I enjoy John and Alex's content quite a lot. However, THIS is the stuff that got me to subscribe. Great video and thanks for taking the time to do this analysis. Having done similar work myself, I know it's time-consuming and tedious.
Excellent analysis as always Richard. A point that may be missed is memory timings on the graphics cards, despite being able to match bandwidth, the memory has timing straps which are tied to frequencies range. I would guess that downclocking memory on a NAVI GPU would simply have it used very conservative timings compared to GDDR5. By any means this does not invalidate any of the phenomenal research done here, but it is just another point to be considered. The thing that called my attention the most was GCN 1.0 poor scaling with modern APIs. There has to be some sort optimization done on the consoles for the games to keep up performance, especially on Microsoft DX12 only titles, wonder why it didn't make it to the AMD Driver on PC. Wouldn't we be able to get them to comment on this? I'm a Patreon backer, but never found a way to contact the team for other interesting stuff. Keep up the good work!
@Mitch what he wants to say is peak floating point calculation power is far telling the whole story about how GPU will perform in some software especially when comparing 2 different architectures.
Had a few other comments after my other comment. 1) You were so close to it! But I wish you'd explicitly noted that the extra CUs going from RX 570 to RX 580 offer almost nothing in performance because games are never limited by compute. In fact, as you said at the end, teraflops are a "difficult measure" of gaming performance because they really have almost nothing to do with game performance. Games are unlikely to use even half the compute power available on high-end GPUs and modern games still usually scale pretty damn close with fillrate and memory bandwidth. 2) That's the central, primary reason why Navi is faster: it has twice as many ROPs as the other cards. Simple. 3) In fact, that's true even in Ghost Recon Wildlands which you say is compute-limited but which actually sees an even smaller gain in performance from RX 570 -> RX 580 than Assassin's Creed did. The real difference between these cards in every game is almost certainly down to their boost algorithm and the RX 570's lower TDP. RX 570 is limited to 150W while the RX 580 is a 185W part; it will boost more aggressively and be less likely to downclock, even when you have the clock rate manually configured. The RX 580 also likely has more aggressive memory tuning, which brings me to 4) While you equalized memory *bandwidth* by fiddling with the clocks, the memory access times are quite different. GCN 1.0 will have quite high memory latency owing to its very low transfer rate, because it has a 384-bit memory bus where the rest of the cards in this comparison have 256-bit buses. 5) On Ultra settings, Witcher 3 heavily uses tessellation even with Gameworks disabled. That's why Tahiti runs it like a dog and doesn't see any gain in performance from dropping the resolution. You noted yourself that Polaris is over 5 times faster than Tahiti at tessellation, and that's why you see the "strange" results there. 6) GCN 1.0 hardware does not support async compute. That's a big part of the reason why it suffers in DX12 and Vulkan titles. GCN 1.1 (Hawaii/Bonaire) does. This may (?) be a driver limitation rather than a hardware limitation, but it does exist. 7) Strange Brigade is a Vulkan exhibition title, and it has unbridled triangle counts as a result. GCN 1.0's poor geometry performance is hurting it here as well. Anyway, I still enjoyed the video and the data is really useful and interesting. I would have enjoyed a big takedown video where you call out these companies for focusing on compute performance. Compute performance is completely and utterly irrelevant to game performance; GPUs with more compute perform better in games because they are simply bigger. GPUs are modular devices and adding more compute means adding more TMUs, more ROPs, and so on. In fact, I would really like to see a video that calls out AMD and Nvidia on the fact that they no longer make graphics cards -- they make compute accelerators that also happen to do graphics. You could lop half the shaders off an RX 5700 and it would run just the same as long as you kept all the other functional units. I guess that's a bit outside the domain of DF, though.
I see what you're saying here but since ray tracing and mixed precision shading are starting to become utilized, GPU compute will become much more important. However, the current designs of compute units are not made for these types of workloads so I guess it depends if AMD will actually optimize their compute units for them. Seeing that Nvidia has already started maybe that will force them to do the same.
"games are never limited by compute" that's an assumption, which was anyway tested by the results, which was the point of the test, did you even watch the video!?
and compute is not free, 90% of graphics tasks are compute for e.g lighting, culling, particles etc, on the contrary there is not enough compute especially for lighting even when you use RTX hardware you still need a lot of compute to interpret those rtx results..
Jaguar is 20% ipc increase over Vishera (fx 8350), ryzen was 53% ipc over fx 8350 (33% over jaguar), Zen+ is 3% ipc over Zen, and Zen 2 is 13~15% ipc over Zen+. So that would put Zen 2 like 78%~ ipc over fx. Zen 3 (with increased clocks), should absolutely decimate anything intel has to offer though.
This was good... Concrete evidence we can use to actually comprehend these numerous cards. I'm looking forward to your next gen comparison video. Also... Its not lost on us how big the mountainous workload was that you took on to make this video happen. We appreciate it. It's what puts Digital Foundry into its own league here on youtube. Thank you.
@@rogehmarbi My man. Gamers Nexus and Hardware Unboxed are the two tech channels I'm a patron of. I find I don't have time for GN much these days though, too long and technical, teardowns and reviews of a million different models of GPU that I'm not in the market for anyway. Mostly just watch the major product launches and HW news. DF and HWU are more interesting lately imo.
@@Malus1531 well said. I can see where you come from about GN. I usually only listen to them, I'd play their videos on the TV or the PC and do something else. Even without actually watching them I can still learn a thing or two from their videos..
@@Lukasz3332x no, 4K60 with meaningful RT is not happening any time soon on consoles. There will probably be some games that technically have RT at 4k60, but it's unlikely to be a very meaningful difference. Btw, I would be rooting for 120/144Hz support. RT I can live without. 4k is a neat upgrade, but doubling the framerate makes a world of difference. I can never go back again on PC.
Witcher 3 is heavier on geometry and raster than it is on compute, at least to a point. Tahiti has 2 geometry engines and has no DCC. This means that Polaris with 4 geometry engines and its 36% gain in effective bandwidth will beat Tahiti by a lot, since it leaves more bandwidth for the ROPs. That is why you see that result.
@@tortugatech Delta-Colour Compression. It is a bandwidth (NOT VRAM!) saving feature that is in-hardware. Example: a Polaris RX 580 gets effectively, on average (since some scenes compress a lot, others almost nothing) 36% more bandwidth to play with. Its DCC is that good. A RDNA RX 5700 XT? A lot more. Same for a Turing card or Ampere. But for example, if you got an old 7970 or a GTX 580 to play with - they dont have any DCC.
One other important note to keep in mind when speculating about future console hardware; the Xbox One X is based on Polaris (close to the RX 580), the One S and One are based on the R7 260/360 as Rich mentioned. The base PS4 is a bit slower than a 7850 (R9 270), while the PS4 Pro has graphics hardware closer to an underclocked RX 480. For those who weren't following GPU tech over the past seven or so years, the big improvement in geometry hardware for AMD was with the R9 285 (Tonga), which became the R9 380. The R9 280/280X were leftover Tahiti chips from overproduction during the first Bitcoin boom, while the R9 290/290X (Hawaii) were given an extra 4 GB of RAM and sold as the R9 390/390X with none of the Tonga improvements.
