It's just the classical socratic method applied to characters that you particularly don't identify with. Fasterthanlime's bear and why's guide have often been praised here for the same thing (the interludes break up the long, boring blocks of text that deep dives often lose readers on,) and it's trivial to use a Reading Mode or skip them for the legitimately interesting technical information within.
Unfortunately reading mode didn't remove the characters for me, it actually made them occupy the whole screen LOL (Fennec on Android if anyone is wondering). I don't have that much of a problem with the conversational format as long as it's used in an appropriate context, but this one was really intrusive and broke the flow of the text, since I'm reading mostly for entertainment I decided it wasn't worth the time.
I'll do what I can. I fear the problem will boil down to not having control of CSS in reader mode (which is understandable, really) and I may need to admit how my CDN serves the sticker files to serve lower quality assets. I hope it's a simple fix. Gonna do the fix live on Twitch tomorrow: https://twitch.tv/princessxen
Tested against dataset consisting of only multiple choice questions and was only able to achieve 67% accuracy on MedQA (the medical licensure examination.) LLMs alone are not the way forward in this area (and in this instance not generalizable, nor capable of handling necessary edge cases the job demands.)
These are far inferior to professionals and just like what we saw with Watson, hyping them up as job replacements or assistants when they're clearly not up to snuff is not only foolish for the perception of the industry as a whole, but wildly irresponsible.
The idea that they can't improve the UI/UX to better inform to the people who repeatedly, accidentally send insecure messages/sms (ignoring the existing words "Unsecured SMS" in the chat field, the unlocked lock near messages, the unlocked lock next to the phone, or the giant banner that occasionally drops down that tells you the % of secure messages you can be sending if you pester a contact into grabbing signal) as one of the reasons for this change is frankly bullshit.
Changing the Send button's icon to "SMS" or a color/border change ala iMessage are ideas off the top of my head and I'm sure they've got designers significantly more talented than I am that can think of better ones. We've seen very little iteration there that's indicated the significance of that problem...and frankly if they highlighted this as a tactic vs endless spam texts more people would be receptive to this news. As it stands I think this is going to significantly reduce their number of casual users. In fact I'm willing to bet that the cohort of users who are used as justification are the least likely to convince their contacts to switch to Signal.
Don't get me wrong, their real desire to increase the amount of people sending secure messages via Signal alone + resource mgmt in the face of a recession are valid. But acting as a unified messenger (with better link unfurling, threaded replies, and reactions after Google killed Allo vs the default messenger that spent years getting them) was the trojan horse onto many of my friends' and colleagues' phones. Now that there's parity I can see more people just opting into the default messenger/FB Messenger + Whatsapp combo because more people exist there and we're all just lazy.
I've followed Hail and applaud the Broad Institute's work wrt establishing better bioinformatics software and toolkits so I hope this doesn't come as rude, but I can't imagine an instance in a real industry or academic workflow where you need 300ms feedback from an experiment to "maintain flow" considering how long experiments on data that large (especially exome sequencing!) take overall? My (likely lacking) imagination aside I guess what I'm really saying is that I don't know what's preventing the usecase you've described from being performed locally considering there'd be even less latency?
300ms is my ideal latency, but we donβt achieve that under all circumstances. Even for blob storage, I see as much as 100ms latency. That said, my laptop has maybe 8 cores. Even if I had 0ms reads from an SSD, Iβm compute bound for some tasks.
Moreover, I think we have differing definitions of βexperimentβ. In the context of a sequencing study, I think an βexperimentβ can be as simple as answering the hypothesis: does the missingness of a genotype correlate with any sample metadata (e.g. sequencing platform). You might try to test that hypothesis by looking at a PC1-PC2 plot with points colored by sequencing platform where the PCA is conducted on the 0/1 indicator matrix of missingness.
In the dry lab, that is what I mean by experiment. By that definition, a scientist does many experiments a day. Particularly for sequencing studies, these experiments are data-intensive, I need to run a simple computation on a lot of data to confirm the hypothesis.
These prices are beyond insulting and frankly I'm glad they're going to take a hit competing with the flood of miner cards entering the market.
Also note that nothing is preventing Optical Flow Acceleration [0] (and subsequently the DLSS 3.0 models that they claim are exclusive to the 40 series) from running on either 2/3 RTX cards. Just like RTX Voice and other gimmick "exclusives" I expect it to be available to older cards the moment they realize their backlog of 30 series cards aren't clearing as quickly as they thought.
They're competing against an over-bloated secondhand market, AMD, Intel, much better integrated GPUs, and comparatively cheaper consoles that maintain sky-high demand with subsidized games via new subscription programs. They're vastly overestimating their brand loyalty (think Microsoft v Sony after the 360) and EVGA's exit makes more sense now than ever.
IMO they still have too many 3xxx cards and don't want to cannibalize sales, hence adding them to the top of the stack instead of replacing the 3080/ti (and renaming their 4070 as 4080 12gb).
The feeling I get from the whole thing is NV doesn't care that much about consumer graphics anymore, and doesn't care about providing competitive value. More of their revenue is coming from business applications, all of which is growing, and gaming looks like its down to 1/3 and shrinking.
Me and most of my PC gaming friends gave up on this hobby in the past two-three years due to the insane GPU prices. I moved to the Oculus Quest 2 and they moved to PS5 and Steamdeck since the price of a modern GPU alone buys you a whole stand-alone gaming console today. Crazy stuff.
I assume the 4000 series is basically a paper launch designed to make the existing 3000 inventory look attractive and make the people who were waiting on the 4000 series being good value rethink and just buy a 3000 instead.
Just to echo the other children of this comment; I'm pretty lost as to the reasoning here.
I game at 1440p on a GTX 1080 from launch day in 2016. Cost is essentially no object for me, as gaming is my primary hobby and I can easily spend 60hrs/week playing video games, so the return would be more than justified in my mind.
But I haven't upgraded in 6 years because there has been no reason to. I'm running my favorite games at 144fps, and the newer super-high fidelity stuff I can run at 60fps, which is more than reasonable.
I will say that I was excited for the 40XX lineup, but at these prices I'm more inclined to hunt for a cheaper 2080 or 3080 solely for RTX functionality, but that's an enthusiast feature even for someone like me who spends a huge amount of time gaming.
