AMD’s RDNA 2: Shooting for the Top
chipsandcheese.com
chipsandcheese.com
Last week I had occasion to dig deeply into the microarchitecture of the Apple M1 and M2 GPUs (which are very similar to the A series on mobile). I found the metal-benchmarks[1] repo to be very useful, and, relevant to this article, it contains a comparison to the RDNA generations. From that it appears that GPU microarchitectures are starting to converge, even if terminology is not. The standard configuration these days seems to be blocks of 128 threads (WGP's in AMD terminology, "cores" in Apple), organized into 4 SIMDs. Resources like workgroup memory (LDS in AMD lingo) are shared among each of these 128-blocks.
The linked repo also has links to other resources, including the incredible work being done by the Asahi Linux people to reverse engineer M1 and deliver working open source GPU drivers.
> To simplify, RDNA and GCN are the codenames for both ISA and Microarchitecture for AMD GPUs. https://www.gpumag.com/rdna-vs-navi-vs-gcn/
And WGP is "workgroup processor". https://en.wikipedia.org/wiki/RDNA_(microarchitecture)
With that much background it's no longer alphabet soup and I can judge whether to read it.
i.e., TLA is an acronym
> Note that the word acronym is also sometimes used to mean "initialism."
What do they mean by a "SIMD", is it a independent core with its own instruction pointer(s) etc?
Nice to read a rare explanation mof what the actual HW functionality that can help with ray tracing actually is:
> At the CU level [...] the texture unit can now perform ray intersection tests, at a rate of four box tests per cycle or one triangle test per cycle. Box tests take place at upper levels of the BVH, while triangle tests happen at the last level. A BVH, or bounded vertex hierarchy, speeds up raytracing with a divide-and-conquer methodology
I wonder what else BVH speedups could be used for in eg GPGPU applications. DB queries that have 3 dimensions eg timespan, attribute set, value span?
Has it's own instruction decoder and multiplexes upto 16 threads of executions (wavefronts) depending on how many registers each thread needs. The SIMD has it's own register files and execution units (ALUs/FPUs).
What these SIMDs share with the other 3 SIMDs in the workgroup processor are the caches and memory subsystem (which includes texture units and RT hardware)
CUs don't really exist anymore as of RDNA 1, replaced by the WGP.
In some marketing materials, AMD brand the workgroup processors as "dual CUs". But that's mostly so people can do per-CU preformance comparisons across generations. Because the number of lanes per SIMD doubled from 16 to 32. Both the WGP and CU have 4 "SIMDs"
But the SIMDs in GCN used to be less independent, it was only the vector ALUs and vector registers that were in the SIMDs. All instruction decoding, the branch prediction and scalar ALUs per CU. But to really confuse issues, each SIMD still had its own set of instruction pointers (10 of them for upto 10 hardware threads) and could branch independently.
The CU would do a round robin, executing an instruction one of the SIMDs each cycle, so that each SIMD executed an instruction once every 4 cycles.
I guess the SIMD name made more sense back in the GCN days were it really was just the vector units, vector registers and a set of IPs. But now RDNA moved the full instruction decoding and branch unit and scalar registers into the SIMD.
It's a CPU arch term. Single instruction, multiple data. So essentially, 1 instruction pointer that operates on >1 independent pieces of data each cycle. https://en.m.wikipedia.org/wiki/Single_instruction,_multiple...
Useful to increase performance at lower hardware cost when you're doing the same thing at the same time over a large number of items (i.e. graphics).
> I wonder what else BVH speedups could be used for in eg GPGPU applications. DB queries that have 3 dimensions eg timespan, attribute set, value span?
There was some cool DB acceleration stuff at GTC ~5 years ago(?) doing this with the then-current hardware accelerators.
Unfortunately, it gets into the specialization vs more general hardware tradeoff: it's usually cheaper to keep software the same and throw more hardware at a problem vs specializing software and hardware.
> Each WGP, or workgroup processor, features four SIMDs. [...] Each SIMD has 32-wide execution units for the most common operations, a 128 KB vector register file, and can track up to 16 wavefronts.
> What do they mean by a "SIMD", is it a independent core with its own instruction pointer(s) etc?
Basically I'm interested in how independent these are, can they each go off doing their own thing or not, and what resources are shared with the parent units (wgp and cu)
For me, it's easier to think about GPU arch as trying to realize a never-ending stream of data being continually processed, with all other components of a processor redefined to ensure The Stream Never Stops.
