AMD Reveals the Radeon RX 6000 Series, Coming November 18th
anandtech.com
anandtech.com
- Driver stability. Nvidia is not perfect but the 5700 series was awful from AMD (had to return mine). They need to at least match Nvidia.
- DLSS. It started out as a gimmick but with DLSS 2.0 it's just 70%+ free performance improvement with the only downside being somewhat limited game support.
- Video Encoder. Nvidia's encoder has very high quality and is well supported by streaming and recording software. I wonder what sort of improvements the RX 6000 series has?
- CUDA, Tensor Cores, Ansel, etc. I don't really use these things but if I'm paying the same amount I want similar capabilities.
It's kinda crazy that they're using the same silicon in all three cards; the 6800 has 1/4 of the CUs disabled! It is a big chip for 7nm, though. AMD has said that availability won't be an issue (which could give them a leg up on Nvidia in the near term), but I do have to wonder about it.
The 80CU 6900XT was rumored months ago and there was a lot of speculation and incredulity about the memory system to feed all those cores. GDDR6X is Nvidia exclusive so some said a 512-bit bus, or HBM2, or even a combination of HBM and GDDR6. The big cache is an interesting solution, I'm curious how it'll perform. It makes sense in a gaming context where a lot of buffers and textures get reused over the course of a frame. I'm a bit worried about tail latency affecting the 99-percentile frame times which have a huge impact on the subjective experience of smoothness. Also bad for massively parallel compute work.
IMO, VRS (variable rate shading), which is in all major GPU providers (NVidia, AMD, and even Intel iGPUs / Xe), provides the "upscaling-ish" performance gains that we want for the next generation.
Its much harder to see the difference between native and VRS.
https://devblogs.microsoft.com/directx/variable-rate-shading...
-----
Its not a "competitive" feature, because everyone is doing it. But it basically accomplishes the same thing: carefully upscaling results (2x2 instead of 1x1) in locations where the gamer probably won't notice.
DLSS is also a rather heavy computation, taking 1-2 ms per frame on high end GPUs. That's a serious amount of processing power to put into a display and doing so makes it exclusive for that use. On the GPU those tensor cores can be used for other tasks when they're not doing DLSS.
Because that's where the difference gets important. If I compare to my PS4 Pro, many games "look" better because they drive the UI at full native resolution while using checkerboard rendering to keep the framerate up. The same game on my 970GTX will chug, drop frames or have all the text and UI look blurry because I need to drop the resolution.
If DLSS 2.0 fixes this issue, it's a massive improvement.
DLSS 2.0 is a great addition to that field, but it's far from being the only good option here. AMD has also done some things in this area, too, like Radeon Boost which was the idea of dynamically reducing resolution in response to player inputs, with the idea being if you're sitting still-ish you're OK sacrificing FPS for better quality, but as soon as you start whipping about you want the FPS more. Game engines also do similar things, and combined with temporal upscaling & antialiasing are quite compelling.
Care to explain more or have some good resources where I can learn about it by myself?
CUDA might be a bit tough. It's hard for an equivalent to catch on if everyone is doing CUDA. It's the chicken-and-egg problem. There's ROCm* if that matters.
Software devs might consider it as a factor because they might want to play around with it, but for everyone else is a irrelevant feature.
Any chance AMD would hire some engineers to upstream rocm support into tensorflow/pytorch?
Raw performance looks fine, its all the lack of news on the features and zero acknowledgement that people play games other than the latest AAA games that get benchmarked and they need to run well too. Not to mention all the bugs, if I buy an AMD card I can guarantee based on history something weird will happen for the life of the card, it always has for the past 20+ years, I haven't yet had a clean experience with an ATI/AMD card and I have owned a bunch of them.
What I really need AMD to say and do is acknowledge how important a good customer experience is with their driver stack, how every game and API matter, that they understand that customer experience is reality and are now committing it to history for good. Then they can sell me a high priced card, until they do well its got to be a lot cheaper to be worth the risk.
Try that under Linux sometime. It's a different song there thanks to Mesa.
>opengl
Is not the most popular API, and AMD had, until recently, limited economic resources, which they mostly directed towards their CPU team and towards hardware. For their proprietary windows efforts, they needed laser focus on what's most important: The newest and most popular APIs.
There's money these days, and they multiplied the size of their driver team, but money spent on OpenGL would be indeed wasted, as there's a pretty compelling solution for this API from the Mesa guys, in the form of an opengl implementation that's on top of Vulkan: Zink.
I expect future GPUs and drivers (for all platforms) to absolutely not bother with opengl and simply ship Zink instead.
https://devblogs.microsoft.com/directx/in-the-works-opencl-a...
The old ICD interfaces that OpenGL drivers rely on, might not be around forever, and so far Windows doesn't support them in sandboxed processes.
I'm really not finding anything that supports this claim, can you substantiate it? I see scattered sub-60fps reports in Minecraft for both Nvidia & AMD. It seems to just be a "minecraft's engine is shit" thing, not an AMD thing?
But I'm not seeing any actual comparisons or head-to-heads or anything showing AMD in particular is struggling with Minecraft in particular. This unknown quality benchmark is the only thing I'm finding for Minecraft: https://www.gpucheck.com/game-gpu/minecraft/amd-radeon-rx-57... and it's definitely not struggling to hit 500 fps numbers at 1080p, either.
And other OpenGL games & apps are not showing any severe AMD issues, although the only significant sample of those are on Linux so that also has a different OS, so there certainly doesn't appear to be anything systemically wrong with AMD's OpenGL as you're claiming. Unless for some reason they just never brought over the way better Linux OpenGL driver to Windows, which would be highly unexpected & definitely something that needs evidence to support.
> zero acknowledgement that people play games other than the latest AAA games
This is just completely false and seems to just be because your pet peeve is Minecraft in particular? Both Nvidia & AMD discuss popular games not just the latest AAA games. Usually in their latency section, since that's where it comes up as more relevant (esports games), but it's definitely not an ignored category entirely.
When you're using AMD or Nvidia's proprietary drivers, which are way faster, then you're not using anything interesting from Mesa.
