Blackmagic Forum

MishaEngel wrote:We are in the lucky situation that we can test almost every piece of computer hardware before we purchase it and we do. It used to be AMD+NVidia then Intel+NVidia(for a very long time) now it's AMD+AMD and I have no idea what it will be in the future. For our current workflow(Davinci Resolve) this combination suites us best. We have no brandloyalty whatsoever, we simply don't have the budget.

We shoot with RED Helium, UMP G1, Micro Studio 4k Sony A7S2 and eng-style news camera's at 1080i50.

Exactly. During the Bulldozer era it would have been extremely hard to recommend AMD for gaming or video. But now, since 2017 and Zen1, and now Zen2, they have caught up and surpassed Intel. Zen 2 is a better architecture than anything Intel has to offer, except in some niche cases like gaming @ 720p with a 2080Ti.

Talking about stability regarding AMD CPUs in 2020 is a complete joke. Its like they are completely oblivious regarding the CPU industry the past 3-4years.

John Paines wrote: I wonder if you could say the same?

That depends on the type of knowledge they seek.

anyway...back to the topic at hand, the announcement of the new GTX range....

I'm picking that even though its PCIe 4.0, it will have backward comparability with PCIe 3.0 and will run - question is, how well. Better then a 2080ti? I know we will probably have to wait to find out but I'm curious

Jeff Brass wrote:anyway...back to the topic at hand, the announcement of the new GTX range....

I'm picking that even though its PCIe 4.0, it will have backward comparability with PCIe 3.0

Just remember to force PCIe 3.0 in the bios if you have any stability issues if you run the card in a 3.0 board. PCIe 4.0 could act weird on some 3.0 boards, be it Intel or AMD.

I remember reading somewhere that the bandwidth limit of the PCIx gen3 was approached by the 2080ti, but cant recall in what workload, if in gaming or whatnot
Assuming that is true, 3070 and maybe 3080 might be still safe on gen3
3090 better be on gen4
But on Resolve might that be also tied to the source video bandwidth x resolution and effects applied?

kobayashi wrote:I remember reading somewhere that the bandwidth limit of the PCIx gen3 was approached by the 2080ti, but cant recall in what workload, if in gaming or whatnot
Assuming that is true, 3070 and maybe 3080 might be still safe on gen3
3090 better be on gen4
But on Resolve might that be also tied to the source video bandwidth x resolution and effects applied?

PCIe 3.0 8x. 16x not yet. Need to wait for benchmarks and see if the uplift in performance is close to saturating 16x. My personal opinion is ...not yet. Should be negligible differences in most cases from 3.0 to 4.0 with 4.0 having a small edge. There was a youtuber that made some tests and 4.0 helped cards with like 4GB. Like the 5500XT.

Hi.

Here is a link to a nVidea Q&A on Reddit:

https://www.reddit.com/r/nvidia/comment ... _answered/

It have some nice informations:

Quote: '[Tony Tamasi] One of the key design goals for the Ampere 30-series SM was to achieve twice the throughput for FP32 operations compared to the Turing SM. To accomplish this goal, the Ampere SM includes new datapath designs for FP32 and INT32 operations. One datapath in each partition consists of 16 FP32 CUDA Cores capable of executing 16 FP32 operations per clock. Another datapath consists of both 16 FP32 CUDA Cores and 16 INT32 Cores. As a result of this new design, each Ampere SM partition is capable of executing either 32 FP32 operations per clock, or 16 FP32 and 16 INT32 operations per clock. All four SM partitions combined can execute 128 FP32 operations per clock, which is double the FP32 rate of the Turing SM, or 64 FP32 and 64 INT32 operations per clock........
More details on the NVIDIA Ampere architecture can be found in NVIDIA’s Ampere Architecture White Paper, which will be published in the coming days.'

So I expect, that we all have to wait, to find out what the fp32 Performance are for Resolve?

Regards Carsten.

So we can say the real CUDA cores count is half of what announced, with (hopefully) twice the performance in most scenarios?
Huh, marketing

kobayashi wrote:So we can say the real CUDA cores count is half of what announced, with (hopefully) twice the performance in most scenarios?
Huh, marketing

Mm, from what do you deduce this? Description above is about instruction execution, it doesn’t say anything about the actual core count. I understand it as same amount of cores being able to double fp32 performance.

