Big Navi might still be faster than the 3080

[CarlBar]

55 minutes ago, AluminiumTech said:

Yeah that's what I was alluding to. A Titan version of the top Quadro card.

 

And how often has that happened in the past? I don't see NVIDIA doing it unless they get desperate.

[cj09beira]

53 minutes ago, AdmiralMeowmix said:

Don't hype up based on these alone because old Navi had better specs on paper than nvidia's offerings, by al laccoutns it should have performed amazingly well. But it didn't and was performing way udner spec. So I cant trust AMD's "specs" claims.

 

no it didn't, it had much lower specs than radeon VII, and performed close to it, has similar specs to the 2070 super and performs close to it, what did it not do really?

[AlTech]

11 minutes ago, CarlBar said:

 

And how often has that happened in the past?

GTX Titan

GTX Titan Black

GTX Titan Z

 

GTX Titan X (Maxwell)

GTX Titan Xp

GTX Titan X Pascal

 

RTX Titan

[cj09beira]

guys this should be obvious but..

50% better perf/W doesn't mean much higher performance, only much lower power, as to increase performance the power needed is exponential the question is how exponential it is.

from consoles power consumption we know they pretty much achieved their 50% claims, as to how fast we will have to wait.

 

btw one of the leakers now said that the 256bit bus might have been for emulation, though i doubt it as we saw the pcb with a 256bit bus which fitted the 3 fan cooler

[AlTech]

4 minutes ago, cj09beira said:

guys this should be obvious but..

50% better perf/W doesn't mean much higher performance,

 

only much lower power, as to increase performance the power needed is exponential the question is how exponential it is.

from consoles power consumption we know they pretty much achieved their 50% claims, as to how fast we will have to wait.

 

btw one of the leakers now said that the 256bit bus might have been for emulation, though i doubt it as we saw the pcb with a 256bit bus which fitted the 3 fan cooler

 

Idk if anybody else has picked up on this but TechPowerUp is now updating their GPU database with 3 new RDNA2 cards although TechPowerUp says these specs are a work in progress.

 

RX 6900XT (which I've said how much I hate the name for this as a 3080 competitor) with 80 CUs 16GB G6, 256 bit bus.

 

RX 6800XT with 60CUs, 12GB G6, 192 bit bus.

 

RX 6700XT with 40CUs, 10GB G6, 160 bit bus.

[porina]

5 hours ago, Jet_ski said:

Nvidia like Intel has reached some sort of tech limitation where they are resorting to making the chips bigger to get more performance while drawing much more power.

I don't recall the numbers but I don't think the Ampere chips are that different in size from Turing. They are big, but they are also doing a lot.

 

5 hours ago, Deli said:

Tech Yes City mentions undervolting the 3080. He drops the core clock by about 100MHz. Easily reduces power consumption by 100w, to ~220w. Performance only drop by about 3%.

 

The question is why Nvidia needs to go way outside of its norm of 250w TDP with the 3080. They can keep the same old 250w TDP and the 3080 would be almost as fast.

The biggest problem with undervolting is the same problem with overclocking, you are operating the device outside of its intended operating area. If it works for you and your use cases, great. But the device as sold has to work in every possible scenario, some of which may need that extra voltage to be stable.

 

AMD have forced Intel and nvidia to run closer to technology limits, and I think that is a negative for those of us who value stability more than absolute performance. Backing off even a little bit from the curve helps reduce power consumption a lot more than the difference in performance.

 

Having said that, this problem also exists with AIB factory overclocked GPUs. I recall my XFX 280X was always borderline turning itself into a fireball until I removed the factory OC through bios mod, when it ran much cooler. I also have an EVGA 970 which is fine in gaming loads but totally unstable in compute, again unless I remove the factory OC.

[AlTech]

Just now, porina said:

I don't recall the numbers but I don't think the Ampere chips are that different in size from Turing. They are big, but they are also doing a lot.

They're a bit smaller but they would have been much smaller on TSMC.

Just now, porina said:

The biggest problem with undervolting is the same problem with overclocking, you are operating the device outside of its intended operating area. If it works for you and your use cases, great. But the device as sold has to work in every possible scenario, some of which may need that extra voltage to be stable.

 

AMD have forced Intel and nvidia to run closer to technology limits, and I think that is a negative for those of us who value stability more than absolute performance. Backing off even a little bit from the curve helps reduce power consumption a lot more than the difference in performance.

I disagree. Nvidia did this to themselves. Nobody grabbed a gun and pointed it at Nvidia's head telling them to use an inferior 8nm Samsung Low Power node as opposed to the perfectly good TSMC N7 or N7P 7nm nodes.

 

At this point if you want a 30 series then what you're buying are cards pushed past their efficiency curves because of the decision to use Samsung. The clockspeeds are wayyy too high for any meaningful kind of increased efficiency from Turing.