EDIT: Also Mesa's open source Radeon driver came from AMD.
Haha, no.
(Which is not the case for RADV, the Vulkan implementation)
The driver bugs are what I'm most worried about though. I remember when I got a 4670 in 2008 there was a cursor corruption bug (google it, it's common) and if you didn't shake the mouse enough for it to fix itself it'd crash. Then I built a new PC with a 5770 and I still had it. Later a new PC with a 6950 and still had it. Then 2x280X and still had it. It even occurred on Linux once. Then I went Nvidia for a few generations and it seems AMD finally fixed it. It seems fixed now (maybe because of new hardware) based on internet searches but from 2008-2015 the bug was there. And this is just one example of things.
Not that Nvidia drivers have ever been perfect but I never ran into this kind of shit so often for so long on their cards. Hopefully I don't end up regretting the 6900 XT because of it.
This cursor corruption bug was ridiculous. I ended up not playing some games (I think Dota) because I would get this bug. I went with an Nvidia card as my next one because of that as well.
It's interesting how something that seems relatively minor could have such an impact on purchasing decisions.
You can see problems today like the mouse cursor not respecting Windows 10's night-shift color filter for the same reason.
This pointer not respecting night-shift hits the trifecta hat trick. There's a registry fix, of course, but the first time I saw it, I thought "huh, that's weird". How did this get out without somebody on the release train thinking to themselves the exact same thing?
But my simple 3400G based rig using the integrated Vega11 plays stock minecraft just fine @ 1440p and 20 chunk viewing distance with everything else set to max.
Watch Dogs Legion comes out tomorrow and I've been benchmarking it today, as have many others in the subreddit. DLSS is sometimes improving things but is also quite often leading to worse performance, especially if RTX is set to ultra. I have no idea if the issue is game specific or not but I'd be curious to know which games use it well.
In addition, my CPU seemed to be overheating and was likely throttled. I tracked the issue down to my new RTX 3090 which was dumping heat inside the case in a noticeable way.
I re-applied thermal paste to my CPU and replaced my watercooler fans (due to rattling) and I’ve improved fps by 15%.
The latest nvidia driver also increased fps by 12%.
Use case: iRacing on triple screens, it offloads the CPU, in a CPU limited title.
Meanwhile I'm loving my 3900XT Ryzen CPU.
Seriously. I have the RX 5700 XT and at least once a week the driver crashes with the same error on `dmesg`, with the only solution being to reboot the machine because the 5700 XT for some ungodly reason doesn't support software reset. I love the card and I feel like I got what I paid for in terms of performance, but the driver instability is absurd.
Ironically, the exact opposite is the reason I use AMD on Linux. It "just works" without installing anything and everything is super stable, while Nvidia is a huge proprietary mess.
That, and Proton in Steam have made it possible to play some old favorite games as well as new titles with impressive performance.
I will say that real-time ray tracing has me at least considering one of those new cards...
this is also one of the mechanisms that Wayland is built on top of.
Edit: Here’s the video: https://youtube.com/watch?v=TY4s35uULg4
Something is very definitely wrong.
It is not new and untested, it's been in use since 2018.
But on a serious note - why are they doing so? Surely, they have good predictions of supply and demand - why are EBay resellers reaping the profits from the gap between the two? Why aren't these cards retailing for $1000, with price reductions happening as demand drops below supply?
And I do not feel sorry for them.
Personally, I think its less due to these and more due to people using bots to buy up stock for scalping. The same stocking issues have happened for the new Xbox and Playstation, and even for non-tech hardware, like the newest Warhammer 40k box set. The bot scalping issue is just becoming more pervasive.
Back then these highend-GPUs used to be prestige projects that mostly existed for the marketing of "We have the fastest GPU!", the real money on consumer markets was made with the bulk of the volume in the mid-range.
3090 is priced at $1500 for its 24GB RAM which enables ML & Rendering use cases (Nvidia is segmenting the market by RAM capacity).
AMD's 6900XT has the same 16GB of RAM as 6800XT, with less support on the ML side. Their target market is gamers enthusiasts who wanted the absolute fastest GPU.
RTX Titan for example was not as good in FP64 calculation compared to RTX Quadro. While first Titan was good at both. Nvidia kept changing what Titan means in order to extract most amount of money without hurting Quadro sales.
Even at its peak certain things were software limited for Titan, but not for Quadro. Nvidia has tons of artificial limitations for each of their graphics card once you go beyond gaming.
I also expect been a kind of struggle between their consumer side who need some other reason to sell their top end GPUs beyond just the whales when game developers aren't really that interested in pushing out features that will require it given the niche ownership, while the datacenter people want to protect their Quadro margins. This to me seems their most likely reason for the on-again/off-again relationship with pushing their high end cards for compute usage vs gaming usage.
I'm saying the only reason Nvidia didn't call it a Titan was because they knew it wasn't going to have an unassailable advantage over AMD.
A Nvidia card is Titan if it has the name and the drivers, and not if it doesn't.
AdoredTV made an entire video[0] to demonstrate this is absolutely baseless conjecture that NVIDIA Marketing tricked people into inferring.
It absolutely destroys the Titan idea, leaving no room for doubt.
curious what is that?
An official firmware release could change this but they're likely saving it for the Titan cards down the road.
The Reddit AMA implies the Ampere architecture supports this configuration but is software limited under the FP32 section: https://www.reddit.com/r/nvidia/comments/ilhao8/nvidia_rtx_3...
To epistemologically correct, I don't believe one will ever see a statement from NVIDIA confirming that this is a software limitation. I believe it's just inferred.
edit: This is close to confirmation: https://imgur.com/a/RH8vyz9 -though I may suspect there are non reversible hardware fuses at play too.
I mean, I really hope that it's just in the driver so that an enterprising reverse engineer can hack the driver and re-enable full FP16/FP32 accumulate :)
TITAN X got access to pro drivers shortly after AMD announced the Vega Frontier Edition.
So maybe not 100% designed for ML, but better than anything else out there unless you want to sell a kidney.