Teraflops is just a measure of "how fast we can run the same fast instruction over and over again", and not very good at all at telling what a graphic card can do, as the test don't even touch the video card memory for example, or some of the more complex instructions or how efficient the card is at dealing with code branches.
here we go again, some wizkid who thinks he knows best..... but more sounding like and AMD FANTARD...... it's AMD who throws with these numbers in the first place and fail miserably. Second, the test doesnt have to touch memory cause it's been equalized but hey.... seems you don't understand the meaning of these tests at all.... do ya?
@@vforvendetta6193 It's not about Nvidia vs AMD here, it's about people just running around with a quite useless number, pretty much like the bits of the old days. And as you pointed out, its not even a valid "nvidia vs AMD" discussion as both sides use the same stupid marketing number extensively. And finally, the video pretty much shows what i mean on the other parts, such as how the better thread scheduling, better dealing with branches etc.. make the GPU be better even if the tflops numbers and even memory bandwidth are normalized.
The problem is that fanboys want to condense the performance of a GPU architecture down to a single number, which is not possible considering the complexity of modern a GPUs and rendering techniques. Its easier to get involved in "my number is bigger than yours" dick swinging contest for both fanboys and even company PR and marketing.
Videos like this is why I love this channel ! Amazing work ! New consoles will be sweet for sure, esspecially cause of much better cpu. I would love a comparison like this but on cpu side, jaguar vs ryzen.
FLOPS are a pretty useful metric - if you're comparing like for like. But it doesn't take efficiency into account. If you're comparing different architectures, one chip might simply need more operations to get to the same result, or, as has happened with GCN and it's four way parallelism, developers simply have a very hard time keeping the chip fed.
Give me ONE reason why you should buy a measly Sony 32 bit machine instead of the TWICE as powerful N64. Bet you can't heh... Keep your wipeouts and tomb raiders, once Mario 64 and Ocarina of Time are out you'll see true next gen gaming.
The fact that a 7970 OC’d to >1GHz can still get over 30fps in really demanding situations is pretty insane. Talk about a good buy (especially the 6GB Sapphire variants) from 2011-2012. It’ll be interesting to see if RDNA pulls off something similar; AMD tends to be pretty good at longer-term future-proofing. Thanks for another fascinating video, DF!
Incredible video. I truly appreciate everything digital foundry does in all of the testing. Thank you for the hard work. I really loved this video comparison and how you tried your best to normalize the results.
Lol, ATI was used in GC, sure AMD bought out ATI, bought the CG was hardly a RX 290 lol, and yes I know it was a joke, but too many people on here wom't know.
@@UltraCasualPenguin Lol, I think you read my comment too fast, I was talking about his comment, the GC was great, even more powerful than the PS2 in just about every way, I still have an old grey GC here with a few games.
As someone still running a 280X (though I have been thinking of getting a V56) im happy to see my card in the ring. Some notes (some of which I think were explicitly mentioned in the video) and suggestions: any sane person running tahiti would prefer DX9 over DX10/11 or 12 any day. Even with the same in-game settings, you would usually see better performance using 9 (tests on DX9 wouldve been very interesting here). Some of the massive drops in performance of tahiti vs even GCN2+ as far as I remember were introduction of hardware acceleration for tesselation and texture compression (especially the former makes tahiti performance drop off a cliff). Still thanks for this video, it was fun to see how well the "fine wine" of ATI/AMD really aged.
GCN 1 in Tahiti predates the 2013 consoles' GCN 2 which brought a 4x increase in ACE units for singificantly boosted async compute performance. That's probably why you're seeing such a large difference in compute performance between Tahiti and Polaris.
I actually read the transcript online and since I've watched so many of these videos I read it in your style! This has been one of my favorite channels on youtube, keep up the good work !
To be fair, HD 7850 class hardware in 2013 was modern Mid-range hardware, not low end. 290/290X and 780/780 Ti was enthusiast class 280/280X and 770 was high end class 7850/7870 and 760 was mid range class. 7750-7770-7790 and 750/750 Ti were low end class
Super cool summary of AMD's past and current GPU architectures. Very in depth with nice testing methodology and game selection. I would love to see you guys tackle the topic of drivers for nVidia, AMD and Intel. The whole process of find a driver download, installing them, getting to the settings for the driver suit software etc. How well they support Vulkan, DX11, DX12, OpenGL (AMD is pretty bad here) and how is the image quality? Testing some games on OpenGL like Quake 3 Arena, Wolfenstein: New Order and newer stuff like Gears 5 on DX12.
5 років тому
I don't understand half of the things mentionned here but at this point I just enjoy whatever Digital Foundry does. Simply one of the best channels on UA-cam.
As much as you might want to consider some of the GCN 1 results bogus because of potential bottlenecks elsewhere causing an 'invalid' comparison of Tahiti's FLOPs and Navi's FLOPs, I think the comparisons are even more valid given those wide differences in performance. It's simple really. In situations like the Novigrad test in The Witcher 3, GCN 1 gets decimated by newer revisions of GCN, let alone RDNA. To me, that's a valid test of architectural FLOPs. That Tahiti chip is still very much capable of a theoretical 4.1 TF of compute, but the architecture design is imbalanced in a way that games like The Witcher 3 see very low performance scaling with either clock speed or memory bandwidth, as determined by your 1440p performance being basically identical. The Radeon geometry pipeline has improved dramatically over the years, and newer architectures clearly get more out of their TFLOP rating when faced with geometry and tessellation heavy workloads. It would have been cool to see both the R9 380X and R9 390 included in this comparison. The R9 380X uses the Tonga chip based on GCN 3 with 2048 shaders, while the R9 390 uses the Hawaii chip based on GCN 2 with 2560 shaders. In the same way you projected the results of GCN 1 vs GCN 4 onto RDNA, you could have potentially then projected the same results against GCN 2 and GCN 3, but it would have been a **lot** of extra work for a purely academic result.
I'd really love to see a companion video to this with comparisons between a few generations of Nvidia GPU architectures! Thanks for another wonderful video, DF! :)
A 390 isn't gcn 1.0, neither is a 280 btw, I don't think they know that despite the internet fury the 280 is not the exact same as the hd 7000 gcn 1.0's.
@@Folsomdsf2 What changed in 280 compared to 7000? As for 390, I don't care if it's GCN1.0 or not, it's far better perf/TFLOPS comparison and gives a much better picture of RDNA improvement over GCN.
Really looking forward to seeing what these new consoles can deliver, the price though is going to be £499 minimum probably higher but when you compare that to high end smartphones costing around £1000 I don’t think we can complain too much.
You can't just look at the performance and technology leaps in GPU. You have to also consider the performance gains of going from a 8 core mobile APU to a full fledged current gen, multi-threaded 8 core 16 thread CPU. And New speed of memory. It really is and amazing leap.
I really love ur content. I was calculating the Teraflop difference in Architecture between Nvidia´s GPUS vs AMD GCN and AMD RDNA to get an Idea what the next gen console will deliver with an expected amount of 10,8 Tflops. That was the Perfect video for me thanks:D.
Richard, you are quite skilled. Shout-out to Gamers Nexus. I like your all things being equal approach. The conversation in the CPU market seems to be 5 GHz is a rough upper limit on clock speed, and more cores / smaller dye is the right way to go. I'd really like an education on what a GPU actually is today, because x86 / x64 I understand. Thumbs up!