[hollyh88]

Yeah lets first see if they can even launch a card without its bios completely broken.

[porina]

 

7 minutes ago, AluminiumTech said:

I disagree. Nvidia did this to themselves. Nobody grabbed a gun and pointed it at Nvidia's head telling them to use an inferior 8nm Samsung Low Power node as opposed to the perfectly good TSMC N7 or N7P 7nm nodes.

There's a lot more to the decision process than going that's a good one, we'll go with that. You might as well say they should have elbowed Apple out of the way and stuck it on TSMC 5nm, cos that'll be even more better.

 

7 minutes ago, AluminiumTech said:

At this point if you want a 30 series then what you're buying are cards pushed past their efficiency curves because of the decision to use Samsung. The clockspeeds are wayyy too high for any meaningful kind of increased efficiency from Turing.

I kinda agree with this, but we've had this from all sides especially since Ryzen era. Most consumer grade hardware is pushed for performance not efficiency. 

[cj09beira]

32 minutes ago, AluminiumTech said:

 

Idk if anybody else has picked up on this but TechPowerUp is now updating their GPU database with 3 new RDNA2 cards although TechPowerUp says these specs are a work in progress.

 

RX 6900XT (which I've said how much I hate the name for this as a 3080 competitor) with 80 CUs 16GB G6, 256 bit bus.

 

RX 6800XT with 60CUs, 12GB G6, 192 bit bus.

 

RX 6700XT with 40CUs, 10GB G6, 160 bit bus.

ya those are just wrong, as the 40 cu card is the one on 192bit bus confirmed on the linux kernel, and i dont think amd would release a X9XX card if it wont at least clearly beat the 3080 which with hbm it should, but with 256 bit bus i dont know (both mentioned on the linux kernel).

its normal for them to do this though, they go with what is rumored then later on add the official specs, they could have done a better job though

[Blademaster91]

I doubt it would beat a 3090, but I'll wait for benchmarks.

Also the rumored 6800XT, 192bit bus seems wrong to be competing with the RTX 30 series, unless AMD plans to use HBM that would drive up cost way too much to have 16GB.

[ravenshrike]

9 hours ago, Rune said:

You only get coolers if you preorder. Excellent I like it.

In fairness, if they started selling naked boards that would allow for true third party coolers and easier access for the plebs to aftermarket water cooling. Which would almost certainly mean a proliferation of 120mm fan compatible air coolers instead of the 2 that are on the market.

[BuckGup]

I was going to get an RTX3090 but I think I might go with AMD this time because I don't care at all about ray tracing

[tim0901]

9 hours ago, epicdotexe said:

A friend of mine got his 3080 today. We were trying out some games and benchmarks, naturally. Every single application crashed or had stuttering issues.

There have been reports of people having problems, in particular on factory overclocked models, suggesting it's a board partner problem.

 

Some have had success by simply installing the studio drivers, instead of the game ready drivers. Others have found installing the broadcaster software seems to fix it. The most reliable fix seems to be dialing back the core frequency by ~100MHz - not ideal, but at least it'll be useable while waiting for a proper fix.

 

That being said, most only had problems in select titles, not in literally everything they tried. Your friend could have a PSU issue or a corrupt driver, or potentially just be unlucky with a dud card.

 

Until we know more about the issue, I wouldn't go making sweeping statements that the cards as a whole are 'barely funcitonal'.

[Rohith_Kumar_Sp]

Good, i hope the prices of 3080 go down even more because of this.

[Bombastinator]

One thing worth remembering is the 3080 and the 3090 use the same die.  The 3080 is just crippled a bit, and the 3090 may be capable of more speed than it claims.  It’s clocked lower than the 3080 and runs quiet.  I don’t think it’s possible to know what the top end of ga102 might be until the 3090s get quite a bit of messing with.  
 

there is also the driver deficit.  AMD is hampered by its small size.  The cards aren’t a s commonly bought which means fewer opportunities for faults to come up and fewer people to work on solutions even when faults are found.  Nvidia can actually release more buggy stuff, but because they can fix it faster they have the power of opinion.  By the time Navi is ready to release the faults now being found in amphere will be fixed and forgotten about.  Nvidia is getting to release buggy cards with no non buggy card to compare to.  Navi will not have that luxury. AMD could release what was actually a less buggy card but it would still be thought of as being more buggy. 

[schwellmo92]

8 hours ago, AdmiralMeowmix said:

Don't hype up based on these alone because old Navi had better specs on paper than nvidia's offerings, by al laccoutns it should have performed amazingly well. But it didn't and was performing way udner spec. So I cant trust AMD's "specs" claims.

 

Wtf are you talking about? Navi specs line up very closely to 2070S specs and 2070S is 6%~ faster 1080p and 7%~ faster 1440p.