Here’s a dirty secret: there is a ton of ML prototyping done on GeForce level cards, and not just by enthusiasts or at scrappy startups. You’ll find GeForce level cards used for ML development in workstations at Fortune 50 companies. NVIDIA would love everyone to be using A100s to do their ML work (and V100s before that,) but the market isn’t in sync with that wish. The 2080TI remains an incredibly popular card for ML even with only 11GB. Upping to 24GB, even with the artificial performance limitations for certain use-cases, enables new development opportunities and use-cases to explore.
When it comes to product stratification, the hard rule according to driver EULAs is that GeForce cards can’t be used in data centers. For serious ML development at scale, NVIDIA has their DGX lineup. In the middle are the Quadro cards, but they tend to be a poor value for ML. The cost differential with Quadro is largely due to optimizations and driver certification for use with tools like Catia or Creo (CAD/CAM use,) which don’t intersect with ML.
The Titan RTX may not have the gimped drivers, but the 3090 beats the Titan in many benchmarks nonetheless. Is the 3090 the best NVIDIA PCIe form factor card for ML? No. The A100 is still king of the crop and is the only Ampere card with HBM memory, and even the A6000 will outperform for many use-cases with 48GB of RAM. Still, the 3090 will be the optimal card for many.
I’m one of the lucky few to have a 3090 in my rig. I lead of team of volunteers doing critical AI prototyping and POC work in an industry give-back initiative, and price was not a leading factor in my decision to procure a 3090 over a Quadro. I chose the 3090 principally because I didn’t want a loud blower card in my computer (and I don’t need 48GB.) If someone donated an A100 to our efforts, I’d gladly take it, but it wouldn’t replace the 3090. It’s not a graphics card and it won’t play games, which indeed is an important value-added benefit of the 3090 :)
Aka, it's a more expensive 2080 Ti, not a cheaper RTX Titan.
AMD performance is on par or slightly higher. Power efficiency is much higher.
And costs $500 less.
A slower card that uses way more power and costs $500 more is really hard to sell, even with NVIDIA marketing team being as strong as it is. At those prices, few people are going to automatically buy the product without exploring their options.
Are there benchmarks showing this?
And they've obviously used the same preset for both their cards and the competitors'.
There's some more details in the press kit, but I do agree with the principle that decisions should be withheld until NDAs expire and third party benchmarks are available.
What's clear is that, with the information in hand, buying NVIDIA Ampere cards is simply not sensible. Waiting for RDNA2 reviews is.
Still it's more of a Zen 2 moment than a Zen 3 moment. From the lack of comments on RT performance compared to the competition (just that it was added to the hardware) it seems extremely unlikely the RT performance is at the same level. The cards also lack the dedicated inference hardware for features like DLSS or voice/video filtering. And the card still has less VRAM than the 3090. These are all minor but if you put them together it seems really unlikely we'll call the 6900 XT the absolute best performing GPU of the generationjJust like Zen 2 didn't topple Intel's claim of "best gaming CPU". We'll have to see 3rd party reviews and benchmarks to find out for sure though. What it does represent though is a huge upset in the 3070/3080 area where most cards are sold and a hint that there may be an Zen 3 moment coming for GPUs in the next generation where AMD really drives top tier performance to a new level after long stagnation instead of "just" coming close to taking the crown dead even.
Personally (and this part isn't going to be reflective of the average person) I was going to be a 3090 whale and I probably would still be if it weren't for Nvidia's shit stance on open drivers in Linux (one of my biggest gripes with my 2080 Ti). However with AMD being so close this round and me not having liked DLSS or RT on the 2080 Ti I'm willing to trade off for the 6900 XT. The $500 is a nice bonus but not really what's coming into play, like I said if I were trying to get perf/dollar the 6800 XT makes WAY more sense. This is similar to what happened with Zen 2, I was planning on getting the better Intel CPU for the couple extra FPS but I was fed up with meltdown type issues and Zen 2 was really damn close. Now I'm really excited for Zen 3 though :).
Which it doesn't. Thanks to AMD having some new, large "Infinity Cache" feature which they adapted from their Zen CPU architecture.
At the end, performance is what matters.
Edit: should clarify that I’d really love to get a quadro or one of their data center cards, which aren’t gimped in certain non-gaming workloads... but I’m not made of money :)
Would you still do it without hesitation if the money was coming of your own pocket?
Now on the other side of the world, where a 3090 is several times your rent, you really need to think thrice about buying one.
AMD GPUs have near zero support in major ML frameworks. There are some things coming out with Rocm and other niche things, but most people in ML already have enough work dealing with model and framework problems that using experimental AMD support is probably a no go.
Hell, if AMD had a card with 8GB ram more than nvidia, and for 500$ cheaper, I would still go with nvidia. Everyone wish AMD would step their game up w.r.t ML workloads but it's just not happening (yet), Nvidia has a complete monopoly there.
* It seems there was not a single commit in the past ~6 months which by itself is already a deal breaker.
* Documentation is lackluster
* You need to use Keras, I use PyTorch. This is not a deal breaker, but a significant annoyance.
* Major features are still lacking. E.g. No support for quantization (afaik), which for me is fundamental.
* Most importantly there seem to be no major community around it.
It feels a bit bad to say that, because clearly a lot of work went into this project, and some people have to start adopting it to drive the momentum, but from an egoistical point of view, I just don't have the courage to deal with all the mess that comes with introducing an experimental layer in my workflow. Especially in ML where stuff can still appear to "work" (as in, not crashing) despite major bugs in the underlying code, leading to days or week of lost work before realizing where the issue is.
Check out keras-helper.. I made it to switch between various backend implementations which are non NVidia specific.
Pytorch may eventually need porting, but for now I don't need it. I've been trying out Coriander and DeepCL now but I decided to stick to Keras, which seems to be a decent compromise. Not using 2.4.x though, do not need it.
OpenCL based backends are cutting it for me, running production workloads without needing to install CUDA/ROCm is the best way to go.
Spending a week/year working around ROCm would already cost you 5k$ plus the opportunity cost. For a whole team that’s a money sink.