From that perspective, things like wavefront-attached instructions, data scheduling, execution masking, or bastardized conditional support make more sense.
Think of "a cycle" as a set of data first, rather than an instruction/state.
https://developer.amd.com/wp-content/resources/RDNA2_Shader_...
The ideal for a gamer is that every year the GPU maker pulls another rabbit out of their hat, and doubles the performance from the previous year, with the same manufacturing and 50% less power consumption. Anyone who can remotely afford it and prioritizes gaming could upgrade.
The ideal for the GPU maker is that every year, gamers double in number, and double in hunger for higher performance, and any incremental gain in performance can be sold for double the price of the previous year, all while R&D and manufacturing costs are dwarfed by sales volume.
There were years where it felt like the gamers were winning. You could buy a mid-range card for a couple hundred US dollars that outperformed the previous flagship. Amazing! The past couple of years shifted the other way. Between cryptocurrency mining and machine learning, the appetite for GPU hardware created a seemingly insatiable seller's market.
Fortunately, the mining craze has all but vanished, and GPU hardware is much more powerful than it was a year earlier. However, the GPU makers are still on a high from the wave they were riding.
If you're a gamer (and not a machine learning hardware consumer), you're not in a terrible place right now. There's lots of hardware that can run modern games in 1440p and 4K resolutions with lots of eye candy and plenty of performance for $500 available to buy right now. If you can't afford $500, you can buy a $200 card and easily get excellent 1080p performance. But things aren't advancing as fast as they had at some point in the past, and there is no new midrange ($250-300) card yet available from the current generation that can embarrass old flagships (that used to cost $500-600). So we can be a bit unhappy about that, but I don't think things are terrible. And I think with a little patience, we'll still see performance increase, and affordable options continue to improve.
Except that's the price of a full gaming console that can also get me 1440p/4k. It's ridiculous that I have to pay the same ammount of money for a GPU to get similar experience to my PS5.
Consoles were always cheaper than PC gaming, but midrange GPUs costing more than a console is a new phenomena and I doubt it's sustainable for gaming industry.
There's a mix - some PC gamers just prefer the experience enough to justify the cost. But most people have some other reason to own a PC anyway, so paying for the GPU makes sense for them instead of having a separate console. And then there's the ability to build, customize and upgrade your system, compared to going out and buying a new console entirely. There's also, in general, much more backwards compatibility in PC gaming, though it's not without flaws.
If you isolate it to "playing games on a screen" then the console will almost always seem like the better option. But there are more factors that go into those decisions.
As for the cost comparisons, consoles have always been $200-500, and most GPUs were often in that same range. In other words, you'd almost always have to spend more on a GPU than a console for the higher end PC gaming experience, but that has yet to kill the PC gaming market.
In 2006, a Playstation 3 was $600, but by 2010 it was a $300 system. Video cards were $500 at the top end (GTX 480). So you could argue that the console was a vastly better buy, if you didn't mind it having 4 year old hardware.
Consoles have held their pricing much longer lately, so the $500 PlayStation 5 is still largely $500 over 3 years after launch. Meanwhile, you can get some great graphic cards in the $500-$800 range.
Overall, that's not a vastly different ratio. It's not until you look at the $1000+ video cards that it seems quite skewed, but no one needs those cards to have a good PC gaming experience. Those are extravagances.
I think this is an oversimplification, or a particular perspective. Why is your hardware becoming obsolete every 2-3 years forcing an upgrade if you want to keep playing newer titles an ideal for gamers?
in a backhanded way, the stagnation of the jaguar consoles/quad-core desktop worked out great for consumers, as long as you didn't buy at the very tail end of it. You could spend $200-250 on a CPU and 5 years later the best consumer processor available at the same price point would only be like 20-30% better, meaning you weren't locked out of anything you wanted to play.
Contrast that to the early-mid 2000s in the GPU market where the GPU API specs (DX versions, OpenGL shader model, etc) were moving so fast that 2-3 years later your crappy old GPU might not even be able to launch the game, or would require some hack that allowed you to launch but completely ruined the graphics etc. I remember when hardware T&L became a hard requirement and a bunch of stuff pretty much just stopped working, and there were similar cliffs a couple times with various shader models.
In a technology sense yeah, stagnation is bad and we want it to move forward, but at a consumer level? It's actually kind of good if I don't have to buy a whole new PC every couple years just to keep where I started, the PC market of the 2000s was a massive Red Queen's Race.