 

5700 XT vs RTX 2070S

Transistors: 10.3b vs 13.6b

CU/SM: 40 vs 40

ROPs: 64 vs 64

Mem B/W: 448GB/s vs 448GB/s

Power: 222w vs 213w

 

With Ampere it is going to be hard to compare CU to SM as they have doubled the cuda cores per SM so I would assum a NVIDIA SM is more performant than an AMD SM (to what degree we wont know till we see the AMD offering). But keep in mind the 3080 has 68 SM's and "Big Navi" is purported to have 80 CU's (SM's). "Big Navi" is also looking like it will have less traditional memory bandwidth to GDDR VRAM however leaks suggest it may have an "infinity cache" which we don't know much about. I wouldn't say 3080 performance is out of the question, but we'll have to wait and see to find out.

[Samfisher]

25 minutes ago, epicdotexe said:

I would go that far, and every consumer they ship a barely functional card to will also go that far

 

This launch has been a catastrophe. The card was tested by us under Windows 10 (we were going to try Server 2016, of course Nvidia has made Server OS drivers a "Quadro Feature" out of the kindness of their hearts), then when that failed under Debian (what we both run) and Fedora when that failed as well.

 

Specific problem applications we experienced were Quake 2 RTX (the Nvidia demo for the card, lol), Blender using Cycles render, LuxMark, LuxRender and LuxCoreRender (decades of compatibility broken!), and SpecViewPerf. That's problems with BOTH CUDA and OpenCL compute! Impressive!

 

I wound up running the same compute tests on my Quadro M6000 24GB to validate that the software we were using wasn't the problem. They all ran perfectly.

 

The fact that reviewers were somehow able to benchmark their cards successfully makes me think they received specifically validated (read: working) hardware while consumers are receiving the barely functional slop units that have real, observable problems.

 

These cards are Nvidia's Vega: hot and buggy.

So you have a single card and you're generalizing for the entire world.  I like your sample size.

[Briggsy]

7 hours ago, BiG StroOnZ said:

 

Yeah, I agree.

 

Also (and obviously this is speculation and could be completely off base), but I think there might be a few others based on the same silicon [GA102]; with varying frame buffer sizes (VRAM) being the real differentiators between SKUs/Models (and slightly less CUDA Cores when compared to the full fat die/chip/GPU/offering).

 

RTX 3090                            [  10,496 CUDA Cores,   24GB    ]

RTX 3090 Super                [   10,560 CUDA Cores, ¿xxGB?  ]

RTX 3090 Ti                      [    10,688 CUDA Cores, ¿xxGB?  ]   

RTX "Ampere" Titan         [     10,752 CUDA Cores,    48GB   ] 

There may not be an Ampere Titan unless it's using TSMC 7nm or Samsung 5nm.

 

With the Exception of Turing at 280 watts, all Titan cards have had a 250 Watt TDP. Nvidia as a company have a lot of pride, and I doubt even Jensen wanted to have the stain of a 350 watt Titan on an otherwise great track record. The speculation I've heard is an 8nm Ampere Titan would have to use regular GDDR6 non-x memory to keep the power below 300 watts, along with a decrease in base and boost clocks. Anyone thinking of buying that for gaming has rocks in their head for sure XD. not even worth thinking about.

[schwellmo92]

-delete-

[ravenshrike]

2 minutes ago, Samfisher said:

So you have a single card and you're generalizing for the entire world.  I like your sample size.

Well, GN hasn't run their workstation tests yet because they were too busy ripping nVidia's marketing department a new asshole for their 8k gaming claims.

[Samfisher]

Just now, ravenshrike said:

Well, GN hasn't run their workstation tests yet because they were too busy ripping nVidia's marketing department a new asshole for their 8k gaming claims.

Like all marketing claims, it's always "best case scenario".  Like their claim the 3080 was double the performance of a 2080.  It was.  In 1 game with specific settings, in RTX.  Technically true, but misleading all the same.

[AndreiArgeanu]

28 minutes ago, Samfisher said:

So you have a single card and you're generalizing for the entire world.  I like your sample size.

There are widespread issue with these RTX cards, so it's not exactly a 1 in a million case.

 

[RejZoR]

RTX 3080 results were not misleading. It actually scores double the framerate in basically all modern games. Not so much in older ones where it seems to be much more CPU bound, but still scores very respectable numbers. You just have to understand that new games are apparently coded differently and they achieve higher framerates when you have the GPU power and doesn't fall down to CPU as much as it did in the past. And that's great. It means your GPU matters more instead of also having to pair it with stupid expensive CPU just to fully use already stupid expensive GPU.

[Spotty]

16 minutes ago, AndreiArgeanu said:

There are widespread issue with these RTX cards, so it's not exactly a 1 in a million case.

There were also issues with Nvidia's previous gen at launch as well. Have people forgotten the space invaders memory artifacts?

Don't be a beta tester.