I couldn't imagine giving up my access to old titles, console emulators, mods, itch.io experiments, and low-budget indie games on Steam to save a few hundred bucks in the short-term to buy a console.
YMMV if you're not paid well in a HCOL area, or if your location offers alternatives such as PC cafes.
Youβre missing the part where you get pipe cred by posting timespy scores in Discord after installing your 4090 and dedicated ac unit.
I historically have bought 1-2 gen old cards and it has worked well for me too.
The big filter right now is 4k 120hz. If youβre aiming for that you need a recent card with hdmi 2.1, which is like 3k nvidia 6k amd series. I upgraded displays recently and it finally forced me, begrudgingly, onto a more current gen card. Really wanted to wait for 4k/7k lines to ditch my old 1080Tis.
No. I am almost certainly an outlier, even among my peer group.
No kids, taking a break from work, very few responsibilities.
When I have a full time job my playtime drops, but 1-3 hours on weeknights plus large binges on weekends can keep the number surprisingly high, though.
This is mostly a matter of multitasking (which I know HN is fairly divded on).
I play games when I watch TV, and I also use games to socialize with my friends.
If an average person added up ~80% of their socialization time and all of their TV/movie/audiobook/podcast time, I suspect that it'd look pretty close to the amount of gaming I do when I have a full time job (probably fluctuates between 10-40 hours per week).
Weekends could easily be > 1/3 of that time (say 24h over 2 days),leaving a reasonable ~5 hours per week day. I don't know about typical, but I've had such a schedule at points in my life when playing games I really enjoyed.
Yes, for some games, but not on a consistent basis.
I still enjoy games, but with adult life I'm playing either after diner (if I'm not too tired) or on the weekends, and gaming competes with other entertainment (series, movies, going out with friends/dates with SO, etc). So as I've grown it takes less time overall.
I am, however, making a schedule for Homeworld 3 and KSP2 when they launch (:
- buy a $800 GPU, keep PC gaming and all the fun maintenance associated with that
- buy a cool new toy like a Xbox Series X/PS5/Steam Deck
- don't do anything, game at 720P with no AA (this sounds sort of like when I had lunchables for dinner yesterday)
Your alternative isn't wrong but they weren't looking to optimize for staying on a PC, but rather, for gaming, and were already thinking in terms of "how much of my disposable income do I want to spend on my hobby?" rather than "should I spend my disposable income on my hobby?"
I had a PS5. I sold it after playing a few exclusives because there was just nothing else to play. Games are more expensive on consoles than PC, no mods, have to pay to play online, and so on.
Plus, the very fact that it's a general purpose computer I can use for work, programming, and cool GPU stuff like Stable Diffusion are what made me get a PC over a console. In other words, I'm gonna need a computer anyway, why not get one that can play games over getting a computer or laptop and also a games console that doesn't do everything as well as a gaming PC?
I'm sure in the long term one would actually spend less on a PC than a console, given all these constraints.
Dual classing a pc as a tv and desk PC can be pretty nice once you get it set up. Something like Nvidia Shield, Moonlight Gamestream software (if you have an Nvidia card), and Playnight launcher can work great if you can hook it all together with ethernet or MOCA over COAX.
Having run an over powered desktop to enable my gaming hobby, I have been toying with putting a "thin client" on my desk and my beefy rig in the living room.
Computer performance has felt ~stagnant for the past decade. Given the number of platforms that build to be mobile friendly, there is increasingly less need to be constantly running a power hungry rig as my main configuration when I am spending most of my time browser the internet. If I want to "desktop game" I could stream from the living room to my desk in the event it is not a controller friendly experience.
> PC -> TV experiences are also not great in general.
Why not? I recently bought an inexpensive gamepad, hooked up TV via HDMI and added a wireless keyboard+touchpad to the mix (I can't get used to Steam console mode). Overall experience is pretty good, effectively indistinguishable from PS4 we had in the office.
It drops into desktop, you have to drag out the keyboard / mouse, it's connected as another monitor on hdmi, so if someone else turns on the tv, that starts interacting with your computer as a second screen. Not to mention the extra annoyances when the tv is in another room.
Streaming boxes like the nvidia shield or similar have their own issues, along with dropping to desktop and so on.
While a dedicated TV console has none of those issues.
I just do both really. There are some games that are just better sitting on the sofa that can be started fast and no fuss. Then there are some games that are maybe more of a slow Sunday afternoon at a desk vibe.
First point is disingenuous, itβs not like you only have that one option. Thereβs plenty options at lower price points which are still decent. And thereβs the second-hand market for the budget-conscious. You can get a decent pc for around 1k, and much lower with second-hand market options.
For the maintenance, itβs not like youβre buying a car. Unless youβre enthusiastic about your pc, you can get a pre-built, plug it in and use it, same as console.
Not to mention that games are cheaper and thereβs more cheap indie titles available, so if one is optimising for max gaming enjoyment for the least money, pc looks like a better option. But if the plan is to play through the best titles here and there, then console is a good option as well.
There is far more maintenance or rather troubleshooting with a PC. The Steam Deck too TBH (even on "Verified" games you run into the occasional problem with more frequency than a Switch). You're just far more likely to run into some weird problem or issue with a random title especially that 10 year old game you bought for $1.
On console you match the label to the machine and it works.
This is my story. I've been a PC-first gamer since I was a kid. Diablo was my first "big" game. The pricing of PC components and GPUs in particular have forced me to take a step back and move to console first with a Steam Deck supplementing.
For the price of one mid level 40 series card I can buy two consoles (Xbox Series X + Digital PS5). I know that the games themselves end up costing more, but I don't really care about the majority of new games coming out and the ones I do are often on game pass.
"I know that the games themselves end up costing more"
Just got PS5 and the PS Extra membership whose library has almost every game I wanted. Maxed out my SSD storage on day 1. And I think the XBox subscription is even better. I mean PC is still cheaper but the difference is not that big anymore.
I know, I have a 2080 at the moment. But I don't like where the PC industry is headed so I am not going to support it going forward unless something changes.
You don't have to buy 40 series. Actually, you even don't have to buy 30 series. Even more actually, for gaming only you may buy AMD card, more or less th same that used by Xbox and PS5.
