and GPU did it first.
With CPU's it was always an option also. Nothing technically stopped eg. Intel from making dual Pentium Pro/II/III if they really wanted. They did not want it because at this time processor clocks were rising in such fast pace it would not make economical sense to double/quadrupple transistor budget and run in to TDP issus (which would be silly compared to today...) on something most software/markets could not utilize, especially when they already had the easier means to increase performance with clocks. Also most software people used could not utilize it at OS or any level and for server and workstation markets you could always put more processors.

With GPU's because driver could easily split workload between cores and which was transparent from software point of view and clock scaling was much harder it made more sense to go with multiple cores/pipelines.
 
I was hoping we'd get into a chirping contest about pipelines.

XoR_ has the right idea :p
 
Scan-Line Interleave only worked as well as it did because games were so simple to render. And all the geometry was handled by the CPU.

You still had to have separate 4MBx4MBx4MB 192-bit memory spaces reserved for each card: so it ended-up being VASTLY more expensive than the 16MB 128-bit Voodoo 3 2000 (same performance as 24MB Voodoo 2 SLI).

The OP asked about why we don't have multiple dies on the same video card, and the response is: you're still going to have to provide dedicated memory for each die. Given the small market for such EXTREME performance/pains getting it to scale in modern games, most people are just going to wait around until the next generation single-die cards release, so there's no inherent demand.

People were a lot more willing to experiment in the the 1990s, but now, they know a bad value when they see it (you buy SLI now, you wait twice as long to get new features. And in the meantime, you have to keep buying an SLI-capable system every tme you upgrade.) SLI on a single card didn't improve performance, nor did it lower the costs of the rest of your system.
 
Last edited:
The OP asked about why we don't have multiple dies on the same video card, and the response is: you're still going to have to provide dedicated memory for each die. Given the small market for such EXTREME performance/pains getting it to scale in modern games, most people are just going to wait around until the next generation single-die cards release, so there's no inherent demand.
With AMD approach (and Nvidia in high end) of integrated memory and GPU interposers, it is now becoming less of a barrier to entry. Intel is already banking on this approach and it's only a matter of time. I hope it works, would be good to see some reasonable pricing for Nvidia again.
As process costs increase and scaling begins to trail off, discreet mGPU (invisible to host system) will become the way forward.
 
Yup, I think eventually intel's approach with much more capable stacking and interposers could lead to some very interesting gpu designs, with potentially lower costs than the traditional monolithic die.
 
Yup, I think eventually intel's approach with much more capable stacking and interposers could lead to some very interesting gpu designs, with potentially lower costs than the traditional monolithic die.


HBM has already solved the "how do we feed this thing, in incredibly tiny form-factor" problem for the last five years.

We still have yet to find a mainstream consumer-oriented l use-case that will give those multiple GPU cores on the same package a reason for their existence. That is the reason why nobody is paying the premium fon consumer-level HBM cards anymore

Even Intel had to dream-up a use-case nobody cares about like Lakefield to justify all their FOVEROS work.

Even Ice Lake doesn't make use of the external eDRAM anymore. Their stack engineers were bored and looing for something to do.. why not create pointless competitions for ARM mixed-core chips?

Until you have a plan, you're just shooting the breeze. That is what Intel has been doing, as there is no mass-market use case for CPU die stacking, much like there is no easy use case for stacking multiple GPUs on the same package. The only viable die stacking has been the traditional horizontal CPU placement used by Zen 2/Pentium 4/Core 2 Duo.

Leave package stacking for complex circuits that actually benefit from it, not for inventing some new DOA mobile x86 device product category like MIDs was?

https://en.wikipedia.org/wiki/Mobile_Internet_device
 
Last edited:
Having more dies in the package (realized in any way) would not necessarily translate to reducing cost or any performance gains for end user. We are still mostly TDP and memory (bandwidth and cost) limited. You cannot increase TDP indefinitely because sooner than later you run in to much higher card manufacturing cost and higher failure rates which are not good for business.

Also keep in mind that most revenue for eg. Nvidia is from HPC markets where margins are much higher than for consumer products and that there is limited manufacturing capability in manufacturing fabs to produce physical dies. It would make absolutely zero sense for them to lower prices for consumer products and risk having issues with chip shortages. If Nvidia had unlimited production capabilities it might make some sense to be more aggressive with pricing to get most out of consumer GPU market. Because of all that I would not expect RDNA2 to be cheaper either. Maybe AMD will have more incentive to be cheaper to regain some market share but it won't be by much.
 