The catch is that ML software stacks have had hundreds if not thousands of man-years of effort put into things like cuDNN, CUDA operator implementations, and Nvidia-specific system code (eg. for distributed training). Many formidable competitors like Google TPU have emerged, but Nvidia is currently holding onto its leadership position for now because the wide support and polish is just not there for any of the competitors yet.
The analogy is that everyone should get NVIDIA Ampere units (non consumer) units worth $30k because it's fast and you'd rather be spending less time in a lab with millions of dollars in funding. insert don't be poor T Shirt reference
PlaidML is not ROCm. Nobody needs ROCm, what people need is just linear algebra well implemented with OpenCL primitives. That's what PlaidML is. And it works quite well, even on those integrated Intel GPUs on most laptops.
Have you also looked at DirectML and WSL2? They seem to be running tensorflow quite well too. Those things may be the key to bringing these in adoption outside the well paid class of data scientists you came up with.
_If_ AMD has made a sufficiently powerful GPU, that will add a lot of incentive to ML frameworks to support it. But it's going to have to be a big difference, I imagine.
Given how active AMD is in open source work, I'm a little surprised they haven't been throwing developers at ML frameworks.
So AMD got a lot of heat for not supporting Navi (RDNA) with ROCm, but it seems that they are weeding out the things keeping people from running it ( https://github.com/ROCmSoftwarePlatform/pytorch/issues/718 and the links in that look like gfx10 is almost there for rocBlas and MIOpen). We'll see what ROCm 3.9 will bring and what the state of big navi is.
Why would I get a Radeon VII when used nvidia cards for machine learning are extremely cheap, and then I don't have to worry about experimental stuff breaking one day before deadline lol
That is incorrect - I have been running Tensorflow on RX5*0 cards for close to 2 years now. I even transitioned to TF2 with no problem. Granted, I have to be extra careful about kernel versions, and upgrading kernels is a delicate dance involving AMD driver modules and ROCm & rocm-tensorflow. My setup is certainly finicky, but to say AMD GPUs have near zero support is false.
gimme 48 GB gimme 128
edit: Also CUDA is just too important to switch to AMD.
This would make sense if these were Titan cards with ML drivers. Instead, they are not, only the regular drivers are available, and FP32/64 performance is artificially capped.
These aren't cards for ML.
The ML driver and FP32/64 performance capping aren't really issues since in reality we rarely hit those limits.
Are individuals buying graphics cards for ML? I would think that it makes more sense to provision scalable compute on a cloud-plaform on an on-demand basis than buy a graphics card with ML capabilities?
So yeah, the huge gap ML capabilities between AMD and Nvidia are a selling point, but probably for a small enough group that it doesn't make a difference.
Of course this calculation depends on how bursty your compute requirements are and how much you pay for electricity (datacenter cards are more power efficient)
I guess if you're rolling your own compute clusters, you've probably rolled your own storage solution too?
Using cloud services for work would mean using the exorbitantly expensive services from Azure, to say nothing of the painful and annoying experience that using Azure is at the best of times.
Instead, I can spend a moderate amount, build a machine that is more than adequate for work requirements will last ages, write it off on tax and still end up spending way less than I would have renting a cloud machine for a couple of months.
[1] https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/a... (Appendix A)
BTW - Whether it "could" be faster is indeed relevant because some of us are holding out for a Titan GPU next year with this unlocked. If you have unlimited budget or are under time constraints then by all means get the 3090, it is a beast. But if one has a 2080 TI then it's an important consideration.
If you can wait until next year you should always wait until next year, because there will (almost) always be something better than what is currently out. That's unrelated to whether or not the 3090 is good for doing ML research; it objectively is.
The 3090 will probably stay at a premium due to a few factors - I don't think ML performance plays into this at all though:
DLSS - There's a reason AMD cites "raster performance" since with DLSS enabled the 3090 has a major advantage.
AMD-specific optimizations - near the end of the presentation AMD disclosed that with all the optimizations on (including their proprietary cpu->gpu optimizations, only available on the latest gen cpus)- they could pass 3090's raster performance in some games.
I think for these two reasons, and the fact nvidia can't seem to get cards to vendors (and customers), there won't be a price drop on this SKU. They may however release a watered down version as a 3080ti and compete there.
That's also the target market for the RTX 3090, all the Nvidia marketing material describes it as a gaming card, it's Geforce branded and Ampere based Quadro's will be a thing.
There was/is also this whole thing: https://www.reddit.com/r/MachineLearning/comments/iz7lu2/d_r...
Also, would the TI versions actually help? Even a 100$ markup on a 3080TI would bring it close to the 6900 XT pricing. And the 3080TI cannot come close the 3090/6900 XT performance for that markup or otherwise it would risk cannibalizing Nvidia's own products.
Nvidia's only hope at this point is that either AMD fudged the benchmarks by a large margin or AMD gets hit with the same inventory issues.
(In the way they use their devs to help AAA games to fix some issues under some circumstances, there had been cases where optimizations speed up Nvidea but hindered AMD due to architectural differences but, surely it was all accidentally).
https://techreport.com/news/14707/ubisoft-comments-on-assass...
"Radeon HD .. gains of up to 20%.... Currently, only Radeon HD 3000-series GPUs are DX10.1-capable, and given AMD’s struggles of late, the positive news about DX10.1
Ubisoft’s announcement about a forthcoming patch for the game. The announcement included a rather cryptic explanation of why the DX10.1 code improved performance, but strangely, it also said Ubisoft would be stripping out DX10.1 in the upcoming patch
Ubisoft decided to nix DX10.1 support in response to pressure from Nvidia after the GPU maker sponsored Assassin’s Creed via its The Way It’s Meant To Be Played program."
https://techreport.com/review/21404/crysis-2-tessellation-to...
"Unnecessary geometric detail slows down all GPUs, of course, but it just so happens to have a much larger effect on DX11-capable AMD Radeons than it does on DX11-capable Nvidia GeForces. The Fermi architecture underlying all DX11-class GeForce GPUs dedicates more attention (and transistors) to achieving high geometry processing throughput than the competing Radeon GPU architectures."