Even at 300 dollars, a 350w 3080 still looks pretty bad. I'd much rather watch prices of the 6800xt, at least until AMD makes their RDNA3 announcement. It also runs (well) on linux too :).
I occasionally play games, including some recent AAA releases. Everything plays fine at 1k/medium-high settings or 4k/medium-low settings on my GTX 1060. No AA, but I don't care that much. The age of needing a new graphics card to play the new games is basically over. IIRC the GTX 1060 is still the most popular GPU on steam. I just "upgraded" to a Radeon Pro VII (for RoCM), with basically equivalent graphics performance, and everything still plays very well.
>the price of a modern GPU alone buys you a whole stand-alone gaming console today.
This has been true for as long as Iβve seen. Mid-range GPUs have historically tended to cost about as much as whole consoles. Pascal, one of the most well-received lines in terms of pricing, sold the GTX 1070 for $400 MSRP, the same price as buying a whole PS4. These new prices being the equivalent of buying two whole machines is whatβs truly intense.
Tom at MLID pretty much confirmed as much in his recent two broken silicon podcasts. Nvidia is pricing this way so they can run the 4k and 3k lines concurrently.
I believe that you're looking at the same chart that I was thinking of when I saw this line, which likely incorporates crypto GPUs into the "gaming" segment.
The 3080 released for $800 USD in Jan 2021 which one calculator[1] says is $905 USD with inflation. So a 4080 is $900. At least they are consistent.
edit: 3080 12 GB, that is. The original 3080 was $700 USD in Sep 2020 which is $800 in 2022. So the base 4080 is priced above that.
They're competing against an over-bloated secondhand market, AMD, Intel, much better integrated GPUs
For the 4080/4090 pricing, the only competition is AMD.
Just glancing at some benchmarks, the Intel ARC A380 is worse than a GTX 1650[2] and the best integrated graphics I know of is the Ryzen 7 5700G which is worse than a 1050[3]. I don't see why these would affect 4080/4090 pricing at all.
The $900 one is the 12GB version, which is basically a 4070 compared to the 16GB version, because the memory is not the only difference, but also a different core count, memory bus size and clock speed, which is frankly even more insulting.
The 3080 (10GB) and 3080 (12GB) also had a core count difference, although they did not release simultaneously.
But I realized I did give the price for the 3080 (12 GB) model, which I corrected in an edit. So I think you're sort of right in the sense that the 4080 (12GB) is priced above the 3080 (10GB) in real terms, and that the 4080 (16GB) is priced way above the 3080 (12 GB) in real terms.
Theyβre different markets though. Itβs like $50 bluetooth keyboards for the everyday user versus $200 mechanical keyboards with replaceable keycaps. The latter intentionally requires tinkering because itβs geared towards an enthusiast that enjoys the tinkering.
Outside of the true exclusives, console games are just a subset of PC games. So as a PC gamer, you get access to console games + the majority of MOBA, MMO, RTS, 4X.
I really miss keyboard and mouse on gaming console. It's weird that devices originally for non-gaming is better than a gaming oriented device, but gamepad is really painful for some games like FPS/TPS.
No they aren't. All signs point to ARC's first outing being a swing & a miss, and Intel themselves didn't even have any plans to come anywhere close to competing at the top end. Intel's highest end ARC card, the A770, was positioned against the RTX 3060. And Intel is both missing any semblance of a release date it set for itself, but the drivers are also a disaster and rumors are Intel is going to shutter the entire division, focusing only on data center compute cards only.
So used cards yes are a factor, yes, but if this really is a 2x performance jump then those aren't going to be a factor here. None of the $1000+ GPU buying market is going to care about saving $100-300 to get half the performance after all.
That just lives AMD as competition, and only if RDNA 3 comes out a winner. Which the rumor mill is saying it will be, but AMD's graphics track record is spotty to put it mildly. I'm sure Nvidia's internal plan is to just cut prices if RDNA 3 surprisingly competes at the top end instead of just the midrange, but in the meantime is happy to take those fat fat margins.
> That just lives AMD as competition, and only if RDNA 3 comes out a winner.
The RX 6650 XT is in the same price bracket as the RTX 3050. I'm not sure they would be considered competition even if RX 77whatever were in the same price bracket as the RTX 3050. Nvidia just seems to push just the right buttons on the software features which AMD just cannot seem to get.
Don't get me wrong. I would love AMD to be competition. They should frankly team up with Intel. AMD has the hardware and Intel can provide the software. Don't see that happening considering AMD went alone on their FP8 proposal and everyone else joined up on a separate paper.
Intel's GPU software is currently, and always has been, a complete disaster. AMD's software is miles ahead of Intel's. Nvidia does have more stuff if you care about it though, although the suite of that is increasingly becoming more annoying than valuable. Like GeForce experience constantly demanding to be logged in & collecting telemetry so I can do... Driver updates? So dumb.
But there's occasionally the random great piece of software like DLSS 2 or RTX Voice, definitely. AMD kinda tries with stuff like FSR or Anti-Lag, it's just not as well done as Nvidia's version.
Bingo, Nvidia isn't making enough A100, A10G to fill demand from GCP/Azure/AWS. The idea that they are worried about their position in the gaming industry is off mark.
What's happening is that with mining dying they know the gaming market won't magically expand, there will be a few years of reduced interest. The only way they can make it look good on their balance sheet is if the entreprise segment grows and that means reassigning their capacity towards it even if it means reducing volume on the consumer side.
From a business point of view this makes perfect sense.
My guess is, the problem is the lower end. I'm guessing they bought a lot of fab capacity for the 40 series given fab capacity was highly constrained over the recent years and they couldn't produce graphics cards fast enough.
Now masses of 30 series are being dumped onto the used market from a combination of ex-mining card and ex-gaming cards being sold in preparation for the 40 series. Gamers who were looking for 3060/3060ti/3070's can probably pickup a used 3080 for comparable money on ebay.
So what does the demand for the 4060 look like? I'm assuming a 4060 is a binned down part that didn't make the cut for higher models? You wonder if nvidia is left with a lot of chips that would've gone to the 4060 which they can't sell.
In 2019-2021, IMO gaming market was important because new PS and Xbox were start selling . It was time to get customers to their ecosystem before they could get new console. Now shortage is end, so selling latest model is no longer important.