Current owner of an RTX 2080, and i have to say judging from what Richard said at the end, it is true! consoles simply give amazing results for their price and hardware and the games i show to everyone new to gaming to impress them are God of War and Gears 5, very amazing and visually stunning games, and each running on my pro and my x
Facinating stuff, this must have been a ton of work rich :) it feels like the next gen consoles will use some kind of Navi refresh but this this is a very good hint
The next-gen consoles will be hotrods, with custom Zen/Navi providing gigantic boosts to clocks and efficiency (CPU) and huge bandwidth gains (GPU) combined with hybrid SSD drive setups -- holy shit, can't wait. Equally excited about innovations on PC. What's nice is that multiplatform games will no longer be so held back by the lead SKUs (having to run correctly on base PS4/XB1) so the baseline spec is going to bump up. Finally, we'll escape from the weak 2012-era laptop grade CPU architecture (bye bye, Jaguar)! The baseline for CPU-dependent draw calls, world simulation, physics, logic, AI, etc. will progress! And that will lead to new innovation in *gameplay* - new interaction, new world sim, exciting AI, etc.!
So, in other words, VESA Adaptive Sync Compliant/AMD FreeSync V1.0 Compliant Edit: So overall, very good news indeed, I'm surprised such a feature wasn't released when the TV's launched
Conclusion is : Teraflop will not be a buzz word for next gen console wars. Because measuring teraflop number is not related to power, performance and technological advantage. Next gen will be how good the exclusives and quality of the developer is.
If we could harness the power of Wichards hand gestures, Iam sure we could achieve a breakthrough in discovering a raw power when it comes to hand gestures. Basically wichard can power an entire personal computer with the power of his hand gestures.
Okay so this might be a personal preference sort of thing and not applicable to everyone but I'll say it anyways. I love what you all do with the real-time FPS charts along the bottom and the averages along the top-right, but something I'd really like to see implemented is when you're touching on the conclusions, such as the improvement between Navi VS Tahiti at 1080p/1440p, could you show us those as charts as well instead of just text? Seeing a bar chart or similar of these figures is *far* more digestible to me than just the text alone, and sometimes I'll find myself going back in a video to grok something you've said that I otherwise could have kept up with if the information was presented a little more "legibly". Again, maybe a highly personal preference, but something I'd ask you to consider all the same.
So this is basically a video showing off the equivalent of "GPU IPC?". Not sure what a proper term this would be but I really like these types of videos.
While MHzs can not be directly related to performance out of context, FLOPS can. This metric shows how many floating point operations per second can be possibly carried out. So you can't call it a "myth". But it can be indicative only of computing workload performance, not games.
@@BlackShadowLegion yeah at least it provides how much floating points it can calculate... but in games context... its basically the same thing as mhz myth
Lol you can literally overclock your gpu and cpu and get fps gains wtf are you talking about mhz dont matter? Yet intel is still leaps ahead of amd in benchmarks simply because of frequency
@@BlackShadowLegion Tflops can only be directly related to performance when comparing two or GPU that use the SAME architecture. And even then it is not accurate and the only way to truly compare them is with real-world and synthetic benchmarks.
@@WH250398 Oooo let's have... 2002-2015: The Core Wars Who had the most CUDA/GCN cores, or... 2002-2015: The Transistor Wars I remember transistors been a bragging point at one time.
Genius idea, mind blown. I wonder if the in house team at AMD even know exact answers to the questions Richard is going for or have done these kind of tests?
What I think is that the games aren't just optimized for Tahiti, specifically. CCN have this problem, right? It's a "monster", but it requires very well made code to prevent stalls inside the pipeline. Isn't that's why it always failed to deliver the promised TFlops? One of the biggest changes in Navi is for that reason, improve efficiency and keep the pipelines going when executing games.
We need MORE competition in PC/window space. Microsoft has been almost dictating what the standard is for pc space for last what 10+ years with windows/directx, and IMO it’s become bloated and MS has gotten lazy because of very little competition, until recently with vulkan api. But when comes os, windows is still the standard. But windows/ms does not see vulkan as threat, or Linux (yet) or any games made under different os. Very interesting to see and it’s shows that RDNA 1.0 is promising. Hopefully RDNA (Navi 2?) will even be better and hopefully it help make console little better, I hope we see more 60 fps as standard on console. Yes I like both console and pc. Console because you DONT have deal with windows Bullshit lol Also note. Game engine also make huge different and some game engine are a lot better optimizes for some hardware, but some older hardware may not run as well. Some dev optimizes more or fav Nvidia, some would optimize for AMD, Hopefully RDNA will sound like it’s giving much better results, and once next gen console come out. I think we will see more games optimizes better for Ryzen amd Navi hardware. As said I was love see these test under Linux but that would be be almost “ impossible” task/challenge because of limited software/tools and even games. We also know most games under OpenGL are very horrible ports and don’t even match window version ( most, but not all) thus why I hope we see more devs use vulkan api. Good video Rich and keep up good work :)
One possible confound is the 280X having basically zero delta color compression, hence the expensive 384 bits pipe was likely needed. So it could look worse than in reality - although the Battlefield Dx11 vs. Dx12 results may speak against that.
Old AMD drivers used more CPU than it should be. A process maybe I forgot what it was exactly but it was inherited and it couldnt be easily fixed. Great video.
Thanks, Rich. Brilliant stuff. I’ll be saying goodbye to my PC when PS5/Nextbox land. GPU/CPU power is hitting a plateau with the TVs/ displays. It’s going to be 4K for a few years now and the next gen consoles will deliver all the resolution/performance that even the enthusiast gamer will need for quite some while. In a quiet little box for $500 max. Yes please. When the world goes 8K/120 I’ll think again.
This is why people shouldn't worry about PS5's 10.3 Tflop GPU. RDNA 1 was already a huge boost over GCN 1.0 which is what's inside the base PS4, but PS5 will be using RDNA 2 which will offer another boost in performance over RDNA 1. All said and done we should expect at least 7x the GPU power over base PS4 and over 3x the power of the Pro.
You know I love technology I really do. I love seeing new technology. but what I really hate is when I buy a video card for 500 + dollars two and a half years ago and now it's almost obsolete JS.
Nice video, what you could have mentioned is that the consoles use GCN1.1 or 2nd gen GCN, not the first. It might go some way to explaining the dx12 results.
I agree with most of this video, but with the BF1 I strongly disagree. For this whole gen, DICE's DX12 implementation has been disastrous, always being slower than DX11 regardless of architecture or OEM used.
Yeah its crazy how something like Uncharted 4, GoW or Horizon can run on custom 7850. Cant wait for the next gen. Even if the consoles will have only something like RX 5700 GPU.
One thing to especially keep in mind regarding memory underclocking is that this somewhat breaks the comparison to a certain extent. Due to how closely GPU operations are tied to memory performance and latencies, underclocking memory has far reaching consequences on the peak performance of the core itself, in simple terms a bottleneck. Certain compute and cache operations are closely tied to memory operations. So underclocking and overclocking inherently accrues penalties/gains for the architecture. Overall what this means is that we can't compare the GPU's true architectural performance ceiling against the benchmark pair card we are testing against we can only asymptotically approach the truth.