Only "educated speculation" on my part,
I noticed the number of cuda cores in the official specs are exactly doubled from what most past leaks indicated.
Now I read that the performance in some workloads is doubled per cuda core.
Maybe the marketing department has reconsidered arbitrarly what constitutes a cuda core? I dont know
Also there are teraflops all over the place in the spec sheet...
We need benchmarks!

All things aside, I think both GPU teams will deliver with the new releases this year.

After browsing various threads on the Internet it looks like the extra cores are for real not just marketing gimmick - but we have to wait for the benchmarks and breakdown.

The icing on the cake - the Ampere series is going to support SR-IOV which means a VM will get full access to the GPU acceleration. We will now be able to run a Linux box with a Windows VM and get access to GPU h/w from within or vice-versa.

This means I will be upgrading as soon as the dual fan series shorter cards are available locally to fit my chassis.

G0bble wrote:After browsing various threads on the Internet it looks like the extra cores are for real not just marketing gimmick - but we have to wait for the benchmarks and breakdown.

The icing on the cake - the Ampere series is going to support SR-IOV which means a VM will get full access to the GPU acceleration. We will now be able to run a Linux box with a Windows VM and get access to GPU h/w from within or vice-versa.

This means I will be upgrading as soon as the dual fan series shorter cards are available locally to fit my chassis.

Depends on how you look at things. They changed the way the SM works from Turing. Turing could do int32 and fp32 ops per clock while Ampere can do int32/fp32 OR 2xfp32. However the SMs haven't changed between Turing and Ampere. They just got more stuff added to it. We'll need to see how this actually translates into real world performance when benchmarks arrive. It won't be linear that for sure.

Rumours are that the doubling of shader performance TFlops counts for around 30% in real FP32 workloads(like resolve).
(RTX3090) 35.6 TFlops shader performance would be around (35.6/2) x 1.3 ~ 23 Tflops (FP32), that is a lot. Even more important with effects is the memory bandwidth of 936 GB/s.

MishaEngel wrote:Rumours are that the doubling of shader performance TFlops counts for around 30% in real FP32 workloads(like resolve).
(RTX3090) 35.6 TFlops shader performance would be around (35.6/2) x 1.3 ~ 23 Tflops (FP32), that is a lot. Even more important with effects is the memory bandwidth of 936 GB/s.

Assuming the 35.6 figure is correct in the first place. Which is where the doubt comes. The compute ampere A100 has only 19TFLOPS from 6912 non doubled cores. Somehow the 3090 has more FP32 if you do /2 (*2) than the big brother compute card with 1664 less cores? Highly doubt that.

Looks like the new cards are faster, but the number of CUDA cores is not an apples to apples comparison. Despite having double the cores of the 2080 Ti, the 3080 gets about 140% performance. For me, the higher memory bandwidth is more interesting.

https://videocardz.com/newz/nvidia-gefo ... benchmarks

Jack Fairley wrote:Looks like the new cards are faster, but the number of CUDA cores is not an apples to apples comparison. Despite having double the cores of the 2080 Ti, the 3080 gets about 140% performance. For me, the higher memory bandwidth is more interesting.

https://videocardz.com/newz/nvidia-gefo ... benchmarks

People who sold their 2080Tis for peanuts thinking 3070 will destroy it might be in for a nasty cold shower.

Hi.

Videocardz.com have published rumors of the RTX 3080 CUDA & OpenCL Performance. And are Expecting Reviews of the RTX 3080 September 14th, with custom models reviews 3 days later.

[url][/url]

From: https://videocardz.com/newz/nvidia-gefo ... benchmarks

Regards Carsten.

G0bble wrote:The icing on the cake - the Ampere series is going to support SR-IOV which means a VM will get full access to the GPU acceleration. We will now be able to run a Linux box with a Windows VM and get access to GPU h/w from within or vice-versa.

Would you mind to share the source of that info?
I'm genuinely interested, and hope that will be true, but I strongly doubt that Nvidia will be unlocking workstation/datacenter class features on regular desktop parts.