AMD's decision to have different architectures for gaming and datacenter is still a major mystery. It's clear from Nvidia's product line that there's no reason to do so. (And, yes, Hopper and Ada are different names, but there was nothing in today's announcement that makes me believe that Ada and Hopper are a bifurcation in core architecture.)
Moreover, CDNA is not a new architecture, but just a rebranding of GCN.
CDNA 1 had little changes over the previous GCN variant, except for the addition of matrix operations, which have double throughput compared to the vector operations, like NVIDIA did before (the so-called "tensor" cores of NVIDIA GPUs).
CDNA 2 had more important changes, with the double-precision operations becoming the main operations around which the compute units are structured, but the overall structure of the compute units has remained the same as in the first GCN GPUs from 2012.
The changes made in RDNA vs. GCN/CDNA would have been as useful in scientific computing applications as they are in the gaming GPUs and RDNA is also defined to potentially have fast double-precision operations, even if no such RDNA GPU has been designed yet.
I suppose that the reason why AMD has continued with GCN for the datacenter GPUs was their weakness in software development. Until today ROCm and the other AMD libraries and software tools for GPU computational applications have good support only for GCN/CDNA GPUs, while the support for RDNA GPUs was non-existent in the beginning and very feeble now.
So I assume that they have kept GCN rebranded as CDNA for datacenter applications because they were not ready to develop appropriate software tools for RDNA.
Some guy on Reddit claiming to be an AMD engineer was telling me a year or so ago that RDNA took up 30% more area per FLOP than GCN / CDNA.
That's basically the reason for the split. Video game shaders need the latency improvements from RDNA (particularly the cache, but also the pipeline level latency improvements, each clock an instruction completed rather than once every 4 clocks like GCN).
But supercomputers care more about bandwidth. The once every 4 clocks on GCN/CDNA is far denser and more power efficient.
But what kind of latency are we talking about here?
CDNA has 16-wide SIMD units that retires 1 64-wide warp instruction every 4 clock cycles.
RDNA has a 32-wide SIMD unit that retires 1 32-wide warp every clock cycle. (It's uncanny how similar it to to Nvidia's Maxwell and Pascal architecture.)
Your 1/4 number makes me think that you're talking about a latency that has nothing to do with reads from memory, but with the rate at which instructions are retired? Or does it have to with the depth of the instruction pipeline? As long as there's sufficient occupancy, a latency difference of a few clock cycles shouldn't mean anything in the context of a thousand clock cycle latency for accessing DRAM?
EDIT: So it looks like my memory was bad. I could have sworn RDNA2 was faster (Maybe I was thinking of the faster L1/L2 caches of RDNA?) Either way, its clear that Vega/GCN has much, much worse memory latency. I've updated the numbers above and also edited this post a few times as I looked stuff up.
The weird part is that this latency difference has to be due to a terrible MC design by AMD, because there's not a huge difference in latency between any of the current DRAM technologies: the interface between HBM and GDDR (and regular DDR) is different, but the underlying method of accessing the data is similar enough for the access latency to be very similar as well.
Or... supercomputer users don't care about latency in GCN/CDNA applications.
500ns to access main memory, and lol 120 nanoseconds to access L1 cache is pretty awful. CPUs can access RAM in less latency than Vega/GCN can access L1 cache. Indeed, RDNA's main-memory access is approaching Vega/GCN's L2 latency.
----------
This has to be an explicit design decision on behalf of AMD's team to push GFLOPS higher and higher. But as I stated earlier: video game programmers want faster latency on their shaders. "More like NVidia", as you put it.
Seemingly, the supercomputer market is willing to put up with these bad latency scores.
When we look at RDNA, we can see that many, many latency numbers improved (though throughput numbers, like TFLOPs, aren't that much better than Vega 7). Its clear that the RDNA team did some kind of analysis into the kinds of shaders that are used by video game programmers, and tailored RDNA to match them better.
> I've never seen any literature that complained about load/store access latency in the shader core. It's just so low level...
Those are just things I've noticed about the RDNA architecture. Maybe I'm latching onto the wrong things here, but... its clear that RDNA was aimed at the gaming workload.
Perhaps modern shaders are no longer just brute-force vertex/pixel style shaders, but are instead doing far more complex things. These more complicated shaders could be more latency bound rather than TFLOPs bound.
Nvidia have been making different architecture for gaming and datacenter for few generations now. Volta and Turing, Ampere and Ampere(called the same, different architectures on different node). And Hopper with Lovelace are different architectures. SMs are built differently, different cache amounts, different amount of shading units per SM, different rate between FP16/FP32, no RT cores in Hopper and I can go on and on. They are different architectures where some elements are the same.
No, the NVIDIA datacenter and gaming GPUs do not have different architectures.
They have some differences besides the different set of implemented features, e.g. ECC memory or FP64 speed, which are caused much less by their target market than by the offset in time between their designs, which gives the opportunity to add more improvements in whichever comes later.
The architectural differences between NVIDIA datacenter and gaming GPUs of the same generation are much less than between different NVIDIA GPU generations.
This can be obviously seen in the CUDA version numbers, which correspond to lists of implemented features.
For example, datacenter Volta is 7.0, automotive Volta is 7.2 and gaming Turing is 7.5, while different versions of Ampere are 8.0, 8.6 and 8.7.
The differences between any Ampere and any Volta/Turing are larger than between datacenter Volta and gaming Turing, or between datacenter Ampere and gaming Ampere.
The differences between two successive NVIDIA generations can be as large as between AMD CDNA and RDNA, while the differences between datacenter and gaming NVIDIA GPUs are less than between two successive generations of AMD RDNA or AMD CDNA.
Turing is an evolution of Volta. In fact, in the CUDA slides of Turing, they mention explicitly that Turing shaders are binary compatible with Volta, and that's very clear from the whitepapers as well.
Ampere A100 and Ampere GeForce have the same core architecture as well.
The only differences are in HPC features (MIG, ECC), FP64, the beefiness of the tensor cores, and the lack of RTX cores on HPC units.
The jury is still out on Hopper vs Lovelace. Today's presentation definitely points to a similar difference as between A100 and Ampere GeForce.
It's more: the architectures are the same with some minor differences.
Turing is an evolution of Volta, but they are different architectures.