I didn't specifically address core clocks and the various operation & issuing latencies, pipeline stalls, cycles per ops, cache speeds, queuing and thread scheduling etc. But overall clock for clock comparisons are very foggy endeavours that reveal little useful actionable information. One could do the same with a matched power envelope performance/watt comparison and it still wouldn't provide very much useful information beyond being an exercise in academic curiosity that merely demonstrates that yes indeed the new architecture is better at the same clock speed/power envelope. Sometimes we are even ignoring that the higher clock speeds and lower power profiles of the new chip may often actually be a result of the newer architectures design itself. So you have a bizarre chicken and egg comparison problem.
The thing is, everytime we only count as valid those Tflops and bandwdth numbers , without even counting that most games are still TMUs and ROPs limited.
See those generational gains? All directly related to console hardware implementation. Go look at the specs of the new xbox project Scarlett and new PlayStation. RDNA will be the powerhouse for the next gen of consoles.
I have a 1.1 ghz at 800mv setting on my own 5700xt and the power figure reads at 50w-80w depending om usage in game then. AMD advertised same speed at 50%, power usage 14nm vs 7nm and i can confirm thats not far from the truth.
@7:37 your percentage gains are off a bit. i.e. you state Navi vs Tahiti is a 62% gain in performance (you typed +62% 1178 pts vs 655 pts). This is actually incorrect. 655 > 1178 is a 79.8% increase, or going from Navi down to Tahiti, is a 44.4% decrease.
Seems there may be an elephant in the room that nobody is addressing. For years reviewers have said AMD cards (I.E.: Polaris, Vega) are designed with more of a work station performance and Nvidia (I.E.: Pascal) was more about gaming performance. Yes, even crypto miners loved the AMD cards because of this. But this was always pointed to when we discussed gaming performance numbers. "The games are optimized for Nvidia because they hold the market share". Navi launches and reviewers say, "AMD have switched and the GPU will concentrate more on gaming aspects with fewer workstation assets". Well we know games are optimized for Nvidia and guess what. Turing is a workstation card. That is what it first launched as and that is what it is. First reports were that Turing will not come to gamers. So if games continue to be optimized for Nvidia first. Then it would seem that when AMD jumped over the fence and broke away from GCN. Nvidia jumped to the side AMD had been on. So as Nvidia influences new games to take advantage of that workstation based Turing the RDNA performance gains will dwindle and so will Pascal. People may want to hold onto those Polaris / Vega cards. Just put it on a shelf because you may need it within 2 years.
awesome analysis richard ! one of the best videos on DF so far, in my opinion. Would it be possible to make a similar analsys with nvidia and even intel graphics ?
I enjoy John and Alex's content quite a lot. However, THIS is the stuff that got me to subscribe. Great video and thanks for taking the time to do this analysis. Having done similar work myself, I know it's time-consuming and tedious.
Indeed. Rich is great. Truly old school in the best way. 💛
While this is great, DF started out as a benchmarking channel for console games and not PC.
@@BoshBargnani ...Your point..??
@@BoshBargnani
Truly?
Damn and I was always sure they started with PC games.
That said,so what?
I like Digital Foundry's neutrality.
I really wish they had done an RX 550 vs Radeon 7750 video to compare GCN 1.0 vs Polaris. Same number of SPs, TMUs, ROPs.
Excellent analysis as always Richard. A point that may be missed is memory timings on the graphics cards, despite being able to match bandwidth, the memory has timing straps which are tied to frequencies range. I would guess that downclocking memory on a NAVI GPU would simply have it used very conservative timings compared to GDDR5. By any means this does not invalidate any of the phenomenal research done here, but it is just another point to be considered. The thing that called my attention the most was GCN 1.0 poor scaling with modern APIs. There has to be some sort optimization done on the consoles for the games to keep up performance, especially on Microsoft DX12 only titles, wonder why it didn't make it to the AMD Driver on PC. Wouldn't we be able to get them to comment on this? I'm a Patreon backer, but never found a way to contact the team for other interesting stuff. Keep up the good work!
Yeap all Teraflops are not equal thanks for this video Digital Foundry.
Considering that NVIDIA often matched and beat AMD GPUs with 25-50% Tflop deficiency no one thought they were...
@Mitch what he wants to say is peak floating point calculation power is far telling the whole story about how GPU will perform in some software especially when comparing 2 different architectures.
only for games, for productivity apps TFLOPS and VRAM are everything look at radeon 7 with 16TFLOPS still kicking ampere's ass...
Had a few other comments after my other comment.
1) You were so close to it! But I wish you'd explicitly noted that the extra CUs going from RX 570 to RX 580 offer almost nothing in performance because games are never limited by compute. In fact, as you said at the end, teraflops are a "difficult measure" of gaming performance because they really have almost nothing to do with game performance. Games are unlikely to use even half the compute power available on high-end GPUs and modern games still usually scale pretty damn close with fillrate and memory bandwidth.
2) That's the central, primary reason why Navi is faster: it has twice as many ROPs as the other cards. Simple.
3) In fact, that's true even in Ghost Recon Wildlands which you say is compute-limited but which actually sees an even smaller gain in performance from RX 570 -> RX 580 than Assassin's Creed did. The real difference between these cards in every game is almost certainly down to their boost algorithm and the RX 570's lower TDP. RX 570 is limited to 150W while the RX 580 is a 185W part; it will boost more aggressively and be less likely to downclock, even when you have the clock rate manually configured. The RX 580 also likely has more aggressive memory tuning, which brings me to
4) While you equalized memory *bandwidth* by fiddling with the clocks, the memory access times are quite different. GCN 1.0 will have quite high memory latency owing to its very low transfer rate, because it has a 384-bit memory bus where the rest of the cards in this comparison have 256-bit buses.
5) On Ultra settings, Witcher 3 heavily uses tessellation even with Gameworks disabled. That's why Tahiti runs it like a dog and doesn't see any gain in performance from dropping the resolution. You noted yourself that Polaris is over 5 times faster than Tahiti at tessellation, and that's why you see the "strange" results there.
6) GCN 1.0 hardware does not support async compute. That's a big part of the reason why it suffers in DX12 and Vulkan titles. GCN 1.1 (Hawaii/Bonaire) does. This may (?) be a driver limitation rather than a hardware limitation, but it does exist.
7) Strange Brigade is a Vulkan exhibition title, and it has unbridled triangle counts as a result. GCN 1.0's poor geometry performance is hurting it here as well.
Anyway, I still enjoyed the video and the data is really useful and interesting. I would have enjoyed a big takedown video where you call out these companies for focusing on compute performance. Compute performance is completely and utterly irrelevant to game performance; GPUs with more compute perform better in games because they are simply bigger. GPUs are modular devices and adding more compute means adding more TMUs, more ROPs, and so on. In fact, I would really like to see a video that calls out AMD and Nvidia on the fact that they no longer make graphics cards -- they make compute accelerators that also happen to do graphics. You could lop half the shaders off an RX 5700 and it would run just the same as long as you kept all the other functional units. I guess that's a bit outside the domain of DF, though.
I see what you're saying here but since ray tracing and mixed precision shading are starting to become utilized, GPU compute will become much more important. However, the current designs of compute units are not made for these types of workloads so I guess it depends if AMD will actually optimize their compute units for them. Seeing that Nvidia has already started maybe that will force them to do the same.