Wasn’t single-user direct hardware access also possible before? Afaik grid etc cards offer hardware virtualization where multiple vm-s get direct hardware access of same card and this differentiation would probably continue.

Yep, but I'm aware its avilable only on datacenter class cards, Nvidia Tesla, Radeon MI50 and the likes.
Dont know about workstation class like Quadro and Radeon Pro

kobayashi wrote:

G0bble wrote:The icing on the cake - the Ampere series is going to support SR-IOV which means a VM will get full access to the GPU acceleration. We will now be able to run a Linux box with a Windows VM and get access to GPU h/w from within or vice-versa.

Would you mind to share the source of that info?
I'm genuinely interested, and hope that will be true, but I strongly doubt that Nvidia will be unlocking workstation/datacenter class features on regular desktop parts.

https://twitter.com/RyanSmithAT/status/1301993263835086849
Not sure if there is a mixup between A100 datacenter gards and GA10* consumer cards but this is the source.

That would be really nice
hoping that they wont limit to the flagship 3090, like it seems to be for nvlink

We won't know for sure until AMD releases the first RDNA2 cards, but every rumor I have seen to date is that AMD won't even try to compete on performance with Ampere. RDNA2 is also going to trail Ampere by 4 - 6 months. Where RDNA2 might be useful is in the hardware acceleration for codecs as Ampere really isn't doing any thing amazing on that front. If RDNA2 had AV1 encode it would certainly get my attention but I don't even expect AV1 decode. AMD is now saying that RDNA3 is where the big performance gains will be made but they have been selling the wait for the next release story for as long as they have been around. Go back a year and RDNA2 was going to be the savior. Go back another year and RDNA was going to be.

As I run on Linux NVidia is a real problem for me but so is a crappy graphics card. My RX580 just doesn't cut it any more now that I have moved to the BM6k and shoot RAW. So time is moving like a snail for me while I wait for the details to come out.

lost_soul wrote:We won't know for sure until AMD releases the first RDNA2 cards, but every rumor I have seen to date is that AMD won't even try to compete on performance with Ampere. RDNA2 is also going to trail Ampere by 4 - 6 months. Where RDNA2 might be useful is in the hardware acceleration for codecs as Ampere really isn't doing any thing amazing on that front. If RDNA2 had AV1 encode it would certainly get my attention but I don't even expect AV1 decode. AMD is now saying that RDNA3 is where the big performance gains will be made but they have been selling the wait for the next release story for as long as they have been around. Go back a year and RDNA2 was going to be the savior. Go back another year and RDNA was going to be.

As I run on Linux NVidia is a real problem for me but so is a crappy graphics card. My RX580 just doesn't cut it any more now that I have moved to the BM6k and shoot RAW. So time is moving like a snail for me while I wait for the details to come out.

RDNA2 and Zen 3 are launching this year, its been repeated over and over. RDNA2 before consoles come.
out. RDNA2 in consoles is 2080Super performance (as quoted by devs) so what makes you think they won't compete?

Also this whole selling the wait is nonsense. Wait for what? They didn't have a competitor for 1080ti / 2080ti in Games....who cares about that? In the segments where they had cards they competed well.

RX580 (an almost 4year old cards is still relevant today) was faster and with more ram than the competitor 1060 6GB (not including that abomination 1060 3gb that released first)
RX570 still has no competition
Vega64 is faster than 1080 in games now. In compute there's no match.
Vega56 is faster than the 1070 competition.
Radeon VII trades blows in games with 2080 and smokes a 2080ti in compute and has more ram and insane bandwidth. Still not matched in bandwidth. Not even by the 3090.
5700XT is faster than 2070
5600XT is faster than 2060...

Sounds like competition on performance to me...
All they need is to release more SKUs.

Hi.

I want to share this foto of a RTX 3090 in a PC Cabinet from chiphell.com with you. The RTX 3090 is huge:



From: https://translate.google.com/translate? ... id45811329

Regards Carsten.

Carsten Sellberg wrote:Hi.

I want to share this foto of a RTX 3090 in a PC Cabinet from chiphell.com with you. The RTX 3090 is huge:

Ah well, looks like it's time to take a hammer to the ATX-mini case then!