GameWorks slowing down ATI/AMD users by up to 50% https://arstechnica.com/gaming/2015/05/amd-says-nvidias-game... https://blogs.nvidia.com/blog/2015/03/10/the-witcher-3/
Same with CUDA vs ROCm.
And getting ROCm set up is still a buggy experience with tons of fiddling in the deep inner workings of Linux so it is nearly impossible for the average machine learning engineer to use.
It is compelling for certain bespoke projects like Europe's shiny new supercomputer, but for the vast majority of machine learning, it is totally unusable. By now in ML world the word "gpu" is synonymous with "nvidia".
We actually had more issues with nvidia drivers messing up newcomers' machines during updates than with setting up AMD GPUs, but then again n is small (and AMD GPUs were for playing around rather than real work).
Still, a Titan Xp has CUDA support and plenty of memory, but it's better, IME, to upgrade to a model with less memory but higher cuda compute and access to tensor cores.
For the amount invested in the hardware development, the amount AMD have been investing in the software side has been shocking.
How much is this one really about TSMC 7nm vs Samsung 8nm?
Then again, I'm rooting for AMD all the way until they become the new evil.
Isn’t A14 a “5nm” process chip? Why would it be compared to intel and their 14nm++++?
I'd say the crowning achievement for this architecture is the "Infinity Cache": https://twitter.com/Underfox3/status/1313206699445059584
"This dynamic scheme boosts performance by 22% (up to 52%) and energy efficiency by 49% for the applications that exhibithigh data replication and cache sensitivity without degrading the performance of the other applications. This is achieved at a area overhead of 0.09mm²/core."
See also this presentation: https://www.youtube.com/watch?v=CGIhOnt7F6s
TSMC has to spit out Zen3, RX6000, plus the custom versions for XBX and PS5, all around the same time...
How do you even begin to think about testing hardware like that for correctness?
I remember my days as an intern at Intel. I remember someone using genetic algorithms to try to construct a test suite for some chip. But it was no where near that transistor count.
[a]: My knowledge of creating hardware is very limited. So I could be completely wrong.
Could you expand a bit on how that worked? Seems like an interesting application.
From my understanding: Binary decision diagrams on supercomputers. (https://en.wikipedia.org/wiki/Binary_decision_diagram)
Verification hardware is a major business for all these major CPU players today. Fortunately, solvers on this NP-complete problem have benefited not only from faster computers, but huge advancements in algorithmic improvements in the past 20 years.
And for an extra measure of automation there are test suites that are programmed into hardware that can drive chips at line speed on a test bench.
I'm not super familiar with it, but one day perhaps hardware verification becomes as easy as software testing someday?
Both are very different and very interesting on their own. For the design, verification can be done by computers and it's much more powerful than testing anyway. For the hardware, I imagine there are cores inside the chip with the sole task of testing sub-circuits (fab defects are much less varied than design flaws), but I stopped following that stuff at the era of single billion transistors.
Basically chain all the logic together in a special test mode.
Analog blocks get tests written specifically for them and either tested via external test hardware or linked internally.
For example, if you have a DAC and and ADC, provide a link internally and drive the ADC from the DAC to test both. You can also test comparators etc from using a combination of the DAC and ADC, trim bandgaps etc.
If you’re real smart, you do this at probe (wafer), massively in parallel.
They've got no ML drivers and capped FP32/64 performance.
Anyway, ignoring all that, IMO gaming is still more affordable than it's ever been - you can get a used Xbox One S for about $100 on eBay, or for PC gaming, you can get a used RX 470 or GTX 1060 for a bit less than that, which are fine GPUs for 1080p gaming performance. Also, even relatively modern AAA games can now be played at low-medium settings at 720 or 1080p30 without a dedicated GPU on a late-model Ryzen APU.
[0] https://www.theregister.com/Print/2007/11/12/review_nvidia_g...
[1] https://www.hl.co.uk/tools/calculators/inflation-calculator
If you look at gaming builds from the last few years it's always a 1080/2080. These cards are essentially the equivalent of the 8800GT back then, ie. the one with high performance that you got if you didn't literally have money to burn. But how people can afford this stuff unless they have no other hobbies I don't know. Maybe the answer is they can't afford it? I'm not a single penny in debt.
By every relative and absolute metric, you are getting way more for your money than you did a decade ago, even vs other technology products (eg CPUs have made significantly less performance gains over the same period [0]) or experience (you can play the latest titles at much better fidelity and frame rates than the 8800 GT could for games of that era).
Have high end card prices gone up? Sure, due to both to increased demand volatility (multiple waves of crypto booms and busts) as well several generations where Nvidia has simply not had viable competition in the highest performance categories, but even with that in mind, performance/$ has still kept climbing in just about every single category. Note: new generation low-mid range cards still haven't come out yet. Expect another bump in price/performance Q1/Q2 next year (GA106 & GA107 for Nvidia, Navi 22 and 23 for AMD, also potentially Intel's Xe DG2).
[0] https://www.cpubenchmark.net/compare/Intel-Core2-Extreme-Q68...
Since you added this after I started writing, I think your idea of what hardware most gamers have is just plain wrong. One only has to look at the Steam Hardware Survey results to see just how few people buy the top-end cards: https://wccftech.com/nvidia-geforce-amd-radeon-discrete-gpu-...
You can corroborate this w/ JPR's market reports which consistently show that mainstream and midrange cards (<$250) account for almost 80% of AIB sales: https://www.jonpeddie.com/press-releases/the-add-in-board-ma...
This also doesn't account, as you've seen from the used market, that most gamers on a budget upgrading can simply sell their previous card and effectively only pay a fraction of the price every time they are upgrading.
I see that I was mistaken, or perhaps misled, about what kind of card is actually required to be able to have a decent gaming experience today.
You've convinced me that the mid to upper-mid tier is not really that much different to how it was back then. I also should consider that while I was earning a pittance back then compared to what I earn now, I lived with my parents and didn't own a car or anything else really apart from my PC.