@@456MrPeople
Mark Cerny has confirmed Hardware Raytracing for Playstation 5 a day or several ago.
"games are never limited by compute" that's an assumption, which was anyway tested by the results, which was the point of the test, did you even watch the video!?
and compute is not free, 90% of graphics tasks are compute for e.g lighting, culling, particles etc, on the contrary there is not enough compute especially for lighting even when you use RTX hardware you still need a lot of compute to interpret those rtx results..
@@niks660097 You didn't read my comment. Don't waste my time.
Kudos for calling Steve Burke brilliant.
Any chance of adding a Time Stamp? For ease of finding it :)
@@longnamedude3947 2:18
@@arjunyelamanchili3517: Thx, mate :)
Hell yeah tech Jesus!
All hail Tech Jesus
I'd like to see a similar video DF showing the most important difference between the old Jaguar CPU, FX 8350 and Zen 2.
That would be an amazing video! Especially because the jump would be at least twice as large than the gpu one
Yeah, the jump between Jag and Zen is huge and I'd like to get a lot of it put into perspective.
Jaguar is 20% ipc increase over Vishera (fx 8350), ryzen was 53% ipc over fx 8350 (33% over jaguar), Zen+ is 3% ipc over Zen, and Zen 2 is 13~15% ipc over Zen+.
So that would put Zen 2 like 78%~ ipc over fx. Zen 3 (with increased clocks), should absolutely decimate anything intel has to offer though.
@@dondraper4438 Did jaguar really got 20% improvements over the original fx? Would you mind share a link to that info?
@@dondraper4438 "Jaguar is 20% ipc increase over Vishera (fx 8350)" dude what!!?
This was good... Concrete evidence we can use to actually comprehend these numerous cards. I'm looking forward to your next gen comparison video.
Also... Its not lost on us how big the mountainous workload was that you took on to make this video happen. We appreciate it. It's what puts Digital Foundry into its own league here on youtube. Thank you.
6 seconds in I noticed 6 hand gestures this mans on point!
Dayle Baynes He ordered them hands from ‘CUTTING EDGE GAME-‘ oh...wrong channel
6 seconds in and my panties are already wet
Not his hands, there is somebody behind him poking their arms through.
That’s because Wichard is the Gesture Jester Witcher Master.
And not a single one of his gestures matched any of the words coming from his mouth. It's like he's trying to fuck with us.
Love this sort of content. Can't get it anywhere else on UA-cam, thanks for all the hard work!
Incredible analysis. Only UA-cam channel I support on Patreon and this is why.
Absolutely fantastic - and a great amount of work too I'm sure. Digital Foundry is in a league of its own.
DF, GamersNexus, and HWUB, that's my holy trinity of PC related content
@@rogehmarbi My man. Gamers Nexus and Hardware Unboxed are the two tech channels I'm a patron of. I find I don't have time for GN much these days though, too long and technical, teardowns and reviews of a million different models of GPU that I'm not in the market for anyway. Mostly just watch the major product launches and HW news. DF and HWU are more interesting lately imo.
@@Malus1531 well said. I can see where you come from about GN. I usually only listen to them, I'd play their videos on the TV or the PC and do something else. Even without actually watching them I can still learn a thing or two from their videos..
This is quality content. Thanks DF!
On old gen games: 30% faster IPC GCN1 vs RDNA.
Latest games + next-gen: 88-130% faster IPC GCN1 vs RDNA
Next-gen will be bloody amazing.
do you think 60fps/4k (not necessary native)+RT is possible?
@@Lukasz3332x no, 4K60 with meaningful RT is not happening any time soon on consoles. There will probably be some games that technically have RT at 4k60, but it's unlikely to be a very meaningful difference.
Btw, I would be rooting for 120/144Hz support. RT I can live without. 4k is a neat upgrade, but doubling the framerate makes a world of difference. I can never go back again on PC.
Lets wait and see because this gen was a joke.
@@dushikorsou1 yeah pushing 4k instead of 60fps...
Besides TFs and Power, We need a good games
Witcher 3 is heavier on geometry and raster than it is on compute, at least to a point. Tahiti has 2 geometry engines and has no DCC. This means that Polaris with 4 geometry engines and its 36% gain in effective bandwidth will beat Tahiti by a lot, since it leaves more bandwidth for the ROPs. That is why you see that result.
DCC?
@@tortugatech Delta-Colour Compression.
It is a bandwidth (NOT VRAM!) saving feature that is in-hardware.
Example: a Polaris RX 580 gets effectively, on average (since some scenes compress a lot, others almost nothing) 36% more bandwidth to play with. Its DCC is that good.
A RDNA RX 5700 XT? A lot more. Same for a Turing card or Ampere. But for example, if you got an old 7970 or a GTX 580 to play with - they dont have any DCC.
@@CharcharoExplorer thanks for the info, are you a CS student / dev, or just in the know?
@@tortugatech In the know. I also write articles for Sapphire Nation at times and GND Tech. Its a hobby for me really :P
@@CharcharoExplorer any advice on where can i learn more about GPU / real time rendering stuff?
One other important note to keep in mind when speculating about future console hardware; the Xbox One X is based on Polaris (close to the RX 580), the One S and One are based on the R7 260/360 as Rich mentioned. The base PS4 is a bit slower than a 7850 (R9 270), while the PS4 Pro has graphics hardware closer to an underclocked RX 480.
For those who weren't following GPU tech over the past seven or so years, the big improvement in geometry hardware for AMD was with the R9 285 (Tonga), which became the R9 380. The R9 280/280X were leftover Tahiti chips from overproduction during the first Bitcoin boom, while the R9 290/290X (Hawaii) were given an extra 4 GB of RAM and sold as the R9 390/390X with none of the Tonga improvements.
What I really like is, the RX 580 and the rx 5700 have 2304 SPs, that's almost 50% Performance uplift CU to CU, that is quite the generational leap.
Background music:
Bjørn Lynne - Over Distant Shores
Thank you
I love Bjørn!!
Teraflops is just a measure of "how fast we can run the same fast instruction over and over again", and not very good at all at telling what a graphic card can do, as the test don't even touch the video card memory for example, or some of the more complex instructions or how efficient the card is at dealing with code branches.
here we go again, some wizkid who thinks he knows best..... but more sounding like and AMD FANTARD...... it's AMD who throws with these numbers in the first place and fail miserably.
Second, the test doesnt have to touch memory cause it's been equalized but hey.... seems you don't understand the meaning of these tests at all.... do ya?
@@vforvendetta6193 It's not about Nvidia vs AMD here, it's about people just running around with a quite useless number, pretty much like the bits of the old days.
And as you pointed out, its not even a valid "nvidia vs AMD" discussion as both sides use the same stupid marketing number extensively.
And finally, the video pretty much shows what i mean on the other parts, such as how the better thread scheduling, better dealing with branches etc.. make the GPU be better even if the tflops numbers and even memory bandwidth are normalized.
The problem is that fanboys want to condense the performance of a GPU architecture down to a single number, which is not possible considering the complexity of modern a GPUs and rendering techniques. Its easier to get involved in "my number is bigger than yours" dick swinging contest for both fanboys and even company PR and marketing.