Carsten Sellberg wrote:Hi.

I want to share this foto of a RTX 3090 in a PC Cabinet from chiphell.com with you. The RTX 3090 is huge:

I guess you don't have to worry about card sag if the card is sitting on the drive bays

Hi,

And what about the Quadro and the news RTX30XX ?

the memory ?

regards,

pperquin wrote:Hi,

And what about the Quadro and the news RTX30XX ?

the memory ?

regards,

Hi.

Resolve do all its Image processing in fp32. Quadros have its advantages in fp64. So for Resolve is there no reason to get a more expensive Quadro.


Here is a foto where you compare the size of the 3 different RTX30XX's.




From: https://translate.google.com/translate? ... id45811329

And from the same link an another one, where you compare the RTX 3090 to the size of a RTX 2080:




Quote: 'GDDR6X does come with one somewhat immediate drawback however: capacity. While Micron has plans for 16Gbit chips in the future, to start things out today they’re only making 8Gbit chips'

From: https://www.anandtech.com/show/16057/nv ... 0-rtx-3090

Personally I first expect to see the "twice the vram" cards, when Micron change, from the its new current 8Gbit chips to the next 16Gbit chips. It will take a while. I am sorry, but I expect we have to wait until sometime in 2021.

Regards Carsten.

New leaks from videocards.com on 2080Ti vs 3080 : ''On average than RTX 2080 Ti is 24% slower than RTX 3080, RTX 2080 SUPER is 37% slower, RTX 2070 SUPER – 45%, and RTX 2060 SUPER is 50% slower.''

So much for "Twice the performance of 2080''?

People need to stop selling their 2080Tis for peanuts.

I am waiting for tests from Puget Systems.

Kenzo wrote:I am waiting for tests from Puget Systems.

Puget Systems used to be pretty good, testing hardware with Resolve, these days they mainly use crappy codecs(AVC, H265 etc..) and high frame rates.
I prefer the Slashcam.de methode.

https://www.slashcam.de/artikel/Praxis/Nachgelegt---Tabelle-zum-Blackmagic-RAW-CPU-GPU-Vergleich.html

https://docs.google.com/spreadsheets/d/e/2PACX-1vQYJhAzw-RAQj9TscjFPOu3CcQwL0hahCuW6Wm-VIPLZTjLnY2MSJl6O422Zc7kkxAJ3E_N0nfOLVhN/pubhtml?gid=0&single=true

https://www.slashcam.de/artikel/Test/Der-Preis-der-zusaetzlichen-Leistung---Zwei-und-mehr-GPUs-unter-DaVinci-Resolve.html

Far from perfect, but more useful.

MishaEngel wrote:

Kenzo wrote:I am waiting for tests from Puget Systems.

Puget Systems used to be pretty good, testing hardware with Resolve

Have they? All I remember is them testing/pushing mainly what they sell while for the rest kept making excuses like ''its been discontinued'' (Radeon7). Their test methodology was subpar as well. Using different RAM kits and size and settings for platforms in the same ''test''.

IvanovS wrote:Have they? All I remember is them testing/pushing mainly what they sell while for the rest kept making excuses like ''its been discontinued'' (Radeon7). Their test methodology was subpar as well. Using different RAM kits and size and settings for platforms in the same ''test''.

You may be right, but I think they are more reliable than game tests.

Kenzo wrote:

IvanovS wrote:Have they? All I remember is them testing/pushing mainly what they sell while for the rest kept making excuses like ''its been discontinued'' (Radeon7). Their test methodology was subpar as well. Using different RAM kits and size and settings for platforms in the same ''test''.

You may be right, but I think they are more reliable than game tests.

As Misha pointed out there are other sources that aren't that biased into directly selling you the product they test. And its best to check more than one source when purchasing anything.

IvanovS wrote:New leaks from videocards.com on 2080Ti vs 3080 : ''On average than RTX 2080 Ti is 24% slower than RTX 3080, RTX 2080 SUPER is 37% slower, RTX 2070 SUPER – 45%, and RTX 2060 SUPER is 50% slower.''