That is also why I fundamentally don't buy into the "NVIDIA and AMD don't want to sell you an upgrade" thing. They do, it would be great for them if they could put out a card that was so great that you couldn't even keep your 3080 anymore. That's not the world we live in anymore, moore's law died a long time ago and GPUs fundamentally relied on having more transistors at less cost every product cycle, and we've gotten to the place where all the easy gains have been squeezed. A lot of the more difficult gains were completely shat-upon at first by the public (DLSS and other non-raster tricks were extremely unpopular when the idea was first raised, especially the idea of throwing 5% or so die space at it!) and now they've been largely squeezed too.
Consumers simply just don't want to accept this reality. The entire Turing and Ampere generations were focused on keeping costs down by using cheaper nodes instead of expensive TSMC N7 - that's why AMD (who used TSMC N7) wasn't enthusiastic about competing on price, it's not that "AMD is joining NVIDIA in gouging", it's that those are the actual costs of operating in this market and consumers are still living in 1992 (when they're not making up outright fantasy prices). The x70 segment has always bounced between $349 and $399 even 10+ years ago, with the sole exception of the GTX 970. The "true" 680 was the GTX titan, which launched at $999 (in 2012 dollars!) for a cutdown GK110, a cutdown. Ampere was below the cost curve and a 3070 still cost $499 MSRP, a TSMC GPU would have been even more expensive, and now that NVIDIA is back on TSMC we are seeing prices snap back to the actual cost curve rather than Uncle Jensen's Discount Ampere/Turing Emporium pricing.
https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_proces...
The idea that "NVIDIA just wants to sell enterprise GPUs" is completely insane, both these families share architectures and general architectural work and the economics doesn't work without both of them. Look at the revenue over the last few quarters for what happens without consumer sales. NVIDIA would love if consumers had to buy a new GPU every 2 years to stay in the game. And margins have stayed pretty constant, apart from a secular trend towards more datacenter gpus (and the resulting higher margins). When you consider that NVIDIA is selling 3-4x more datacenter cards than even 5 years ago, there is no evidence of some ballooning margin in consumer GPUs, this is just what they cost now.
(In contrast though AMD is in a very real decision-point between enterprise CPUs and everything else... right now they are trying to cement their gains in the datacenter market before sapphire rapids and other competition finally re-enter the market. So they actually are pushing for datacenter CPU in favor of, say, consumer GPU, along with everything else. Their shipments are massively down and that's why their sales share is under 10% now. NVIDIA though? No, their revenue is in the tank because of consumer GPUs and they very much wish they could sell you something twice as good at the same price as last year.)
But like, people just dive into these bizarre conspiracy theories. It's a margin pump that doesn't show up in corporate margins. It's slow-walking the technology at inflated prices because NVIDIA doesn't like selling you a new card every year. These are consumer goods. NVIDIA makes money by selling them to you as often as possible and by delivering gains that make you want to buy as often as possible. This is just the reality of post-moore's-law technology and spiraling costs from R&D, from TSMC, and on everything else during the pandemic. Yes, volume is way down but they can't sell them to you at a loss either, they'll cut shipments instead.
The good news is prices may finally be breaking. Flash and DRAM are way down and those are the canaries in the coal mine for overall consumer demand. TSMC is reportedly looking to cut prices to try and keep customers from cutting orders. VRAM and power ICs are coming down and stabilizing. All of this will take a long time to trickle through (NVIDIA has been producing Ada for over a year now, a lot of stuff was produced with high-cost components/inputs), and the long-term trend is still upwards (until MCM GPUs really get solved - and by MCM everyone means multi-GCD and not just pulling out memory controllers like AMD did). But there will be some near-term relief over the next 2-3 years until consumer demand starts picking back up again.
Long term though? The first 28nm generation was the GTX 670, launched for $399 in 2012 dollars, for a cutdown 300mm2 die. A 1070 was the next leading-node launch at basically $450 street price until prices finally settled, again for a ~300mm cutdown. A 3070 at $500 was under the average cost-trend because it was a trailing node. Fair price according to cost-trends for a 4070/4070 Ti/whatever you want to call it is probably $649 right now - $799 is still a little inflated but not all that much. Consumers really don't like that fact, they want to go back to imaginary $300 prices (which no x70 card ever launched at). You can't make a ferrari for $20k and you can't even make a yugo for $5k, there are simply some consumers that can't be satisfied and companies have to focus on the ones who can be. "sell at a loss and make it up in volume" is a meme not a business strategy.