A100 and GA102 DO NOT have same core architecture. 192KB of L1 cache in A100 SM, 128KB in GA102 SM. That already means that it is not the same SM. And there are other differences. For example Volta started featuring second datapath that could process one INT32 instruction in addition to floating point instructions. This datapath was upgraded in GA102 so now it can handle FP32 instructions as well(not FP16, only first datapath can process them). A100 doesn't have this improvement, that's why we see such drastic(basically 2x) difference in FP32 flops between A100 and GA102. It is not a "minor difference" and neither is a huge difference in L2 cache(40MB vs 6MB). It's a different architecture on a different node designed by a different team.
GP100 and GP GeForce has a different shared memory structure as well, so much so that GP100 was listed as having 30 SMs instead of 60 in some Nvidia presentations. But the base architecture (ISA, instruction delays, β¦) were the same.
Itβs true tbat GA102 has double the FP32 units, but the way they works is very similar to the way SMs have 2x FP16 in that you need to go out of your way to benefit front them. Benchmark show this as well.
I like to think that Nvidiaβs SM version nomenclature is a pretty good hint, but I guess it just boils down to personal opinion about what constitutes a base architecture.
AMD as well. The main difference being that Nvidia kills you big time with the damn licensing (often more expensive than the very pricy card itself) while AMD does not. Quite unfortunate we do not have more budget options for these types of cards as it would be pretty cool to have a bunch of VM's or containers with access to "discrete" graphics
Nvidia's datacenter product licensing costs are beyond onerous, but even worse to me is that their license server (both its on-premise and cloud version) is fiddly and sometimes just plain broken. Losing your license lease makes the card go into super low performance hibernation mode, which means that dealing with the licensing server is not just about maintaining compliance -- it's about keeping your service up.
It's a bit of a mystery to me how anyone can run a high availability service that relies on Nvidia datacenter GPUs. Even if you somehow get it all sorted out, if there was ANY other option I would take it.
While I do wish they priced lower, I think when you put things into perspective, it blows my mind that I can buy a card with 76 billion transistors(!) for just 1600$. I suspect the demand for top of the line RTXs will come from enthusiast gamers and ML researchers/startups. Most gamers would be more than fine with 3090 which handles almost anything. The market will be very interesting to watch.
Virtually any contemporary game with maximum settings will struggle to maintain 4k120 without a sacrifice on my 3090. Cranking the render scale in WoW to 6K or 8K (retail or Classic) will easily push the GPU to 100% and drop fps to around 60.
Gaming-wise, I think the 3090 struggles at 8k and 4k 144hz gaming, which is very high end gaming experience and people that spend the money on those monitors will spend the money on the 4090.
For other purposes (like scientific computing/ml), while 3090 can do a lot of things that 4090 does, it's half the speed (or even slower) and a lot less energy efficient, say 1x4090 = 2x3090, 4090 uses 450W, dual 3090 uses 700W, of course you can power limit both of them further, but you still get best bang for your buck out of the 4090.
What games are they running?! While I know not top of the line - I can run the HTC VIVE pro 2 at 5k@120hz at max visual settings in beatsaber on a 2060 based gaming laptop.
Beatsaber is a great game but the graphics aren't very stressful (nor do they need to be).
Keep in mind people are also starting to mod VR into more and more new PC games that were never designed for it. Stuff like Elden Ring, FF7R, the Resident Evil remakes, etc.
$1599 for 4090, $1199 for 4080 16GB, $899 for 4080 12GB (aka 4070). See Nvidia's News page for the announcement.
Edit: Apologies, got the 4080 numbers wrong, accidentally used GBP, corrected above. For the curious, the 4080 16GB is Β£1269, the 4080 12GB is Β£949. Meanwhile, the 4090 is Β£1679.
During the shortage, AMD kept their reference units at MSRP (I was able to pick up a 6800 XT at $649 through sheer luck when others I knew were spending upwards of $1200 on GPUs). With supply loosening up, I expect AMD will continue to play the good guy to steal market share and keep prices sane in comparison to Nvidia's play.
You could buy GPUs directly from AMD, which also had measures in place to ensure no scalpers had the upper hand. This would happen at a certain time of a certain day almost every week, so you could get one if you were patient enough.
Meanwhile NVIDIA stopped production pre-holiday season last year in an attempt to bolster their company value to 1 trillion dollars. Quite a risky play that didn't pay off [1].
AMD outsources it to DigitalRiver, NVIDIA has done that in the past but this time they outsourced to BestBuy (with contractual requirements to sell at MSRP). Both BestBuy and AMD take some nominal measures against bots but it's always been a problem for both sites.
> Meanwhile NVIDIA stopped production pre-holiday season last year in an attempt to bolster their company value to 1 trillion dollars. Quite a risky play that didn't pay off [1].
The source for that article MLID and wasn't borne out by actual production numbers. Actually there was a nominal rise in shipments in holiday '21.
Also, we're currently in a huge glut so obviously there was no availability problem in early Q1 '22 or Q2 '22 either - like, why is pulling back production for 2022 a bad thing, when you're already sitting on a huge stockpile of chips? Were they supposed to keep cranking and make the glut even worse?
People take these techtubers as gospel and a lot of what they say just ends up somewhere between slanted perspective and factually incorrect. But people are willing to believe literally anything as long as it paints NVIDIA in a bad light - just like people went from being absolutely sure that partners were scalping gpus and selling right into mining farms, but then the EVGA CEO says he didn't make any money at all during the mining boom and NVIDIA was capping prices and forcing him to sell at a loss suddenly everyone forgets about the board partners selling $2,000 3080s in their first-party stores. Or the "800-900w TGP" rumors for 4080 that ended up being off by double.
There is just a ton of anti-nvidia sentiment in general and it completely distorts everyone's perception of reality. People will believe any random thing, even if it disagrees with common sense, or their own personal observations, as long as it scratches that "green man bad" itch. And there is a whole little cottage industry of people like MLID and kopite7kimi who thrive on catering to it for clicks. Once those negative frames get started, they don't go away, even if the information used to construct them later turns out to be false.
But people rushing to the defense of poor board partners who say they didn't make any money at all during the crypto boom, heck we actually lost money really takes the cake imo, like that's either actual goldfish-tier memory or massive "green man bad" brain, take your pick. They're the ones who handle sales of all the mining cards too - NVIDIA doesn't run that directly either, it's all partners.