So much for "Twice the performance of 2080''?

People need to stop selling their 2080Tis for peanuts.

For Resolve, CUDA score is more apt. That gap seems to be bigger than the gaming results.

MishaEngel wrote: Puget Systems used to be pretty good, testing hardware with Resolve, these days they mainly use crappy codecs(AVC, H265 etc..) and high frame rates.
I prefer the Slashcam.de methode.

https://www.slashcam.de/artikel/Praxis/Nachgelegt---Tabelle-zum-Blackmagic-RAW-CPU-GPU-Vergleich.html

https://docs.google.com/spreadsheets/d/e/2PACX-1vQYJhAzw-RAQj9TscjFPOu3CcQwL0hahCuW6Wm-VIPLZTjLnY2MSJl6O422Zc7kkxAJ3E_N0nfOLVhN/pubhtml?gid=0&single=true

https://www.slashcam.de/artikel/Test/Der-Preis-der-zusaetzlichen-Leistung---Zwei-und-mehr-GPUs-unter-DaVinci-Resolve.html

Far from perfect, but more useful.

Hi.

We don't all read German. Here is a link to Google Traslated versions of Misha first and last link:

https://translate.google.com/translate? ... leich.html

https://translate.google.com/translate? ... solve.html

Regards Carsten.

G0bble wrote:https://twitter.com/RyanSmithAT/status/1301993263835086849
Not sure if there is a mixup between A100 datacenter gards and GA10* consumer cards but this is the source.

Turns out Nvidia has clarified this is incorrect news. While the h/w supports its they have no plans of enabling it in Gameforce drivers. I guess that means not studio drivers as well. Enterprise drivers would not exist in any case as the Nvidia sales agreement prohibits use of GTX/RTX in datacenters.

Hi.

nVidea RTX 3080 Reviews Delayed to September 16th at 6am PT. Here is a link:

https://www.tomshardware.com/news/nvidi ... october-15

Regards Carsten.

Hi.

Here are links to the first Nvidia GeForce RTX 3080 Founders Edition Reviews. Partner RTX 3080 Reviews, will arrive in a few days:

https://www.tomshardware.com/news/nvidi ... 080-review

https://www.pcmag.com/reviews/nvidia-ge ... rs-edition

Regards Carsten.

MishaEngel wrote:

Kenzo wrote:Puget Systems used to be pretty good, testing hardware with Resolve, these days they mainly use crappy codecs(AVC, H265 etc..) and high frame rates.
I prefer the Slashcam.de methode.

You had my hopes up, but these tests are all over the place... different systems used to test video cards, so no consistent baseline at all. The explanations are also deficient.

Would be very happy to see some other benchmark source, to compare to Puget.

IvanovS wrote:New leaks from videocards.com on 2080Ti vs 3080 : ''On average than RTX 2080 Ti is 24% slower than RTX 3080, RTX 2080 SUPER is 37% slower, RTX 2070 SUPER – 45%, and RTX 2060 SUPER is 50% slower.''

So much for "Twice the performance of 2080''?

Raw technical numbers are up to twice as fast. This translates to real-world performance 24-50% faster than gen 2 cards. As usual, marketing plays fast and loose with numbers. This is how it always works. Why are you surprised?

Results seem outstanding from here. Power consumption less so, but I guess that was the strategy... pour more watts into more transistors.

I will be tentatively recommending the RTX 3080 for 4K work. Looking forward to more tests and real user experiences.

robin0112358 wrote:

IvanovS wrote:New leaks from videocards.com on 2080Ti vs 3080 : ''On average than RTX 2080 Ti is 24% slower than RTX 3080, RTX 2080 SUPER is 37% slower, RTX 2070 SUPER – 45%, and RTX 2060 SUPER is 50% slower.''

So much for "Twice the performance of 2080''?

Raw technical numbers are up to twice as fast. This translates to real-world performance 24-50% faster than gen 2 cards. As usual, marketing plays fast and loose with numbers. This is how it always works. Why are you surprised?

Results seem outstanding from here. Power consumption less so, but I guess that was the strategy... pour more watts into more transistors.

I will be tentatively recommending the RTX 3080 for 4K work. Looking forward to more tests and real user experiences.