In some senses the GTX 970 and the Turing/Ampere (particularly Ampere) generations were a colossal mis-step in strategy from NVIDIA, because they mis-calibrated people's pricing expectations. People aren't gonna remember the generation when you saved them from a 15% price increase, they're not even gonna know that, but they're definitely gonna remember it when prices snap upwards 30% in a single generation and eat up all the perf/$ gains. From a market-psychology perspective a steady 10-15% increase every gen is better than 0% and 30% in subsequent generations.
And for as reviled as Turing was - the 2080 Ti at $999 (EVGA sold them as cheap as $949 remember!) is less than the first GK110 cutdowns (GTX titan, the true 680) which launched at $999 in 2012 dollars. Consumers just have no overall perspective on the price trends in the industry, they latch onto the one time you launched at a cheap price and you'll never live it down. Y'all aren't giving NVIDIA any reasons to try and give you a good low-cost generation in the future if it means more generations of mis-calibrated expectations afterwards.
And that's what it comes down to. NVIDIA is obsessed with cloud gaming and other things that don't make sense to ordinary people (like kids cloud gaming in the back seat of a minivan instead of just playing a switch.) Maybe a gamer keeps a GPU busy for a few hours a day, but a cloud GPU could be shared between multiple gamers and train neural networks when gamers are asleep. From a "value-based pricing" perspective a GPU that sells for $1600 and is utilized 10% of the time could be sold for $16,000 to someone who could keep it busy 100% of the time.
By that perspective NVIDIA can't really afford to sell cards to individual gamers.
They can't afford to continue profitable line of business that took them from nothing to multi-billion dollar corporation?
This is not a new problem - why does Intel sell CPUs to gamers for $100 when theu can sell same CPU for servers for $1000?
Narket segmentation, and you try to make sure these markets buy diffetent products that are priced differently. You artifically create differences in priducts. For example Intel withholds ECC memory from consumer products. Whether its rihht or wrong is debatable.
But please let's not use language like 'we can't afford', It implies they are losing money and subsidising gamers.
Whereas in reality they are very profitable and you are saying they could nickel and dime their customers even harder.
For instance on a bad day it seems that Microsoft would be glad to pour lighter fluid all over the Windows ecosystem if it could get just have 10% of the success Android has had with phones. Or look at Meta neglecting it's current products (even market leading VR games) for the sake of one man's dystopian vision.
A common factor in the above is a lack of competition. If Microsoft was really threatened by the Mac or Linux or if there was some real competition for Office, Microsoft wouldn't be able to neglect its customers in favor of daydreams... And the same is true for Facebook. If advertisers voted with their feet you'd see Zuckerberg burn all those VR headsets tomorrow.
Consider the success of the Nintendo switch. The dock for the switch was a bit of a Jedi mind trick, just like the Wiimote was for the Wii. In particular, it changed the competitive psychology of game producers to think that any "big game" that came out for the Playstation or Xbox would also be released for the Switch. Selling the switch as a hybrid device stops gamers and games company from realizing it is what is... the beginning of the end of the TV-attached console.
The Steam Deck is a much more powerful machine for about the same price and it also has a very strong catalog. The Steam Deck can play Control natively and the Switch can't. Yet for the most part the industry is going to keep remaking the same old games and releasing them for the Switch, even most the innovative parts are going to be making games like Genshin Impact for midrange phones.
Sales of the 4090 are driven by prestige, the pleasure you can get out of saying that you own the very best. They are not driven at all by the reality that you can have anywhere near 10x the pleasure from a $4000 gaming PC than from a $400 portable. However a business that revolves around neural networks very well might be able to generate 10x the revenue from a 10x faster graphics card.
It's not a question of what "we" can afford but it is a function of what "they" can justify to their shareholders. I mean, the people in my town could afford 100% fiber optic internet, but the phone company can't afford to stop using the copper network when it can still charge people premium prices for subprime service. If we decided that $100 a month was a fair price for gigabit fiber and that price caps should be scaled accordingly, my telecom would quickly decide that it "can't afford" to offer DSL at $0.10 a month and would start tearing up that copper network pretty quickly.
I have never met these people, and my sample includes four grandmas that are 70 years old. Quality of videocalls so important that 1 Mbps internet is worth $0 to them. Gaming is leisure, that's different - some people buy $1000 golf clubs, and $5,000 bicycles, some don't know what to do with them.
