Microbenchmarking AMD’s RDNA 3 Graphics Architecture
chipsandcheese.com
chipsandcheese.com
A few observations.
First, the astonishingly low latency for LDS (aka workgroup shared memory). Vello uses this fairly extensively (especially in prefix sum / monoid scan operations, plus the stack monoid), so I'd expect performance to be quite sweet on this card. If every card had this, I'm not sure there'd still be a motivation to do subgroups. I also really liked the chart comparing shared memory latency across cards - I already knew that Intel was slow (especially Gen9 and earlier), so it was nice to see quantitative data.
Second, it seems really odd to me that there's chip area for dual-issue ALU, giving over 60 TFLOPS of f32, but the shader compiler only achieves single issue most of the time. I wonder if there are plans to improve performance over time through software updates.
I'd also be really curious whether using Vulkan subgroup size control would help unlock this higher performance. With that feature, the application can query what subgroup sizes are available (32 and 64 here) and explicitly choose one.
I'm really glad this kind of microbenchmarking is being done, we need more of it.
What is Vello?
I'm guessing you mean this: https://github.com/linebender/vello
The AMD RX 7900 XTX is less than half the price of the NVidia RTX 4090 ($999 vs $2300)
EDIT - correction, the MSRP for the 4090 is only $1600
Right now Nvidia is playing a game where the MSRP cards can't be sold at a profit, so you'll only see a token amount of them exist before going out of stock to never return. Hence why even the 4080, which by all accounts has barely sold at all, still has a bunch of out of stock SKUs - coincidentally all the ones that are "MSRP." Realistically the 4090 then also is more like a $1800 card.
Still i think, even with these crazy prices, the price per performance increased compared to the previous generation. The 4090 is a very good card.
There goes the time of the 700 Euros high end models.
here's newegg right now:
https://i.imgur.com/al8N35W.png
stop listening to a bunch of rabble-rousing techtubers lol, the pitchfork brigade isn't always right.
You're kinda proving the point here while framing it as a counter argument lol
Yes, premium card models exist and have always existed, for people who want to pay extra to get a quieter cooler or another few % OC performance.
That's different from "MSRP is fake because those cards don't exist and nothing is available at MSRP".
10-series MSRP is an example of a fake MSRP - FE was priced above "MSRP" and all the partner cards slotted in above FE. 4070 Ti is nothing like that, you can go buy a 4070 Ti for MSRP if you want, they definitely exist, and will likely continue to exist.
I checked the listing for 4080s and there is, unsurprisingly, currently a model available for MSRP there as well. It's backordered until tuesday (a whole 36 hours!), but, I thought you said they never restock after the initial drop?
https://www.newegg.com/pny-geforce-rtx-4080-vcg408016tfxxpb1...
Like yes you're not wrong that partners would rather make the high-margin cards and they do everything in their power to avoid selling at MSRP, but, that's because they're profit-seeking middlemen with a stranglehold on the supply of a scarce resource, not because they can't afford to make cards at MSRP.
Just like Zotac and EVGA - PNY isn't losing money on those cards either. They may not like what they're making, they'd rather sell you a corvette than an econobox, but they're not losing money either. That part is not true and is as hyperbolic as your inventory claims: they still make money, they just don't make as much as they want.
Like yup absolutely the margin is worse on an Accord than a sports car. Agreed. The dealer still isn't losing money on the econobox at MSRP, no matter how much they wail and gnash about it, and it's not good for consumers to just increase the price so middlemen make a fat margin on basic-tier shit.
People generally look back on car dealers scalping Hondas in the 80s (and cars/trucks during the pandemic) as being a bad thing, but, somehow the board partners have gotten consumers into this idea that margins for middlemen are a good thing when it's GPUs.
https://www.wsj.com/market-data/quotes/TW/ROCO/6150/financia...
https://www.igorslab.de/en/evga-pulls-the-plug-with-loud-ban...
Some people really do be seeing EVGA selling a 2080 Ti for $949 and Zotac for $999 and then argue the other partners magically can’t afford to sell for less than $1400. Or think that anything has changed since then.
We call them… gamers.
It’s honestly crazy to see techies embracing the digital-car-dealership model over the Tesla model. You really really want to fight against scalper-bots instead of just putting in an order on Steam for a 3060 Ti at MSRP and having it come in a month or two?
And then everyone acts like rma is only possible with partners… bruh I can rma a cpu to AMD or Intel just fine without a middleman involved.