Can we give it a rest with the ''up to'' cringy stuff? Reviews are up. Its one hot garbage series. 3080 is about 25% faster than 2080Ti with 25% more power consumption. 102die vs 102die. Absolutely no improvement and even some regression in ray tracing.

Only thing I personally was surprised about was the amount of people selling their 2080Tis with 11GB on ebay for peanuts on the marketing bs that a 8GB 3070 is faster than a 2080Ti.

I'm personally keeping my Radeon VII after these reviews...lol.

IvanovS wrote:Can we give it a rest with the ''up to'' cringy stuff? Reviews are up. Its one hot garbage series. 3080 is about 25% faster than 2080Ti with 25% more power consumption. 102die vs 102die. Absolutely no improvement and even some regression in ray tracing.

I'm not sure what you are reading and comparing. But it is more than 25%, especially if you compare it to 2080 (non-Ti).

mpetech wrote:

IvanovS wrote:Can we give it a rest with the ''up to'' cringy stuff? Reviews are up. Its one hot garbage series. 3080 is about 25% faster than 2080Ti with 25% more power consumption. 102die vs 102die. Absolutely no improvement and even some regression in ray tracing.

I'm not sure what you are reading and comparing. But it is more than 25%, especially if you compare it to 2080 (non-Ti).

Why would I compare it with a 2080? Its a 102die vs a 102die. Sorry if Huang bamboozled people with a rename but the 3080 is a 80Ti class. Whatever else they will launch as a 3080Ti (if it will be launched at all)will prolly just have more VRAM and minor adjustments.

IvanovS wrote:

mpetech wrote:

IvanovS wrote:Can we give it a rest with the ''up to'' cringy stuff? Reviews are up. Its one hot garbage series. 3080 is about 25% faster than 2080Ti with 25% more power consumption. 102die vs 102die. Absolutely no improvement and even some regression in ray tracing.

I'm not sure what you are reading and comparing. But it is more than 25%, especially if you compare it to 2080 (non-Ti).

Why would I compare it with a 2080? Its a 102die vs a 102die. Sorry if Huang bamboozled people with a rename but the 3080 is a 80Ti class. Whatever else they will launch as a 3080Ti (if it will be launched at all)will prolly just have more VRAM and minor adjustments.

The price point? I think you are ignoring that huge fact. At the end of the day, die version has little to do with the end-user.

mpetech wrote:
The price point? I think you are ignoring that huge fact.

Ahh yes, the price point. Huang increases the usual 600ish high end stuff to 1200-1500 then he gracefully reduces the price to 699 (799 more likely for AIB stuff) and we all jump and praise this move. Amazing. How is any of this stuff impressive?

At the end of the day, die version has little to do with the end-user.

You got me there. At the end of the day nothing really matters...Deep.

IvanovS wrote:

mpetech wrote:
The price point? I think you are ignoring that huge fact.

Ahh yes, the price point. Huang increases the usual 600ish high end stuff to 1200-1500 then he gracefully reduces the price to 699 (799 more likely for AIB stuff) and we all jump and praise this move. Amazing. How is any of this stuff impressive?

At the end of the day, die version has little to do with the end-user.

You got me there. At the end of the day nothing really matters...Deep.

I'm confused by your answers. So I will let it be...

I guess AMD is going to blow nVidia to dust with new cards, as they do with every release. What was AMDs upper hand again... lower price? Or better drivers, better feature set or better raw power. I guess everyone is dissapointed if AMD doesn’t deliver 2x performance for half the price now, they are the better one afterall.

mpetech wrote:

IvanovS wrote:

mpetech wrote:
The price point? I think you are ignoring that huge fact.

Ahh yes, the price point. Huang increases the usual 600ish high end stuff to 1200-1500 then he gracefully reduces the price to 699 (799 more likely for AIB stuff) and we all jump and praise this move. Amazing. How is any of this stuff impressive?

At the end of the day, die version has little to do with the end-user.

You got me there. At the end of the day nothing really matters...Deep.

I'm confused by your answers. So I will let it be...

What exactly confuses you, maybe I can translate it in simpler terms for you?