@@vforvendetta6193 you're not a clown, you're the entire circus
@@vforvendetta6193 read this comment. Somehow misinterprets the entire comment. Runs it away with a "you're an AMD fantard." 🙃
Videos like this is why I love this channel ! Amazing work ! New consoles will be sweet for sure, esspecially cause of much better cpu. I would love a comparison like this but on cpu side, jaguar vs ryzen.
Teraflops is a modern version of "Bit" in consoles until 64/128 bit era
FLOPS are a pretty useful metric - if you're comparing like for like. But it doesn't take efficiency into account. If you're comparing different architectures, one chip might simply need more operations to get to the same result, or, as has happened with GCN and it's four way parallelism, developers simply have a very hard time keeping the chip fed.
16 bit is doubly better than 8 bit. BLAST PROCESSING!
Give me ONE reason why you should buy a measly Sony 32 bit machine instead of the TWICE as powerful N64. Bet you can't heh... Keep your wipeouts and tomb raiders, once Mario 64 and Ocarina of Time are out you'll see true next gen gaming.
@@wsippel yup, you're right but I meant in terms of marketing material
@@captaincarrot8928 Back then you needed a Jaguar with Tempest 2000 'cause it was 64-bit -- forget about that 3DO with its measly 32-bits! 🆙😀
The fact that a 7970 OC’d to >1GHz can still get over 30fps in really demanding situations is pretty insane. Talk about a good buy (especially the 6GB Sapphire variants) from 2011-2012.
It’ll be interesting to see if RDNA pulls off something similar; AMD tends to be pretty good at longer-term future-proofing. Thanks for another fascinating video, DF!
Incredible video.
I truly appreciate everything digital foundry does in all of the testing.
Thank you for the hard work.
I really loved this video comparison and how you tried your best to normalize the results.
I don't think comparing the new Navi CPUs to the Nintendo Gamecube is entirely fair.
GCN yea lol
Lol, ATI was used in GC, sure AMD bought out ATI, bought the CG was hardly a RX 290 lol, and yes I know it was a joke, but too many people on here wom't know.
@@GTAVREALISM Gamecube was a joke? Sure it didn't sell as many as PS2 but it was a lot faster than PS2.
@@UltraCasualPenguin Lol, I think you read my comment too fast, I was talking about his comment, the GC was great, even more powerful than the PS2 in just about every way, I still have an old grey GC here with a few games.
I dont think you know what the hell you are talking about...Navi is a GPU, gamecube has nothing to do with this video, you must be joking...
This is amazing, the amount of work it must have taken to get all this together is admirable. Looking forward to more like these in the future
Finally the old dude
lmao
XD
😂
He has the best analysis
As someone still running a 280X (though I have been thinking of getting a V56) im happy to see my card in the ring. Some notes (some of which I think were explicitly mentioned in the video) and suggestions: any sane person running tahiti would prefer DX9 over DX10/11 or 12 any day. Even with the same in-game settings, you would usually see better performance using 9 (tests on DX9 wouldve been very interesting here). Some of the massive drops in performance of tahiti vs even GCN2+ as far as I remember were introduction of hardware acceleration for tesselation and texture compression (especially the former makes tahiti performance drop off a cliff). Still thanks for this video, it was fun to see how well the "fine wine" of ATI/AMD really aged.
Tests on DX9 would be pointless, since games don't support DX9 since like 2013.
I clicked thinking GCN stands for GameCube Nintendo
GCN 1 in Tahiti predates the 2013 consoles' GCN 2 which brought a 4x increase in ACE units for singificantly boosted async compute performance.
That's probably why you're seeing such a large difference in compute performance between Tahiti and Polaris.
All us console peeps want is 60fps... in all games standard
Thats never gonna happens, console always will be 30fps focus, with a waste of resources like Ray tracing is a fucking stress GPU power
@@noobdebusca1757 Always? seeing how some XBX games are already 60fps you are very wrong.
Yes, it has already began with this last (HALF) gen... alot of games on X have the option of 30 or 60. I'd take 60 all day long
Then lower your resolution.
@@gordongoodman8342 lol if we could we would XD
I actually read the transcript online and since I've watched so many of these videos I read it in your style! This has been one of my favorite channels on youtube, keep up the good work !
Next, Fermi VS Turing?
I love my GTS 450!
One was built to fry an egg, the other wasn't
Fermi 1
Turing 0
Yes
@@CaveyMoth what
@@parsastrife6629 It's cute
Seeing what top developers are able to coax out of low-end 2013 silicon... I'm dizzy with these prospects.
It's more like... these developers aren't properly utilizing the current gen PC hardware.
To be fair, HD 7850 class hardware in 2013 was modern Mid-range hardware, not low end.
290/290X and 780/780 Ti was enthusiast class
280/280X and 770 was high end class
7850/7870 and 760 was mid range class.
7750-7770-7790 and 750/750 Ti were low end class
This is so awesome. This needs to be done way more often. This has me curious in a direct comparison between GCN1-4 and even Terascale 3 vs GCN1
All this means is next gen is gonna be amazing. coupling RDNA with Zen 2 and a SSD is going to make next gen a power house!
And with the consoles that’s exactly what they did, you Reading the future
Wow, this is a comprehensive study. More of this, please!
Gamers Nexus and DF are maybe the two best tech analysis channels on UA-cam. Cool to here that you help each other out.
THIS is content. I learned so much from this video. Glad my thumbs up got you to 4.8k You deserve much more though.
Super cool summary of AMD's past and current GPU architectures. Very in depth with nice testing methodology and game selection.
I would love to see you guys tackle the topic of drivers for nVidia, AMD and Intel. The whole process of find a driver download, installing them, getting to the settings for the driver suit software etc. How well they support Vulkan, DX11, DX12, OpenGL (AMD is pretty bad here) and how is the image quality? Testing some games on OpenGL like Quake 3 Arena, Wolfenstein: New Order and newer stuff like Gears 5 on DX12.
I don't understand half of the things mentionned here but at this point I just enjoy whatever Digital Foundry does. Simply one of the best channels on UA-cam.
As much as you might want to consider some of the GCN 1 results bogus because of potential bottlenecks elsewhere causing an 'invalid' comparison of Tahiti's FLOPs and Navi's FLOPs, I think the comparisons are even more valid given those wide differences in performance.
It's simple really. In situations like the Novigrad test in The Witcher 3, GCN 1 gets decimated by newer revisions of GCN, let alone RDNA. To me, that's a valid test of architectural FLOPs. That Tahiti chip is still very much capable of a theoretical 4.1 TF of compute, but the architecture design is imbalanced in a way that games like The Witcher 3 see very low performance scaling with either clock speed or memory bandwidth, as determined by your 1440p performance being basically identical. The Radeon geometry pipeline has improved dramatically over the years, and newer architectures clearly get more out of their TFLOP rating when faced with geometry and tessellation heavy workloads.
It would have been cool to see both the R9 380X and R9 390 included in this comparison. The R9 380X uses the Tonga chip based on GCN 3 with 2048 shaders, while the R9 390 uses the Hawaii chip based on GCN 2 with 2560 shaders. In the same way you projected the results of GCN 1 vs GCN 4 onto RDNA, you could have potentially then projected the same results against GCN 2 and GCN 3, but it would have been a **lot** of extra work for a purely academic result.