Unironically you are buying into smoothtalk from brylcreemed CEOs who were scalping 3080s in their first-party webstores for triple MSRP and even scalping directly to Ebay (lol MSI) literally weeks before their media tour, they came crying directly after prices crashed as part of a media push to get NVIDIA to refund their excess mining orders. Gamers are the most credulous rubes on the planet, NVIDIA and EVGA can agree on that much.
https://www.reddit.com/r/hardware/comments/j6idky/msi_scalpi...
and conversely - partners love what NVIDIA is doing with the 30-series price fixing right now! Their margins on 30-series are great, they're not losing money on deep discounts like they are with RDNA2 products. That's not good for consumers, but hey, somehow they've convinced consumers they're on their side instead of a middleman with their hand out.
Again, as you can see from the 2080 Ti example, MSRP isn't "losing money" territory, in fact they can probably even go a bit under MSRP (EVGA knocked another 5% off), they just don't like it, at 5% under MSRP EVGA was probably running actual zero margin. It's when prices go 30-50% under MSRP that they lose money. And NVIDIA is doing a great job with 30-series keeping prices at or above MSRPs and making sure nobody breaks ranks and cuts prices. Trusts are super effective at maintaining margins, and consumers love making sure partners get their margins, right?
So.. during a year of highly inflated GPU prices, GPU vendors who sold everything far above MSRP made a bunch of money?
No shit, Sherlock. What does that have to do with Nvidia's well known, well sourced MSRP games?
And since you're so focused on the 30xx series, reminder that the MSRP of the 3070 was $500 and even now at the end of the product life it's still slim pickings below $600. 3080's MSRP was also $700. But again, even now at EOL there's very little at that price point. The major brands are all $900+
I provided actual sources and numbers here, "well-known, well-sourced" is just "trust me bro" if you don't provide those sources.
Again: 2080 Ti is a great example of how those claims aren't really true. There was nothing stopping the other partners from offering MSRP cards just like EVGA and Zotac. They could even have gone under MSRP, just like EVGA.
Partners won't do it because they don't make enough money, not because they lose money. EVGA and Zotac show they're not losing money doing it, they just don't want to.
It's car dealerships whining they're only making 5% on a budget car instead of 15% on a corvette.
So... No, other companies can't do what EVGA did just like EVGA couldn't keep doing what EVGA did. Hence why they basically went bankrupt doing it (rather, they obviously didn't drop 70+% of their revenue if it was a healthy profit margin as you are so insistent on claiming)
Anyway, I'd still love to see your sources, but, if you can't provide any then I think we're done here. I gave you real numbers and you've ignored them several times in favor of supposition and imagination.
EVGA quit because they knew the mining crash was going to lead to a couple of very slow years, just like after the 2018 mining crash. I've already stated my problems with their argument: they made tons of money during the mining boom, and the claims that they were losing money were limited to the deep clearance pricing they had to do after the mining crash.
You have acknowledged most of these individual points but just don't seem to accept the synthesis of the whole: the EVGA guy wasn't really presenting a fair and honest big picture, that's why GN told you to "take him with a big grain of salt" and Igor's Lab's partners said they didn't understand where he was coming from.
This may be difficult, but sometimes individuals will go to media and present a slanted narrative, especially when they have a financial interest in doing so or are generally upset. What the EVGA guy said wasn't really true, or was very narrowly true.
And to be fair, there is a lot to be upset about, NVIDIA definitely is not gracious to the partners. But if they play it right, it's a guaranteed 10% margin on all their sales. 200k cards a week times $500 ASP times 10% margin is quite a bit of money, and that's EVGA alone.
Do not seem to be practical training runs.
The number of enthusiasts is at an all time high, but the number of in depth tech articles seems to be at an all time low. Everything is just high traffic cursory reviews that often border on clickbait.
Leaving the content completely aside, the future of content production is MLID-style stuff, he's been very successful at commercializing his work via the patron model. Some focus videos, some weekly news videos, some viewer "call-in" (chat, discord, etc) interaction videos. You can see Ian Cutress and others moving towards this somewhat.
Greg's Airplanes and Automobiles (where my greggheads at?) is one of the few who I'd say produces content that is comparable to long-form written content, and honestly he suffers from the bloat problem. His series on the P-47 is eight parts and none of the parts is less than 40 minutes a video, several are almost 90 minutes. And I doubt he makes anywhere near as much as the more "interactive" and broad/shallow content channels from that ten hours of content. And there's probably another 10-20 hours of research behind every hour of content... which is much lower for shallow content as well.
And there is zero chance that written content would make even as much as Greg makes. People aren't willing to pay for it. That's why written tech media (and deep dives) is dying and you get 30 minute GN reviews instead of a 3 page article.
I guess what I'm saying is click that donate button on Chips+Cheese's site. Not that I'm any better about it.
Well that's awesome. Learned a new thing.