Now that consoles are basically gaming PCs, good gaming performance for AAA titles is increasingly anchored by each console generation, so there's a bit of a shift every 4-5 years - one happening now. There's also an insanely large backlog of games and a huge amount of F2P and competitive games that aim to be playable on iGPUs, which have also advanced considerably over the past few years.
Very likely prices pushed back up after Nvidia grabbed any market share they lost to ATI that caused them to give us a good deal.
The NAVI line was never officially supported by ROCm (though ROCm 3.7 in August seemed to get some compiles working for the 5700 XT, a year after its release).
-------
Generally speaking, ROCm support only comes to cards that have a "Machine Intelligence" analog. MI50 is similar to Radeon VII, so both cards were supported in ROCm. The Rx 580 was similar to MI8, and therefore both were supported.
The Rx 550 had no similar MI card, and had issues that are yet resolved today. Rx 570 had no MI card, but apparently its similar enough to the 580 that I'm not hearing many issues.
In effect: AMD only focuses their software development / support on Machine Intelligence cards. The cards that happen to line up to the MI-line happen to work... but the cards outside of the MI line are spotty and inconsistent.
Based on the ISA alone, I'd prefer compute-applications to move to RDNA frankly. Its clearly a better designed architecture.
------------
I can't find any public information on CDNA. I've even emailed ORNL for information, but they responded back saying that NDAs prevent them from saying anything.
Whether CDNA is based on NAVI or Vega will be a big question as Frontier launches next year. I hope its Navi based, I really do, because its just better. But I'd "understand" if ROCm can't be ported easily to Wave32 or other issues (RDNA is a huge architectural change).
Direct ML is still in its infancy and is quite slow right now, AMD really needs to step up their game if they want to compete.
FP32/64 performance is going to be capped, as with prior generations.
Do you mean "enterprise ML"? Like, if I ran an ML company I wouldn't be looking at these cards?
totally incompatible: https://www.itworldcanada.com/article/amd-zen-3-processors-w...
erm, we means totally compatible!!1: https://www.tomshardware.com/news/amd-reverses-course-will-e...
"chipset" nowadays is just a pcie device providing legacy ports (USB 2.0, additional USB 3, SATA, HDA audio controller, PCIe bridge, I2C, SPI, LPC, bunch of GPIOs).
Smart Memory could then be as efficient as cross-CPU EPYC memory access. Which is pretty good.
Makes me wonder if that's related to the improved L3 cache access of Zen 3?
Leaked 3DMark raytracing benchmarks showed Big Navi lagging far behind Ampere so I wonder how that's going to bear out in real games.
I.e. just because games on the consoles can support ray tracing now doesn't mean they actually support it well. I still think it's a generation or 2 away from running well (as a current 2080 Ti owner)
Nvidia did that by enabling DLSS and RTX in everything, that's how they ended up with "Up to 2 x 2080 performance" which in practice only seems to be the case with Mincecraft RTX running on 30 fps instead of 15 fps.
I mean, I get that you can’t enable these on AMD 6000 series because the card doesn’t support these well.
Not every game supports RTX, those that fully rely on it are still performing so badly that a 15 fps increase can be marketed as "double 2080 performance", while saying absolutely nothing about the much more relevant rasterization performance.
Using DLSS can be misleading as picture quality is very difficult to objectively compare. But common sense makes the idea of "More performance with more details!" just sounds too good to be true, even with ML magic involved. Reminds me much more about something like Temporal Filtering: Giving more performance, but a slight cost of picture quality that might not be too noticeable with the extra raw pixels trough upscaling.
I wouldn't be that worried if I were nvidia-- catering to the whales is good business, but I think we're looking at amd winning the lion share of the market
But we'll need to wait for real world testing to see how accurate these claims are.
However, what I'm seeing out of this mess is that AMD is absolutely competitive on a watt-per-performance basis now. The other problem is that AMD is so far behind nVidia in software (read: AI research mindshare) that it's not clear if we'll be able to see that many titles take on raytracing in future titles or adopt the work necessary to do ML-based upscaling with AMD as the baseline software stack rather than DLSS.
Who cares if your average indie game doesn't support it, it probably isn't useful in those cases since you're already running it at 900 FPS
> The point is that looking at DLSS supported games, majority are AAA titles, which is what DLSS is for.
No, that's my point, and it is in part due to lock out with Nvidia's explicit whitelisting system.
And in any case since Unreal and Unity are integrating DLSS and other "GameWorks RTX" features more and more games will be able to implement them with essentially just a toggle.
The graphical fidelity in indie games has increased dramatically mostly due to the fact that Unreal Engine became very indie friendly and Unity has really stepped up their game with what they offer out of the box.
Eye candy is now easy because the game engines have a lot of these effects and the materials required for them built-in and you also can quite easily get really high quality assets, particles and materials in the engine marketplace for very low cost.
5 years ago developing a water shader for physically accurate water rendering would probably take an indie dev months to do complete and probably could've gotten them a speaking spot at GDC, today it's a toggle switch in the UE4 level editor.
That integration (UE4) comes uses the watermarked dll. Unusable until you get explicitly whitelisted by Nvidia. Only a few tens of games have been whitelisted.
The current rumour mill hints at 3.0 when it goes GA it's also when it should be part of unity.
However you can still grab the DLSS branch that NVIDIA maintains and work on it, if the result is good enough I haven't seen any evidence stating that removing that watermark is particularly difficult.
Is it really, though? Consoles have been doing upscaling without it for years, and one has to assume they're still going to be innovating on that front on RDNA 2.0 with the new generation, too.
The DLSS 2.0 mode where it's used as a super-sampling replacement is kinda interesting, but realistically TXAA in most engines is also pretty solid. It seems like a fairly minor extra feature at the end of the day as a result... Cool, but not game changing at all.
EDIT: although AMD did make mention of something called "super resolution" as part of their fidelity fx suite which sounds like a DLSS competitor but there's no real details. And of course the actual image results here are far more important
Considering AMD's core role in next-gen consoles, it's likely that there'll be broad support for raytracing in games (especially cross-platform games). I'd say the question is more whether the 6000 series is anywhere close to RTX in performance for RT (which afaik wasn't shown today).