> I mean, the people in my town could afford 100% fiber optic internet, but the phone company can't afford to stop using the copper network when it can still charge people premium prices for subprime service.
Sounds like a monopoly has your town by the balls, that's not a healthy state of capitalism / free market.
These "artificial scarcity" techniques are cancerous tumours, if we allow them to fester and grow they will bring down the whole system.
Only if you (among other things) ignore the other costs and profit of the cloud gaming service.
How much does continuously streaming 4k bandwidth alone cost over the life of a GPU?
Doesn't matter - the bandwidth used by an average user costs a small portion of a montly subscription fee. High prices for GPUs may be a good way to get more subsribers, who will pay a lot more money over the years.
Citation needed.
> High prices for GPUs may be a good way to get more subsribers, who will pay a lot more money over the years.
And they set price of both markets. Sounds like market manipulation to me.
And that's significantly understating the number of subscribers per GPU, if we're trying to pay off $16000.
Having significantly more subscribers wouldn't give each player much time to play.
Whatever you think the real numbers are, you probably need about half a subscription to pay the bandwidth. I'd be more concerned about the electricity bill, since you'd be in the ballpark of 200kWh a month.
Average driver uses car <5% of the time, but you can't replace 20 cars with 1 uber if everyone needs to get to work at 9 am
https://techgage.com/article/blender-3-4-performance-deep-di...
In 2014, I desperately wanted to buy an AMD card. More perf per dollar, don't feed the nvidia beast, etc etc. AMD had devoted someone to getting the Blender kernels working a few years back, surely now was the time? No, Blender support was still trash. Ok, GPGPU was more important to my use-case anyway. I wasted 2 months trying to write my own compute projects on that AMD GPU. Loads of crashes, leaks (N runs until a black screen and lockup), weird performance issues, debuggers only worked intermittently, it was bad. One time I had spent the whole damn day chasing an OpenCL issue in circles -- only to have the exact, unmodified kernel run flawlessly when I tried it on my buddy's nvidia card. That was it. I sold my AMD card and bought an NVidia one, once more eating the ebay fees, the spread, and the green tax.
10 years later, I hope AMD's success in the CPU space means they now have money and can turn this around. Still, I won't be the guinea pig anymore. I'll buy an AMD card when I see it doing the things I want it to do and not a second earlier. Marketing puffery isn't enough this time. I need to see proof.
You can see evidence of this by looking at the performance in the "No OptiX" chart and seeing the 7900 XTX beats the 3060 Ti there (though maybe not by as much as one might expect).
They certainly did, and what a tragic strategic mistake that was.
AMG GPU's are only good for gaming and nothing else (bad for 3D modeling, bad for GPGPU, bad for studio type work, etc...)
From a pure marketing standpoint, this makes them look like cheap plastic toys, whereas NVidia GPU can not only do gaming superbly, they can also do GPGPU which brings the whole of scientific computing and ML into play and gives them a sheen of seriousness and solidity that AMD simply can't match.
Perception may not be everything, but it certainly is a huge part of everything.
But in the consumer space, yeah, Nvidia is a no-brainer if you want GPGPU.
I thought I remembered that the 7900 XTX trades blows with the 4090 in some games even with both cards at stock settings. (Though of course the 4090 completely blows it out of the water in other games.)
The 7900 XTX will outperform the $1200 4080 in some games using rasterized rendering, but not conclusively across the board.
With ray tracing turned on, the 4080 however generally always comes out ahead in benchmarks, so the 4080 and 7900XTX can definitely be described as trading blows.
There will be almost no benchmark a 7900XTX trades blows with a ~$1600 4090 on (perhaps with exception of performance per dollar...) - the 4090 will substantially outperform a 7900XTX in most things so long as you aren't CPU bound etc - its generally some 30 percent faster than the 4080 which the 7900XTX is roughly equivalent to.
If money is no object for a buyer, the gap between the fastest Nvidia gaming part and the fastest AMD gaming part is pretty big at the moment, assuming you can stomach the MSRP of the 4090. The 4090 is really in a performance league of its own right now.
The 7900XTX's biggest feature today is that its MSRP is 200 dollars less than the 4080, although both cards can be argued to be priced too high this generation relative to the performance offered and price points of prior gens. Rumors abound the 4080 is going to get a price cut soon.