Because the demand for AMD graphics cards has never been all that great. AMD cards languished on the shelves when there was no RTX card to be found. I mean the RX 6650 XT can be found for the price of an RTX 3050. They have been that price for a bit now. If price to performance were a thought to the majority of people buying cards, there would be no way that would be the case. Obviously, gaming performance is not the only consideration. I've been wanting a new graphics card for some time now. I won't pay the Nvidia/miner tax and AMD does not provide one with easy machine learning access.
Not the person you responded to, but rumors, and also, none of AMD's board partners have come out and said they are gonna stop making AMD boards, especially none of the AMD exclusive board partners like EVGA did last week.
I think it's fine honestly - remember we have inflation, semiconductor shortage, and 3090s retailed at the same price like two years ago. Nobody honestly needs this level of hardware for "gaming", like honestly 12-24gb vram? I'm not saying you shouldn't buy it if you want it but it's definitely a luxury product or for non-partnered AI work.
I think this hits the nail its head. Nobody "deserves" a top of the line gaming GPU. And if most gamers can't afford it, game developers will not develop for it as a minimum. Especially for multiplayer games where the total number of players in integral to the game being played at all (which is why many are free to play).
This is not true unfortunately. For high-end 4k gaming you can easily become memory throttled on 8 or 10GB of VRAM. This is the performance cliff you see on cards like the RTX 3070 8GB card. Granted, it's not for everyone, but it's certainly something 4k gamers would want.
Seems like it's higher than it "should" be with GPU mining winding down due to difficulty maintaining profitability. Then again, companies will generally behave with an asymmetric pass-through if they figure it will not impact sales enough by pricing out consumers--prices increase like a rocket, fall like a feather type deal.
I don't know when GPU mining really started to popularize, but there was a significant leap in pricing from the 1080 TI and the 2080 TI. When you roughly account for inflation, it's about on par with current pricing.
Looks like they could trim prices to 1080 TI levels, but I don't see that happening unless sales slow a bit.
Also, I'd like to add, I used general Consumer Price Index to figure this out, so it could vary if inflation did not impact the GPU supply, manufacturing, transport, and other influencing factors similarly.
Is it actually reasonable to be comparing the TI pricing rather than original version pricing? Are they skipping a TI release this series or something? If not then the launch version table looks like this:
Both 3080TI and 4080 are roughly 600mm2+ Die Size. 2080 has a 700mm2+ on a mature node.
1080TI, the only outliner in your example has less than 500mm2 of die size.
And all of that are before taking account of GDDR memory costing, the board, higher TDP requirement and heatsink differences.
While I dont think 4080 pricing is good value or cheap by any standard. I also dont think it is ridiculously expensive. From a transistor stand point, you are getting 2.7x the transistor between 3080Ti and 4090. With 4090 You are buying 76 billion transistor, with 24GB of GDDRX for only $1600.
$/transistor could be the biggest variable here. My understanding is that we haven't been seeing this lower even with node shrinks. Analysts have speculated that this is why Xbox released a trimmed down console (series s)- because a cheaper "slim" wasn't realistic. Similarly Sony raised prices because inflation is outpacing their ability to reduce costs.
>$/transistor could be the biggest variable here. My understanding is that we haven't been seeing this lower even with node shrinks.
Purely from a Wafer, Transistor per Die Space hence Cost per transistor, we are still getting it cheaper per node generation. The ratio, cost per transistor may not be dropping "as much" as what we were used to. But it is still dropping.
However the Initial cost for product development on new node is increasing. This fixed cost, ( think CAPEX ) amortised over "fixed volume" would mean the cost per "die" would also be increasing.
So once you add those together you will see we are at the end of the power curve.
Is it actually reasonable to be comparing the TI pricing rather than original version pricing? Are they skipping a TI release this series or something? If not then the launch version table looks like this:
- 1080 (2016) = $599
- 2080 (2018) = $699
- 3080 (2020) = $699
- 4080 12GB (2022) = $899
- 4080 16GB (2022) = $1199
Doubling the price of the flagship version of the launch product over just 6 years is well, a pretty large jump.
That's non-sense. I omitted inflation data because the parent post had that for similar numbers and did not list the source used. My attempts gave significantly different numbers, so if I gave inflation adjusted numbers, it would be a misleading comparison vs the numbers in the parent's post.
My experience buying my last GPU (1080ti) leads me to fill in the blanks. GPU mining became popular shortly after the 10xx was released. I know this because I couldn't find one in stock and even when I did, it was much more expensive than the launch price.
> Also note that nothing is preventing Optical Flow Acceleration [0] (and subsequently the DLSS 3.0 models that they claim are exclusive to the 40 series) from running on either 2/3 RTX cards.
Are you sure about that?
The page here says it uses "Optical Flow Accelerator of the NVIDIA Ada Lovelace architecture" which sounds like a hardware feature?
You're right, at least for me. Been waiting for a reason to switch, given that 4G decoding capability is available on my AM4 mobo that would give this a further performance boost and available VRAM.
NVidia can suck it and go solely target cloud computing - they already dedicated a significant portion of the announcement to just that. Why didn't the fact A) this is extremely lucrative and B) they dominate there already reduce their price for regular consumers? Corporate greed.
If I see NVidia cards start shipping in consoles, I'll know they decided to rip us off so they could get great deals for the same hardware to manufacturers. A total middle finger to us PC gamers.
I'm getting an AMD vs Intel vibe like I did from years back. I switched then, I'll switch again.
eBay. Make sure the seller has a long history of good reviews and is in your country/region. There are some sellers with over 100 cards theyβre trying to offload.
I think itβs great the high end keeps getting bigger and faster.
Nvidia has primarily been associated with gaming. With Machine Learning thatβs starting to change. With tools like Stable Diffusion every single artist is going to want a $1600 GPU and honestly thatβs a bargain for them!
I think itβd be sweet if you could spend $100 to $10,000 and get a linear increases in cores, ram, and performance. Thatβs not how it works. But philosophically I have no problem with it.
Hopefully AMD steps up and brings some real competition.
>With tools like Stable Diffusion every single artist is going to want a $1600 GPU and honestly thatβs a bargain for them!