I had a dog named Terra. Never occurred to me to measure how floppy she was. She was incredibly floppy though. 😁
GCN: Compute architecture with Graphics tacked on
7970/280x is the best aging GPU of our generation change my mind
I'd really love to see a companion video to this with comparisons between a few generations of Nvidia GPU architectures! Thanks for another wonderful video, DF! :)
Could've done 5700XT vs. 390 for similar unit counts. Here the Navi GPU has double the ROPs.
A 390 isn't gcn 1.0, neither is a 280 btw, I don't think they know that despite the internet fury the 280 is not the exact same as the hd 7000 gcn 1.0's.
@@Folsomdsf2 What changed in 280 compared to 7000?
As for 390, I don't care if it's GCN1.0 or not, it's far better perf/TFLOPS comparison and gives a much better picture of RDNA improvement over GCN.
@@SolarianStrike Negative, it's gnc 1.0v3 You can actually look at the actual differences.
Really looking forward to seeing what these new consoles can deliver, the price though is going to be £499 minimum probably higher but when you compare that to high end smartphones costing around £1000 I don’t think we can complain too much.
You can't just look at the performance and technology leaps in GPU. You have to also consider the performance gains of going from a 8 core mobile APU to a full fledged current gen, multi-threaded 8 core 16 thread CPU. And New speed of memory. It really is and amazing leap.
Finally!!!!! DF's voice of maturity and reasoning. Glad to have you back Rich.
I really love ur content.
I was calculating the Teraflop difference in Architecture between Nvidia´s GPUS vs AMD GCN and AMD RDNA to get an Idea what the next gen console will deliver with an expected amount of 10,8 Tflops.
That was the Perfect video for me thanks:D.
Richard, you are quite skilled. Shout-out to Gamers Nexus.
I like your all things being equal approach. The conversation in the CPU market seems to be 5 GHz is a rough upper limit on clock speed, and more cores / smaller dye is the right way to go. I'd really like an education on what a GPU actually is today, because x86 / x64 I understand. Thumbs up!
Current owner of an RTX 2080, and i have to say judging from what Richard said at the end, it is true! consoles simply give amazing results for their price and hardware and the games i show to everyone new to gaming to impress them are God of War and Gears 5, very amazing and visually stunning games, and each running on my pro and my x
Facinating stuff, this must have been a ton of work rich :) it feels like the next gen consoles will use some kind of Navi refresh but this this is a very good hint
The next-gen consoles will be hotrods, with custom Zen/Navi providing gigantic boosts to clocks and efficiency (CPU) and huge bandwidth gains (GPU) combined with hybrid SSD drive setups -- holy shit, can't wait.
Equally excited about innovations on PC. What's nice is that multiplatform games will no longer be so held back by the lead SKUs (having to run correctly on base PS4/XB1) so the baseline spec is going to bump up.
Finally, we'll escape from the weak 2012-era laptop grade CPU architecture (bye bye, Jaguar)! The baseline for CPU-dependent draw calls, world simulation, physics, logic, AI, etc. will progress! And that will lead to new innovation in *gameplay* - new interaction, new world sim, exciting AI, etc.!
In other news, 2019 LG OLEDS will get a firmware update making them G-Sync Compatible.
Todaywasagoodday.jpg
So, in other words, VESA Adaptive Sync Compliant/AMD FreeSync V1.0 Compliant
Edit: So overall, very good news indeed, I'm surprised such a feature wasn't released when the TV's launched
Whaaaaatttt???? That's crazy. Unfortunately my OLED is 2 years old.
Conclusion is : Teraflop will not be a buzz word for next gen console wars. Because measuring teraflop number is not related to power, performance and technological advantage. Next gen will be how good the exclusives and quality of the developer is.
If we could harness the power of Wichards hand gestures, Iam sure we could achieve a breakthrough in discovering a raw power when it comes to hand gestures. Basically wichard can power an entire personal computer with the power of his hand gestures.
Okay so this might be a personal preference sort of thing and not applicable to everyone but I'll say it anyways. I love what you all do with the real-time FPS charts along the bottom and the averages along the top-right, but something I'd really like to see implemented is when you're touching on the conclusions, such as the improvement between Navi VS Tahiti at 1080p/1440p, could you show us those as charts as well instead of just text? Seeing a bar chart or similar of these figures is *far* more digestible to me than just the text alone, and sometimes I'll find myself going back in a video to grok something you've said that I otherwise could have kept up with if the information was presented a little more "legibly". Again, maybe a highly personal preference, but something I'd ask you to consider all the same.
So this is basically a video showing off the equivalent of "GPU IPC?". Not sure what a proper term this would be but I really like these types of videos.
FLOPS measurement is basically a more modern version of MHz myth
While MHzs can not be directly related to performance out of context, FLOPS can. This metric shows how many floating point operations per second can be possibly carried out. So you can't call it a "myth". But it can be indicative only of computing workload performance, not games.
@@BlackShadowLegion yeah at least it provides how much floating points it can calculate... but in games context... its basically the same thing as mhz myth
Lol you can literally overclock your gpu and cpu and get fps gains wtf are you talking about mhz dont matter? Yet intel is still leaps ahead of amd in benchmarks simply because of frequency
@@malygos9301 you should google mhz myth... its a real term used in early 1990s
@@BlackShadowLegion Tflops can only be directly related to performance when comparing two or GPU that use the SAME architecture. And even then it is not accurate and the only way to truly compare them is with real-world and synthetic benchmarks.
1989-2001: Bit Wars
2016-Present: Teraflop Wars
2002-2015 was a era of peace and prosperity?
@@wing0zero Obviously. I don't remember any buzzwords from that era until the Pro came out.
@@WH250398
Oooo let's have...
2002-2015: The Core Wars
Who had the most CUDA/GCN cores, or...
2002-2015: The Transistor Wars
I remember transistors been a bragging point at one time.
Genius idea, mind blown. I wonder if the in house team at AMD even know exact answers to the questions Richard is going for or have done these kind of tests?
this is crazy because i was asking myself this same question and then I find this video you dropped on the exact same day one year ago lol
What I think is that the games aren't just optimized for Tahiti, specifically.
CCN have this problem, right? It's a "monster", but it requires very well made code to prevent stalls inside the pipeline. Isn't that's why it always failed to deliver the promised TFlops? One of the biggest changes in Navi is for that reason, improve efficiency and keep the pipelines going when executing games.
This deep dive type of video is what got me first hooked with DF. Awesome!
I've been looking forward to this one Richard. Love your hardware Vids.
RaNDY!!!! Love u in southpark bro ;)
Indeed I did. Nice to see some R9 Tahiti action. I still own an R9 270X. Recommended resolution 900p.
We need MORE competition in PC/window space. Microsoft has been almost dictating what the standard is for pc space for last what 10+ years with windows/directx, and IMO it’s become bloated and MS has gotten lazy because of very little competition, until recently with vulkan api. But when comes os, windows is still the standard. But windows/ms does not see vulkan as threat, or Linux (yet) or any games made under different os.
Very interesting to see and it’s shows that RDNA 1.0 is promising. Hopefully RDNA (Navi 2?) will even be better and hopefully it help make console little better, I hope we see more 60 fps as standard on console. Yes I like both console and pc. Console because you DONT have deal with windows Bullshit lol
Also note. Game engine also make huge different and some game engine are a lot better optimizes for some hardware, but some older hardware may not run as well. Some dev optimizes more or fav Nvidia, some would optimize for AMD, Hopefully RDNA will sound like it’s giving much better results, and once next gen console come out. I think we will see more games optimizes better for Ryzen amd Navi hardware. As said I was love see these test under Linux but that would be be almost “ impossible” task/challenge because of limited software/tools and even games. We also know most games under OpenGL are very horrible ports and don’t even match window version ( most, but not all) thus why I hope we see more devs use vulkan api.