Yes, that's the way averages work ;) 4090 is 127.7% of 7900XTX performance averaged across all reviewers at 4K in non-RT titles. So a good game for a 7900XTX is rubbing shoulders with the 4090... and a good game for a 4090 is doubling the framerate of a 7900XTX.
4090 is not lower power than a 7900XTX though. Which is to be expected because it's a significantly faster card. However, the 4090 does undershoot its specified TDP quite a bit, it's 418W average in gaming (on the next page) yielding 107% perf/w of a 7900XTX (so it's slightly more efficient too, just delivers a lot more frames).
https://www.3dcenter.org/artikel/launch-analyse-amd-radeon-r...
But regarding the 4090 you can drop 100W and barely lose any performance at all e.g. https://i.redd.it/d7lq97nttkv91.png. I assume it's at least somewhat similar for the 7900 XTX as well. The power scaling capping out so hard was really quite the disappointment given the new 600W connector.
In both cases though, you are right that they are both absolute hogs. If thermals are a concern, these are pretty much unusable.
across all launch reviews, 7900XTX averages 351W during gaming, while 4090 averages 418W for 127.7% of the performance (in raster-only 4K, on previous page), working out to 107% of the perf/w.
4080 is actually even more efficient, 297W actual power for 96.9% of 4K performance, giving 115% of the perf/w of the 7900XTX.
7900XT actually falls a bit farther behind, at only 96% of the perf/w.
Compared to a 3080 10GB (which is not a power-hog, that's one of the original fairly-efficient configs) being 70% of the perf/w, that makes the 4080 115/70 = 164% of the perf/w, they are significantly more efficient than previous gens, not at all "power hogs".
6800XT (probably the most efficient card of last gen) is 74% of the perf/w of 7900XTX and that puts the 4080 at 155% of the perf/w of the 6800XT.
Here's the 4070 Ti data but it doesn't really change anything (conveniently the 4070 Ti efficiency works out just about exactly the same as 7900XTX): https://www.3dcenter.org/artikel/launch-analyse-nvidia-gefor...
And please don't whip out the "X watts is too much", arguing that products that you don't want to purchase personally shouldn't be manufactured is a shitty thing to do. If you want a 300W card, buy the 4080. If you want a 270W card, the 4070 Ti is right there. Lower-end cards will be launched over time, it's not going to be high-end-only forever.
Ada Lovelace is by a large margin the most efficient generation that has ever launched. The problem is clearly pricing and not efficiency, people just listed to a bunch of dumb techtubers last summer and earnestly believed that a 900W TBP card was going to be a mainstream product and once gamerzzz latch onto an NVIDIA conspiracy theory they're sure as hell not going to renounce it. Efficiency bad, it doesn't matter what the numbers say. Oh, efficiency is good? Well still too much power. Oh there's cards available at the same power as last gen? Listen here buddy NVIDIA bad, did I stutter?
My only concern is the amount of heat - I’m not trying to save some money by buying high-end gaming card.
NVIDIA used older nodes on Ampere/Turing while AMD moved ahead to TSMC N7. NVIDIA is just so far ahead architecturally that they can roughly match AMD's efficiency despite a full-node gap in technology and going against (specifically) the best node in the industry.
The older node let them reduce costs, which is part of the reason why AMD has been perceived as reluctant to play the price wars during Turing/Ampere. NVIDIA was playing with older cheap nodes and AMD was basically using the best, most expensive node available just to compete (not even significantly beat), and a GPU needs a nice big chunk of it that could easily be several high-margin CPU dies instead. Nobody is ever going to publicly attribute a positive intent to NVIDIA ever, but, Ampere and Turing were an earnest effort to keep the cost down for consumers. 2080 Ti was the same $999 MSRP as the first GK110 card 10+ years ago. 3080 was so exquisitely priced that AMD sighed and dragged their feet and finally marked theirs 50 bucks less and threw in the towel and didn't barely produce it for the first year. 3060 Ti should have been the next RX 480 level value-proposition that keeps mainstream gamers happy for the next 5 years. Mining ruined everything - yet again.
The two negative consequences of this were efficiency and top-end performance. Samsung 8nm is a 10+ node, basically a full node behind TSMC N7, and obviously that has perf/w consequences. NVIDIA's architectural advantage mostly held them over but they still slid in a little under RDNA2's efficiency, when if they'd been on N7 themselves they could have easily been 20%+ more efficient. And it resulted in huge dies, because the perf/mm2 is much worse on a 10+ than on N7. So they didn't really have a top-end - a 3090 is really more comparable to what they could have gotten out of a full-die x104 on N7, but 754mm2 is as big as they could go on Samsung 8. That is part of why the 4090 towers above the rest of the lineup in a way the 3090 didn't. 30-series didn't really have a high-end part, they had "castoffs" (full dies) from the mainstream-enthusiast part that happened to have a little extra headroom.