Most artists who are not tech savvy to spin up their local Stable diffusion instances are probably going to pay and use one of the online services like DALL-E 2 or something.
Itβs already been turned into single-click installers and Photoshop plugins.
Packaging Python is a dumpster fire of local environment complexity. But this is an extraordinarily solvable problem. I wonβt call it solved today. But we have Proof of Life of solutions.
I mean shipping people a .zip that has a simple executable that can be reliably run by end users zero external environment variable bullshit or dependencies. No Anaconda, Pip, or any of that non-sense. No fetching of dependencies either. Dead simple turn key usage.
And yeah itβs not super duper hard. Itβs just not how Python projects are structured by default. Simple Diffusion isnβt setup that way. Some random person on the internet spent some time to package it up nicely and in exchange they hit #1 on HN. Possibly more than once!
The 3090s were more than the 4090s, and there's been a great deal of inflation in that time. I won't call them cheap, but it seems like a reasonable price trajectory?
For what it's worth on October 2 2020 I was lucky enough to get a 3080 for a retail price of 699.99. (This was only possible due to using a bot for personal use since Covid may everything nuts). Adjusted for inflation this would be around $800 today. There's two more months until launch so the 4080 card is maybe 80 to 90 dollars above inflation.
my thoughts too. As someone that has been buying generations of video cards from matrox, riva tnt and onward. This doesn't exactly seem off to me, I hope AMD brings some competition with RDNA3 this time around.
For most of that time, the Euro was worth more than USD. The exchange has been dancing around the parity line since June. It may still be roughly the same price in USD->EUR.
Sure, but the salaries in the EU didn't go up 20% in the same time so Nvidia is clearly not targeting the same segment of the population anymore, but a 20% richer segment now. Companies adjust pricing to different markets all the time, but for some reason Nvidia didn't do it this time and EU consumers are much more price sensitive than US ones, so EU will likely turn red or blue in the near future and green will lose its hold. Also with crypto gone and cost of living increasing sharply I think this is a huge miscalculation on Nvidia's side and this will burn them.
NVIDIA not giving you a 20% discount on their hardware to adjust for currency exchange rates isn't really the scandal you think it is. I really really wish during the last 20 years when the euro was higher that I would have gotten a "USD discount" on all my purchases from the EU, but that's not how it works.
Nvidia's manufacturing is in Taiwan. US has traditionally much lower prices on electronics than the EU, now with this change EU is completely priced out. SWengs don't earn 800k there, but typically plateau at around 100k at late stages of their careers. This will have a major impact on Nvidia's ability to service EU market. If you tried to buy something from EU in the past, you probably had to pay large import duties which likely skewed your perception.
People buying expensive GPUs never really struggled with increased cost of living. For the budget sector the difference between a 300 Euro card and a 500 Euro card is significant but for the enthusiasts or business owners it doesn't matter if it's 1800 euros instead of 1500 Euros, they'll still buy it.
Nvidia wants to be the Porsche of cards, not the VW, so they're pricing themselves accordingly.
I am in their "Porsche buyer" category as I bought 1x RTX A6000, 1x Titan RTX and 4x RTX 3090 for Deep Learning, but I am staying away from 4090 until they get some sense with pricing.
From my point of view, 2x NVLinked 3090s give me the same benefit as one 4090 at 150W higher power consumption and nothing is super appealing about 4090 until drivers on Linux and PyTorch/TensorFlow implementation get stable enough for production use which took like 6 months last time, so there is no pressure in buying it.
Why do you need a 4xxx for PC gaming? A 1070/1080 is still a very decent card. There's no game released that won't run at reasonable res/fps on 1xxx series cards.
Agree: the PC gaming market size[1] continues to grow and is definitely very very alive.
I am guessing the dead rumour is because the console market is growing faster. Conveniently ignoring that the mobile market is growing faster than the console market (mobile has 50% of market value by some analyses).
Conversely, AAA games are mostly not exclusive to PC keyboard/mouse control (games market could be split by input devices: mobile touchscreen, console controller, PC keyboard + mouse).
> competing with the flood of miner cards entering the market.
Remember: don't buy miner cards and let them rot. They artificially generated scarcity thus raising the prices for everybody during the pandemic. And don't believe that a card that has been on 24/7 is safe to use. "but it was undervolted!" means nothing.
Just let them become e-waste and buy something new. Turn on the stove and open all Windows when not at home while you're at it. SMH.
And that "not safe to use"? You think it's gonna explode? Worst case it doesn't work, but even that's probably pretty low risk. Hardware doesn't "wear" that way that quickly.
Exactly. The thing that stresses silicon the most is extreme change of temperature. Guess what doesn't change during mining: Temperature.
When mining, you want a steady rate and will most likely downclock and undervolt it in order to get maximum efficiency out of your card.
The only thing I could imagine to break are capacitors, but I still have to run into anyone who has issues like that.
GPU mining is poised to be permanently done. There is no "lesson" to teach miners, because there is no expected repeat of such a run on GPU mining again.
Meanwhile, the implication from your post is that it is better that the tens or hundreds of thousands of GPUs miners currently own and may offload would be better in a dump, just for retribution.
Nicely said. Actually there is a point - vengeance and moralizing. But you're right this probably wouldn't have much effect on GPU availability and reduction of pollution in the near future.
His argument is that by buying a miner card, you're validating strategy of miners, which are considered "bad actors" by gamers and environmentalists. If you intentionally decline buying a miner card, you're helping to make miner activity less profitable, thus doing the "right thing".
>Why don't you go to a beach and scoop up some oil from the BP spill?
If it was trivial, why not? If you told him "don't buy any GPUs" then it would make sense. But either way he will be using some "oil" (GPU) in this case so I don't see why he should waste perfectly good "oil." I suppose you would rather leave the oil in the ocean?
I have never understood this argument. How much environmental damage is done by the mere creation of all television shows cumulatively, the creation of plastic tat given away at conventions, plastic wrapping at grocery stores?
But graphics cards are what we REALLY need to focus on and you better believe you are tuning in for the next big show. It's almost manufactured.
I would actually be quite interested in seeing some numbers to back that up, but I'm not sure if there have been studies on the damage of mining yet.