Good video Rich and keep up good work :)
I've been a big critic of DF in the past but this analysis is simply phenomenal.
The analysis here is excellent as always
One possible confound is the 280X having basically zero delta color compression, hence the expensive 384 bits pipe was likely needed. So it could look worse than in reality - although the Battlefield Dx11 vs. Dx12 results may speak against that.
Very thorough testing. Excellent video.
Oh my god! Those videos are the best ones, EXCITED!
Old AMD drivers used more CPU than it should be. A process maybe I forgot what it was exactly but it was inherited and it couldnt be easily fixed. Great video.
the time and effort in this video is amazing, big like for your guys
Thanks, Rich. Brilliant stuff. I’ll be saying goodbye to my PC when PS5/Nextbox land. GPU/CPU power is hitting a plateau with the TVs/ displays. It’s going to be 4K for a few years now and the next gen consoles will deliver all the resolution/performance that even the enthusiast gamer will need for quite some while. In a quiet little box for $500 max. Yes please. When the world goes 8K/120 I’ll think again.
Tahiti, Polaris, Navi... Irrelevant in the eyes of old school development targeting higher resolutions at 30 FPS on console.
Umm...no? Sorry friend, definitely not irrelevant.
This was interesstttinggggg please do one for Nvidia!
This is why people shouldn't worry about PS5's 10.3 Tflop GPU. RDNA 1 was already a huge boost over GCN 1.0 which is what's inside the base PS4, but PS5 will be using RDNA 2 which will offer another boost in performance over RDNA 1. All said and done we should expect at least 7x the GPU power over base PS4 and over 3x the power of the Pro.
Very good job! Thinking about it last week. Thank you!
You know I love technology I really do. I love seeing new technology. but what I really hate is when I buy a video card for 500 + dollars two and a half years ago and now it's almost obsolete JS.
Just live in the 2000s computer days
Nice video, what you could have mentioned is that the consoles use GCN1.1 or 2nd gen GCN, not the first.
It might go some way to explaining the dx12 results.
I agree with most of this video, but with the BF1 I strongly disagree. For this whole gen, DICE's DX12 implementation has been disastrous, always being slower than DX11 regardless of architecture or OEM used.
true, BF1 dx12 is a mess.
Great job on this one. Now it will make sense when they launch the next gen
TFlops are about to become the next 'bit' 😜 . A simple number to help with marketing that becomes irrelevant after a while.
If gaming is your only concern, sure.
It's amazing that the Polaris architecture gains so much compared to Tahiti at the same clock，even higher than Navi compared to polaris.
Time Spy requires at least 6GB of VRAM to run well, so it's not a good test for the 280X.
Obviously, a teraflop isn't just a teraflop. Informative as this is, I feel like that's talking down to people. Unintended, of course.
Yeah its crazy how something like Uncharted 4, GoW or Horizon can run on custom 7850. Cant wait for the next gen. Even if the consoles will have only something like RX 5700 GPU.
I beg you DF, please test Witcher 3 in Toussaint, at Palace Gardens - as its the most demanding place in game
I believe the reason why they always do the exact same test run through Novigrad is because it allows them to dig up test results from years back.
I got my cycling and tech mixed up and expected a Global Cycling Network video 😂
Gj Richard! Very interesting dive into this topic.
One thing to especially keep in mind regarding memory underclocking is that this somewhat breaks the comparison to a certain extent.
Due to how closely GPU operations are tied to memory performance and latencies, underclocking memory has far reaching consequences on the peak performance of the core itself, in simple terms a bottleneck. Certain compute and cache operations are closely tied to memory operations. So underclocking and overclocking inherently accrues penalties/gains for the architecture.
Overall what this means is that we can't compare the GPU's true architectural performance ceiling against the benchmark pair card we are testing against we can only asymptotically approach the truth.
I didn't specifically address core clocks and the various operation & issuing latencies, pipeline stalls, cycles per ops, cache speeds, queuing and thread scheduling etc. But overall clock for clock comparisons are very foggy endeavours that reveal little useful actionable information. One could do the same with a matched power envelope performance/watt comparison and it still wouldn't provide very much useful information beyond being an exercise in academic curiosity that merely demonstrates that yes indeed the new architecture is better at the same clock speed/power envelope. Sometimes we are even ignoring that the higher clock speeds and lower power profiles of the new chip may often actually be a result of the newer architectures design itself. So you have a bizarre chicken and egg comparison problem.
In short focusing merely on clock for clock comparisons misses the forest for the trees, metaphorically speaking.
The thing is, everytime we only count as valid those Tflops and bandwdth numbers , without even counting that most games are still TMUs and ROPs limited.
See those generational gains? All directly related to console hardware implementation. Go look at the specs of the new xbox project Scarlett and new PlayStation. RDNA will be the powerhouse for the next gen of consoles.
That Navi 5700 @ 1GHZ was prolly using like 80w
I have a 1.1 ghz at 800mv setting on my own 5700xt and the power figure reads at 50w-80w depending om usage in game then. AMD advertised same speed at 50%, power usage 14nm vs 7nm and i can confirm thats not far from the truth.
if you think at full clock is 125watt and undervolted about 100watt
@@razoo911 less than that. At 1ghz to 1.1 ghz read my comment. Scales very well at 1.1-1.6 ghz at stock my 5700xt uses 150w-180w
very illuminating. thank you for the work you put in.
@7:37 your percentage gains are off a bit. i.e. you state Navi vs Tahiti is a 62% gain in performance (you typed +62% 1178 pts vs 655 pts). This is actually incorrect. 655 > 1178 is a 79.8% increase, or going from Navi down to Tahiti, is a 44.4% decrease.
Seems there may be an elephant in the room that nobody is addressing.
For years reviewers have said AMD cards (I.E.: Polaris, Vega) are designed with more of a work station performance and Nvidia (I.E.: Pascal) was more about gaming performance. Yes, even crypto miners loved the AMD cards because of this. But this was always pointed to when we discussed gaming performance numbers. "The games are optimized for Nvidia because they hold the market share".
Navi launches and reviewers say, "AMD have switched and the GPU will concentrate more on gaming aspects with fewer workstation assets". Well we know games are optimized for Nvidia and guess what. Turing is a workstation card. That is what it first launched as and that is what it is. First reports were that Turing will not come to gamers. So if games continue to be optimized for Nvidia first. Then it would seem that when AMD jumped over the fence and broke away from GCN. Nvidia jumped to the side AMD had been on. So as Nvidia influences new games to take advantage of that workstation based Turing the RDNA performance gains will dwindle and so will Pascal. People may want to hold onto those Polaris / Vega cards. Just put it on a shelf because you may need it within 2 years.
awesome analysis richard ! one of the best videos on DF so far, in my opinion. Would it be possible to make a similar analsys with nvidia and even intel graphics ?
That's the type of video that makes me be a subscriber and have the notifications enabled
This was really fascinating. Thanks, Rich!