A little bit of product strategy - you put your cutdowns at the high-value points of your lineup. GTX 970 and 980 Ti being cutdowns, 1080 Ti being a cutdown, etc. In CPUs, that's where AMD puts the 3600 and 3900X, and Intel puts the 5820K, etc. With Ampere, NVIDIA went as big as possible, then put the 3080 as being the GA102 cutdown, to make it a great value. Same for 3060 Ti. And the 3090 and the 3070 are the bloopers that result from the full-die chips being more expensive. But with the AD102 die being 754mm2 already (fucking enormous), that's as good as they could do on Samsung 8nm, so they kind of abandoned the true top-end stuff.
That was the strategy on 8nm - big dies on a shitty node, and make the 3080 a cutdown to take advantage of it. And on Turing it was the 2080 Ti (because 12FFN was the same node as Pascal was on, it's a 16nm with no optical shrink just the "your mom sized" reticle limit) plus pushing some of those "non-raster" performance-scaling approaches like DLSS.
They also needed to bring in a custom memory tech (GDDR6X is NVIDIA-exclusive) to keep costs down - Samsung 8nm has terrible SRAM density compared to TSMC N7 so the "infinity cache" strategy is off the table (or at least you'd need to stack it, and the packaging step drives up costs a lot too). 512b buses are tough to route these days and a lot of people don't think it's possible anymore with tighter integrity requirements (again, maybe it'd be possible with fanout packaging which, quelle surprise, drives up costs). But PAM4 lets you move a lot more data per clock edge using basically the same IC form-factor. The only real alternative would have been HBM2 which would also have driven up costs. So NVIDIA traded a custom memory tech to allow them to use a shitty node that they could get cheaply and produce these massive, fairly low-clocked dies on what amounts to an overgrown smartphone node, all in the name of cost reduction and bending MSRPs downwards ($399 3060 Ti, $699 3080). And miners royally fucked it all up, yet again.
But anyway to go back to the original point - this gap has always existed, AMD just appeared competitive because they were using a better, more expensive node, and they took it on the chin in margins as a result. And now that NVIDIA has moved back to a leading node (very leading in fact) the gap has reappeared. People just didn't want to do the math on comparing products across two nodes and realize the gap was always there, but people like me have been saying something along the lines of "RDNA1 isn't all that great considering it's N7 vs 16nm and NVIDIA still is slightly ahead on perf/w" and "RDNA2 is moving forward but it's still barely beating NVIDIA even though they're using a garbage Samsung 10+ node" quite frequently and we got shouted down.
I agree that it seems like RDNA3 missed somewhere though. The efficiency gains are massively lower than what AMD promised in their earlier slides, the performance seems disappointing. Maybe our expectations were just mis-calibrated, and maybe the marketing slides were just cooked, but it does seem like it missed some target by 20-30% and nobody external knows what the specific reason is.
It's even better than the newest Intel GPUs (except maybe for raytracing, but that's a thing for 'tomorrow')
Interested in the newest available mobile chips? RDNA2 is the latest and greatest.
Steamdeck? Game console? It's the most recent there too.
That said, they've recently released architecture articles about stuff like core from 15 years ago. It's not always about grinding out articles on the latest stuff.
It's even better than the newest Intel GPUs (except maybe for raytracing, but that's a thing for 'tomorrow')
now 0 pts, something fishy at HN HQ ;)
And even further, even on the RTX parts, using RT usually requires rendering at a much lower resolution and then upscaling. With DLSS3, they just go ahead and start inserting extrapolated frames to get the FPS numbers up. It's hard to believe gamers are willing to make that trade to be honest. If it was pitched together as all one package many years ago, they'd have been laughed out of the room.
Hardly, every game I play now has RT and RDNA2 is crap at it. It was their first go at it sure, but RDNA2 is also what the consoles have to put up with so the subpar raytracing performance had an outsized effect on the industry.
>RT in games during RDNA2's lifetime, and even today, is little more than a gimmick
Absolutely, categorically false.
>that is barely noticeable while actually playing.