In some ways, letting a GPU run at a constant load (and thus constant temperature) is less damaging than a typical gaming workload I imagine, where the temperature fluctuates often. In other ways it probably isn't (e.g. atoms of the chip 'migrating' over time with the constant electron flow, though I have no idea how much of an issue that is with modern chips).
> I would actually be quite interested in seeing some numbers to back that up, but I'm not sure if there have been studies on the damage of mining yet.
The few investigations have generally found mining cards are perfectly fine and if anything above-average in used quality. Because yes the primary damage is from thermal cycles, not just being on. And since mining is a constant load & constant environment, it has very few thermal cycles. And miners primarily care about efficiency, so they're power-limiting them[1] which further reduces strain on the cards.
Fans might be closer to death than a typical used card would be, but on pretty much any GPU those are cheap & easy to replace. That's really the only "risk"
1: And before someone inevitable scoffs at this, it's common enough that this is built into tools like nicehash. It's literally "click 1 button" territory, not exactly something a miner wouldn't attempt.
Do you have any idea about where these were kept? Not all mining operations are commercial-like white rooms with people going around in sealed suits. Maybe someone had half a dozen cards in his chicken coop, exposed to dust and moisture and energy fluctuations. You will never know. Would you happily buy that at say, 50% MSRP?
Yup, assuming it works and checks out the first 24 hours I have it.
Solid state electronics are pretty robust. If they survive whatever torture you put them through plus being shipped to my door, chances are it will work just fine. In my experience there are some problematic GPU models that are more prone to early failure under heavy use (HD6990 I'm looking at you!), but they are more rare than legend may lead folks to believe.
At an appropriate discount it's far cheaper to just buy these at pennies on the dollar and just be alright with self-insuring for a rather rare early failure. When you can buy 2 or 3 for the price of new, it starts to get rather compelling.
As long as there isn't any corrosion that will lead to premature failure, I don't care. Dust is easy to clean off and I would expect anyone trying to sell a $500+ GPU to at least try to clean it up a bit. Even if not, I have an air compressor and a blow gun. "Energy fluctuations"...what? Power supplies and VRMs are built to deal with this. Check your home with a power analyzer with a high sampling rate and see what happens when your AC kicks on. At the same time, measure the 12v rail going into your GPU when that's happening.
How do you tell there's no corrosion though? You need to at the very least take off any coolers on the card and look at the PCB itself to have some level of confidence that there's no damage, and even then I've seen corrosion sneak under components where you need to pay really close attention to notice it.
If I can inspect a card in person, I would be plenty happy to buy an ex-mining card, but most of these are going to be sold online sight-unseen. This isn't a problem exclusive to mining cards though; people have been selling damaged GPUs as 'working' for years.
> Are you claiming that it is somehow unsafe to use?
For your wallet, yes. Do you have any information about the environment in which a card was used? Temperature, dust, air moisture were like in a home or there operated in an industrial warehouse? Fan bearing status after all these hours? Voltage applied? Stock BIOS? were they in an enclosure or in a sort of a case?
Thermal paste has been changed? Ever been disassembled?
I would not trust a miner's card even if it was a gift.
Miners screwed the market over and now they want you to think that their used crap is such a bargain. Let them deal with their waste, don't help them.
> Do you have any information about the environment in which a card was used?
That would be even more true for cards owned by non-miners; most miners will operate at a large enough scale to use some best practices, which definitely isn't true about most people who buy GPUs for personal use.
This is but one test, but Linus Media Group did some testing and found that mining GPUs worked just fine. Fan time and thermal pad degradation are about the only real detractors from buying a used GPU. https://www.youtube.com/watch?v=hKqVvXTanzI
Of course, if a card's running at thermal throttle for long periods of time, that has a chance to damage the card, but given that most cards are running underclocked, the likelihood that you'll run into that problem is low.
https://maggieappleton.com/ for her take on a digital garden. She also compiled more below [0] if opening your thought process to the public is your thing (adopting the concept actually got me to add more structure to my serious notes and actually post more)
My overall advice is unless you're a designer flexing your skills is to focus less on the design and more on the actual writing! The internet is littered with the emaciated husks of nice good looking sites hosting nary an entry beyond "How I Made My New Blog With X"
The illustrations on Maggie's website are gorgeous.
I've been focused on content for 5 years (for my website that pays the bills), but the website is so bland. I'd love if people could go "nice" when they use the website.
It seems like Maggie nailed that. Thanks for sharing.
I mean your "data" will be particularly skewed towards one end of the argument seeing as you deal with people with millions and are literally incentivized to note it as a bigger problem than it is. Differences in housing and job availability are bigger differentiating factors when "normal" people decide to move.
I mean if you go through them yourself and you'll see most (if not all) of those are closed. And the majority of the others dealt with composability issues between custom packages/libraries and not base Julia. Honestly it's a tooling + linter problem, and the lack of progress there (plus the general adversarial/celebratory tenor of that thread) made me pretty glad I've just moved onto Rust + python again. It's just too demoralizing from a dev's perspective to see.
Don't get me wrong, I love multiple dispatch and Julia's composability + speed make it a no brainer for science/quant econ teams to adopt (and using it to train models on TPUs is a actually a sneaky way to save time + $), but they've lost the "culture" war and I can't see the entrenched masses here or elsewhere changing their minds on it. Not sure whether it's for job security sake or they got sick of all the "Julia is faster than/will replace X" evangelism or Plots loaded too slowly the one time they tried it years ago or a combination.
No, because of lack of progress in areas I care about. It's just a tool that currently doesn't fit my most crucial needs, so I picked up other tools.
The discussion was further evidence of the reality that I'd already come to terms with; the lack of inroads in building a larger community and general ecosystem is partly because of a lack of focus on things like tooling, linting, documentation standards, etc and partly because of a weird amount of disdain for the language. I can help with the former, but I don't have the time or energy for the latter because it's actually a set of large, long-term marketing, organizational, and psychological problems to overcome. Maybe someday down the road I'll be back, but I've given it a shot for about 8-9 years now.
Love that more of these little search alternatives are cropping up. Try suggesting search queries that do perform well on the page or as a watermark in the search box if you're currently that limited. One thing that definitely turned me off were the silly graphics over the picture under the privacy pledge; it makes it look like a joke when it should be serious because people who care about it really care.