They've definitely been far behind Nvidia for a while, though. I haven't seriously considered an AMD GPU in almost a decade outside of budget gaming rigs, and even then I ended up going with a used Nvidia card. Hopefully this is enough to give them a serious foothold in the high-end GPU market so they can give Nvidia competition for years to come.
So Zen 2 would be when AMD "caught up." Zen 3 would be where they surpass (assuming reviews match the claims etc etc...)
Does AMD allow running in "rage mode" without voiding the warranty? Is that something that a 3rd party mfg will offer to cover?
I also assume that you are going to need a high-end case (notwithstanding the PSU) to provide adequate cooling because, if it is anything like Ryzen, it will react strongly to the cooling available to it.
It will likely be noisy and, as always with OC, not all cards will have the same headroom: it allows them to avoid promising headroom that may not exist on your individual card.
It doesn't. "Rage mode" would be equivalent to Ryzen's PBO. Which is definitely not on by default.
It's likely equivalent to just dragging the power limit slider to the highest it goes on MSI afterburner. Letting boost go out of power spec limits, but nothing more.
Which can still give decent gains (look at how power starved Ampere is, for example). But the only noteworthy thing here is just it's in the official control panel instead of a 3rd party app.
Speculation point: because of this I almost wonder if it's equivalent to selecting a higher TDP rather than a warranty violating OC like PBO was. One of the things people railed on Nvidia for was for ever slightly less perf you could get a much lower wattage card that was quieter, maybe this was the response to that?
The warranty violating part of all of these would be you're technically driving the power delivery system beyond what the spec strictly requires. For PBO that'd be your exceeding the spec that AM4 requires from motherboards, which could put you of spec for a motherboard's VRM solution. I don't know if there's a handshake there between the CPU & BIOS to ensure PBO doesn't exceed a given motherboard's specs, though. If there is it'd be hard to claim this would be an actual warranty-violating usage.
Rage mode almost certainly respects the vbios limits, so you'll likely see lower-margin cards have basically no improvement from Rage mode, depending on how much they skimped on the power delivery. You'll likely not see an actual push to void warranties as a result, too.
Although in both cases (PBO & Rage mode) it's going to be almost impossible for a vendor to actually reject a warranty issue unless you tell them you did this. Afaik nobody is doing something like a blown fuse to keep a permanent record of these feature(s) being used or enabled.
Not sure how in the world they plan on supplying both their new cpu and gpu series with a holiday season launch, though.
Who knows what they'll decide from a business perspective? I wonder how the margins compare between CPUs and GPUs? They could, say, plan on limiting their higher end GPU SKUs which gives them temporary bragging rights in the GPU space but reserves capacity for CPUs.
Of course, even if AMD's benchmarks are representative of how the card will actually perform and even with the lower price, I still have to consider Nvidia for their software features. Shadow Play, G-Sync, and RTX Voice are some nice features. Not to mention DLSS...
With the $1000 price tag on the RX 6900 XT, though, I think Nvidia would be crazy to not lower the price of the RTX 3090.
I current have a small-form-factor case with a 2080-Ti in it. As it stand, the biggest nVidia card I can put in would be the 3080. I'm curious if the 6900 XT is a two slot design, or if it's massive like the 8090 with the three slot requirement.
https://twitter.com/Radeon/status/1321590889396002816/photo/...
AIUI AMD has their own alternatives to all of these (minus RTX Voice, which I hadn't even heard about until your mention). I do not know how good they are, but I expect there to not be much difference.
As DLSS 2.0, they're supposedly releasing a competitor in next month's drivers.
FreeSync is an adaptive-sync solution, but it isn't a complete alternative to G-Sync. G-Sync displays require special hardware from Nvidia, must adhere to a series of certifications from Nvidia, and don't always support adaptive-sync at low framerates.
RTX Voice is probably not super important for a lot of people. However, as someone who prefers to not wear headphones for long periods of time, it sounds like an extremely useful technology.
And DLSS 2.0 is a pretty big deal for anyone with a 4K screen. I don't have a 4K monitor, but I regularly connect my computer to my 4K TV so I would probably leverage DLSS often. I hope AMD's upcoming solution is comparable, but I need to see it before I give them credit for it.
Very random but: Have you considered full-sized open-back headphones? They're at least an order of magnitude better experience, in sound and comfort.
I use/favor/recommend Sennheiser's HD600.
I use HD380 Pro at the office, too. They're comfortable for closed, but they still do get ears/head warm over time. This is in contrast to HD600. The sound is also like night and day difference.
>I've heard they leak audio and didn't want to annoy my coworkers.
At healthy (as in low and plenty sufficient) volumes this is not an issue. The reason I use closed headphones in the office is the other way around: I can hear my coworkers and they're louder than my music is.
>so maybe I should re-consider open-backed headphones if they breath better. Thanks for the suggestion!
You're welcome. I'm confident you can't go wrong with HD600, they're legendary all-rounder, neutral-tuned, uncolored headphones, and at their price they're a steal, but do watch/read some reviews before making the call, for your own sanity.
Do note that HD650/660S are less neutral and near-universally considered worse by reviewers, on both subjective and objective (measurements) metrics, even when more expensive. And the lower end HD599/etc models aren't even comparable on a level field, in neither sound nor build. (HD600 are modular and built like tanks. The parts are compatible with 580/650/660S and sold individually, for the worst case and unlikely scenario of breaking anything)
They're 300 to 600Ω impedance across the frequency range, so they do benefit greatly from a headphone amp, but will still sound great even from an anemic source.
If you end up feeling like buying a dac/amp combo to get that extra, perceivable improvement, do look into "Audio Science Review (ASR)" community for no-bullshit hard measurement driven reviews. There's excellent solutions that do not break the bank, and a lot of hocus-pocus that measures like shit and yet has outrageous asking prices.
Use tiny drivers with higher distortion and frequency response is v-shaped relative to HD600. They're better than e.g. HD579/HD599, but that's a low bar to meet.