This is not true in the slightest.
>It isn't even real ray tracing.
It's true that raytraced effects is not _full path tracing_. It still makes an enormous visual difference via things like RTGI.
>And even further, even on the RTX parts, using RT usually requires rendering at a much lower resolution and then upscaling.
Yes, this is another area AMD was way behind and recently almost caught up. DLSS2.x is amazing, letting me render at 1440p scaled to 4K with an output that _exceeeds_ 4K native quality in many games/contexts (thanks to 4K native also running TAA, which has other issues).
>With DLSS3, they just go ahead and start inserting extrapolated frames to get the FPS numbers up
You're using outdated terms. DLSS3.x includes _both_ the 2.x image upscaling tech and DLFG (deep learning frame generation) that was previously called "DLSS3". DLFG is an optional feature of the 3.x SDK that's only supported on 4000 series GPUs. DLFG also works very well, and both initial reviews and user feedback on reddit has been overwhelmingly positive.
>It's hard to believe gamers are willing to make that trade to be honest.
What trade?
>If it was pitched together as all one package many years ago, they'd have been laughed out of the room.
This entire post just seems like AMD cope, rather than a fair evaluation of the current state of RT. RDNA3 upped RT performance significantly but it's still not enough.
More input latency plus whatever quality loss is incurred to pull frames out of thin air.
The latency is a big trade off for competitive reaction-time-based games.
It's often true that people can't easily feel the difference of a few extra ms or tens of ms in input latency. It is nonetheless a statistical disadvantage, as a players reaction times will be too slow in a higher percentage of cases since they've had a bit of their reaction window robbed by input latency.
Many games do just fine on latency without Reflex.
Remember that framegen's latency is inversely proportional to framerate in general... if you're starting from a (DLSS-assisted) 70fps framerate in a cinematic single-player game, that's probably fine. If you're playing a twitch shooter where the framerate is 200fps, your framerate is already high enough that the extra 0.5ms latency is probably fine.
People have been consistently negative about these new technologies from NVIDIA ever conceivably working or improving in the future, while absolutely fellating anything AMD comes out with. Framegen really is a solid option even right now and it's only going to get better over the next 2 years. The obvious trajectory here is that NVIDIA will continue to get better and people will still pooh-pooh it up until AMD releases FSR3 and suddenly wow, it's great, even though it is worse than NVIDIA's initial offering was. Remember they don’t have the optical flow engine and even nvidias first-gen engine is significantly worse quality.
It's so predictable at this point, I want AMD to do well but the social media story hasn't changed in 10 years to be honest, everything AMD does is amazing and benevolently pro-consumer and is gonna be great in a couple more iterations, and everything NVIDIA does is awful and malicious/avaricious and could never possibly work.
Just wait for FSR3 (with FluidFrames TM!) and RDNA4. Yup, that cycle is starting now too.
https://www.tomshardware.com/news/amd-fsr-3-announced
https://hothardware.com/news/amd-promises-rdna-4-near-future
Like I guess framegen is so awful that AMD is gonna spend like ten million bucks of dev time making their own. That kinda says it all.
Clearly a post made in bad faith. You seem to think I don't have a lot of experience playing these games, if you are going to assert that RT makes so much difference to the player experience (other than reducing performance).
I am certain that the majority of gamers could not effectively discern what "RTX On" actually means without being trained to know what to look for. You hear things like "it's shinier" from people. This is often true even when they are looking at clips or screenshots, let alone while they are concentrating on actually playing. What percentage of players even understand that a game advertising RT may only be using it for shadows, for example?
As we have this discussion, the biggest release is Hogwarts Legacy, for which there are support forums full of people complaining about performance issues so they can see a stained glass window reflected (on a stone floor!?) like its a mirror.
I hope you understand that the marketing spin to which I was referring is the very idea that RT is needed to make a modern game look good in the first place. You seem to really care about it. To me it's just another thing to put on the box like 4k, 8k, HDR, etc. Things that do look good, especially in awe-inspiring screenshots, but have essentially no impact on my actual enjoyment of the game.
You are woefully uneducated on the subject at best, and at worst you’re downplaying the importance of accurate lighting to cover for AMD’s pitiful first attempt.
I’m sure this will all go away when RDNA4 has acceptable RT performance and suddenly it’s the best thing ever.
Granted, at the time I was only targeting 4k60 and now I target 4k120
Honestly though, I feel like most people would rather get higher framerates.