Also keep in mind that most revenue for eg. Nvidia is from HPC markets where margins are much higher than for consumer products and that there is limited manufacturing capability in manufacturing fabs to produce physical dies.

False. Unhappy gamers like to invent the "gaming is just a side business for Nvidia" story because they need some sort of villain to hate. Unfortunately for them, that oft-repeated story is demonstrably false.

https://wccftech.com/nvidia-q3-fy-2020-financial-earnings-strong-gaming-demand/
1590373014582.png
 
False. Unhappy gamers like to invent the "gaming is just a side business for Nvidia" story because they need some sort of villain to hate. Unfortunately for them, that oft-repeated story is demonstrably false.

https://wccftech.com/nvidia-q3-fy-2020-financial-earnings-strong-gaming-demand/
View attachment 248012


Sorta,

Gaming is not a side business for NVIDIA. The total revenue over the course of the year is probably 50% higher than datacenter.

But Gaming has much lower margins than Datacenter. This profitability is NOT covered by NVIDIA's revenue reports.

Most estimate I've read put datacenter profit margins at at least 2x that of consumer cards.
That means you would have to make TWICE AS MUCH YEARLY REVENUE to turn a larger profit from gaming.
 
Gaming is not a side business for NVIDIA. The total revenue over the course of the year is probably 50% higher than datacenter.

But Gaming has much lower margins than Datacenter. This profitability is NOT covered by NVIDIA's revenue reports.

Most estimate I've read put datacenter profit margins at at least 2x that of consumer cards.
That means you would have to make TWICE AS MUCH YEARLY REVENUE to turn a larger profit from gaming.

Take another look. Gaming does pull in twice the annual revenue.
 
Take another look. Gaming does pull in twice the annual revenue.

Not any more. This last quarter gaming only pulled in 18% more revenue than HPC.

https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-first-quarter-fiscal-2021

Gaming
  • First-quarter revenue was $1.34 billion, down 10 percent sequentially and up 27 percent from a year earlier.
Data Center
  • First-quarter revenue was $1.14 billion, up 18 percent sequentially and up 80 percent from a year earlier.
 
With Ampere Cloud and Data Center products soon to be out I would expect an utter explosion in this area for Nvidia. I do not know of much competition coming that compares to Nvidia Ampere products here.

Nvidia Ampere gaming products? Competition will be AMD? As for Intel I see nothing on the horizon or anything that would threaten either one. With AMD future commitments, focus and great success recently, indicating very good management, AMD is likely to affect Nvidia Gaming brand of sells more so then in the past. Any Nvidia unique innovations for the upcoming gaming cards as well as AMD's will have to be seen and then the market to respond.

With the downturn in the World economy I would expect OEM, Other and Auto to decline (Other most likely Nintendo Switch could increase, especially if a newer more robust version comes to market).

The gaming market drives innovation, gives solid base or clout for large volume purchases which will influence foundry support, control and long term standings with a number of manufacturers for best prices and support, gives large presence overall in mindshare and control of resources mostly from developers, OEMs, retailers etc. that will also promote your products, name independently with no additional cost to Nvidia. I expect Nvidia to be a savage dealing with their gaming market, a falter there would have repercussions throughout Nvidia.
 
In theory you could have whole Nvidia GPU with just one SM block and it should work just fine.
The Quadro K420 has just one SM block in it, and yes, it does work well; @ stock clocks, it gets slightly less than 300 GFLOPS FP32, but it is also from 2014.

A $1,200 CPU is going to give you around 200 GFLOP/s single precision arithmetic performance. A 2080 Ti gets around 13 TFLOP/s single precision. I don't know what we're arguing about here. GPU and CPU architecture are completely different and made for different tasks. Besides, as mentioned above a 2080 Ti would have 68 "cores" compared to the 16 you would find in a similarly priced CPU.
Eh, the 6700K @ 4GHz quad-core has 512 GFLOPS FP32 (pre-security firmware patches), and modern $1000+ CPUs (depending on the core-counts) can net upwards of 2+ TFLOPS FP32 - this is peak/optimized performance on what the hardware itself is capable of, and still dependent of software optimizations to take full advantage of that.
Even the Jaguar in the PS4 @ 1.6GHz (8-cores) is capable of ~102 GFLOPS FP32, and that is slightly less powerful than a Haswell i3 dual-core @ 3GHz.

Aside from that, agreed with the rest!
 
Last edited:
  • Like
Reactions: XoR_
like this
Back
Top