>HD6xx
Are just a massdrop/drop brand for HD650, and thus everything regarding HD650 applies to them. All in all, the price difference of 58x/6xx/650/600 does not make it sensible to settle for a worse headphone than HD600. Particularly since it'll last you a lifetime. This is unlike, say, your computer's screen.
And the HD660S is, besides worse, dramatically more expensive.
Crying children, cats (e.g. Burmese outside the closed home office door when your meeting has started), lawnmowers outside, hammering away at the mechanical keyboard - it all just goes away like magic.
I'm in the market for a new high end GPU, but I truly need AMD to come up with their equivalent of RTX Voice. Already got my Ryzen. :-)
Impressive claim.
Considering these NVIDIA cards have capped FP32/64 performance and don't have ML drivers like Titan series do, the one sensible case for buying them is gone.
AIUI FP32 and FP64 are gimped.
"it has half <category> performance" doesn't immediately make it worth 4x the money on an uncapped card, especially if your task doesn't involve a capped feature.
It's the same story on the 3090, you get the 7th compute cluster enabled and they happened to double the memory. Nothing about that more than doubles the price of the card from the 3080, it's the same "flagship cost" driving price.
Even by the time Nvidia got on the scene with the GeForce 256 we were still talking £180 for the SDR model and about £220 for the DDR high end model. The market has got massive as have the silicon dies and the price per generation just keeps climbing.
Also, curious if their APUs next year will be Zen3+RDND2. They have tended to be a year behind when they integrate the two. But I suspect the extra interactions between teams will fix it this time around.
See: https://www.extremetech.com/wp-content/uploads/2020/03/AMD-I...
Worth noting Nvidia also got onboard this train with the 3xxx GPUs as well so it's not some console only proprietary feature.
I started thinking this way after seeing basically no improvement from the 2200G and 2400G to the 3400G.
As for DLSS 2.0, there's supposedly going to be a new scaler to compete with it on the new drivers next month. I don't know which technology will actually look or perform better, but I do not expect the difference to be dramatic either way.
DLSS could be an absolute godsend for VR, if you imagine the next gen of HMDs at ridiculous pixel densities but the game can cheat and basically fake your peripheral vision.
Generally speaking manufacturing proprietary features like that see poor support long term.
Physx, hairworks, SLI all come to mind.
That being said, AMD cards tend to age better than NVidia cards. My R9 290X went from being inferior to a 780Ti to 12-15% faster.
I wouldn't jump to conclusions. I'd instead wait until third party benchmarks.
AMD has the advantage of being in both Microsoft and Sony's new consoles, too, which in practice means games are going to be designed for AMD's RT.
We don't know for how long AMD has been developing their raytracing, or how much effort they've put into it.
What we know, however, is that the consoles use AMD tech this time around, and that's what games that run on both PCs and Consoles are going to be designed for.
This should clean up quickly though.
Would have been happy to see the RX6000 on a Mac.
Which means Mac Pro will get an ARM chip and an Apple GPU.
[1] https://www.amd.com/en/products/graphics/amd-radeon-rx-6800
These are gaming cards. They'll charge you more for the compute cards, when/if they're available.
https://www.evolution.ai/post/benchmarking-deep-learning-wor...
Overall if you need to have a solution that just works with any ML framework out there Radeon GPUs aren’t fit for purpose.
Edit: And some sibling comments have pointed out that it doesn't function on current gen cards.
Nvidia still firmly leads the software game in this regard. I'm really hopeful this will change in the future but I had the same criticisms four years ago when I bought my last AMD card and nothing has changed yet.
AMD had been rock solid for me with Linux ever since I realized that was the issue and bought an AMD card. I'll never go back to nvidia until they make Linux drivers a priority.
I still have the Spanish windows that I use for games with my very old nvidia graphics card. It's time to retire that system soon.
About January, they released a set of updates where all the planets aligned. Went from a grand adventure to a handful of commands. The rest of this year has continued to be a non-event for updates. I'll be looking at one of the ATI cards as lord knows I've failed to find any 3080s, of any make, in stock since launch.
https://www.reddit.com/r/CentOS/comments/eplicu/steam_on_cen...
And NVIDIA's blob driver doesn't work AT ALL on current kernels ATM.
Apparently, it's going to take months this time around.
G-SYNC works out of the box (unlike FreeSync) in Linux, and Nvidia even has a nifty control panel app.
But it is a hassle having to worry about having proprietary drivers in-situ during installation or else your G-SYNC monitor won't work or when your G-SYNC monitor doesn't work if you switch to an AMD GPU. ; p (Yes, the ones with the dedicated G-SYNC hardware, which aren't all G-SYNC monitors, brick when not plugged into an Nvidia card.)
This is intentional, by design. It's called vendor lock-in. NVIDIA abused their position at the time to plant g-sync screens everywhere.
Of course, later they had to adopt FreeSync like everybody else (TVs and Monitors across the industry, with HDMI and DisplayPort, Intel and AMD, on consoles and computers), but there's a significant base of sold screens with gsync modules that will tilt many GPU purchases to the NVIDIA side.
Even when they adopted FreeSync, they pressured screen vendors to use their own "gsync compatible" name for the same technology, by leveraging their market position.
It supports a lot of what the Windows version does, per application settings, frequency / temp / fan monitoring and curve adjustment, etc.
Historically, I did use nvidia when I wanted to do native linux on desktop/media center, because their closed source drivers were supported much better (even FreeBSD had hardware media decoding support). This was before nouveau.
I bought HD4850 on release and had no trouble on Windows.
Linux support wasn't there on release (unlike these days) but came quite fast, a matter of weeks IIRC, with open drivers. Support has been there on release for all my newer AMD cards. All cards still work, and are still well-supported by the open drivers.
In contrast, all my NVIDIA cards from before that ended up as dead hardware, and were a nightmare while it worked, with the moody NVIDIA blob drivers.
And on the topic, the early experiences with the 3xxx Ampere lineup have been horrible, with reliability issues, ridiculous PSU requirements, and high return rates.
This is despite few cards have actually been sold, with availability being extremely low.