Home
Creators
Search Recent Random
Posts
Search DMs Popular Hash Lookup Tags Random
Importer
Import FAQ
Register Favorites Keys Review DMs Logout
Community
Telegram ThePornDude
Home Creators Posts Import Register Favorites Logout

Quantifying what we want from RDNA 4 & Blackwell (Die Shrink Telegrams) (Patreon)

Published:
2024-03-21 20:01:38
Imported:
2026-06
Tags:

Content

Talking to Tim about what value we want to see out of next gen GPUs got us thinking - what do we want to see out of RDNA 4 and Blackwell pricing to make us happy at the end of this year?

Tom & Dan are going to sit down and talk about what they could see these products end up as:

RTX 5090

RTX 5080

RTX 5070 RX 8800 XT

RTX 5060 Ti RX 8700 XT

RTX 5060 RX 8600 XT

RX 8600

RX 8500 XT

...and try to be realistic, and give a best case and worst case scenario. What do we think is possible? What would make us happy? What's a realistic upset scenario for some cards?

Write in below your thoughts about what value we should expect, hope for, and be worried about when it comes to next gen GPUs. Do try to keep your submissions focused and with a real point though. But honestly, anything having to do with next gen performance and competition is fair game for this one! You have a few hours from posting to submit, don't wait!

Comments

CompressedAIBlocks

If top RDNA 4 (8800 XT) is 7900 XT levels of performance in raster, where would you guys place the 8700 XT in terms of raster?

CompressedAIBlocks

How boring will this generation be if VRAM stagnates again for you guys? As an industry we really need 16GB VRAM to be accessible at mainstream price points ($200-300) if we expect to use newer engines and features.

CompressedAIBlocks

What are you guys expecting from the "8500 XT"? The 50 class cards have not been very positively received in quite some time wether due to performance, price or both. Is there a purpose in this level of performance on a card using a new architecture/node that would surely cost more to produce?

JonWickBoomStik

I think if RDNA 4 can provide between 7900XT and 7900XTX raster performance (although more towards 7900XTX) and double the RT performance for $500 they will have a clear winner in the eyes of most gamers out there paying a reasonable price for really good performance. This generation really needs to bring back value for gamers. Especially compared to the current gen. I don't have much confidence Nvidia will give us better price/performance. Given what we've seen they will be charging top dollar for their cards and focusing more on AI sales as it yields more money for their shareholders. As all things time will tell, but without knowing what Nvidia will charge for cards. Intel I'm not too sure about them but I really hope they can show up with good drivers and they have shown they can provide good value once their drivers are good. my hope is with AMD that will provide something great.

Falto

I think from the green side, if they maintain current prices, give enough vram, maintain power usage and improve performance about 50% per tier. That is honestly all we need this next gen for them to do well.

Joseph Kelly

As an AI engineer I'd like to see an absolute minimum of 48gb of VRAM on the 5090 and 96GB of VRAM on the RTX A6000 Workstation Blackwell card. I think gamers deserve VRAM upgrades on all tiers of NVIDIA's 50 series as a whole because games overall will be better for it. AMD's 8800XT should also have 48gb of VRAM too but I think AMD has already failed by cutting their top die regardless or performance but that's because I'm looking for AI cards more than gaming cards and can't afford a $10,000 H100 or A100 card.

Aluveitie

Costs of fabs going up for each generation, and power/frequency stopped scaling as they used to. They buy performance gains with die area, power usage has been rising across all product lines. With Blackwell development reported to cost over $10 billions, I don’t see a way back to price levels and gen over gen performance gain we’re used to. It has already slowed down massively over the last decade… Did you get a general sentiment from engineers regarding Moores Law stopping to hold up?

Jeffrey Reyes

I was highly disappointed to read that Blackwell, which was previously expected to utilize TSMC 3nm, is actually going to use the 4N (5nm) node already being utilized for Lovelace. 1. What could be the reasons for this, does Apple have 3nm on lockdown, or was this actually a deliberate decision on NVIDIA’s part? 2. What kind of performance gains do we expect now what we know it’s just architectural improvements?

Kinihun25

Hey Tom and Dan, I currently have a 4090 and the card is a dream. Does everything I need and more. The only way I can see myself buying a 5090 or 5080. The 5090 has to have a minimum of 32GB of VRAM anything less I don't care, performance means nothing because I can pretty much play every game I want a 4k 120 max setting and it renders my animations at 4k 60fps with no problems, I would prefer 48GB of VRAM because at this Point Nvidia should market their 90 class card as a Prosumer card that is the perfect in between Professional and Consumer, anything less than 32GB of ram won't cut it.,

Manuel Nascimento

After the PS5 Pro leaks, and in particular the PSSR reveal, what do you expect this will mean for ML upscaling for RDNA 4? Will AMD provide a solution that is close to DLSS and make it so DLSS isn’t as much as a selling point as it has been up ‘till now? Cheers and stay frosty :)

Uthleber

5090 (168 SM, 32 GB) 2499 $, 5080 999 $, 5070 599 $, 8800 XT 549 $

AdanFS

I don't really have any expectations for Blackwell, but they really need to give 16GB to more cards. * Best case scenario for Radeon in my opinion is an efficient 16GB 7900xt (Raster) with closer to 4070 Ti Super RT priced as close to $500 as they can. * Worst case probably a little more efficient 16GB 7900xt (No RT improvement) for $600. Most Likely IMO: a more efficient 16GB 7900xt with 7900 XTX RT for $550

Prior is the theme 999

They should just give us xtx/4080super performance for around $600 with 20gb of vram and 4070/7800xt performance with 16gb of vram for $300 and call it a day.

Swiggles

I want top amd to be godly price to performance so I can ignore it through my tears as i pay for my 5090. My first full priced purchase of a top end card ever.

Razgriz III

I can only really see myself going as far up the stack as the potential 8800XT… If either can offer somewhere around 4080 4K (between 7900XT and XTX) performance at ~$680 or less, with anything more than 16GB of VRAM it would be pretty compelling for me as an upgrade from a 3090. If nothing offers that, I’m assuming the 5080 will be the only card above both that performance and vram. It’ll also probably be above $1,000. I’m not willing to break that barrier on a GPU this gen, so at that point I’d be hoping for any flash sales/hunting the used market for an XTX below $700 anyway. On the low end, I really really want to see a low profile or at least single-fan card with 12+GB return to the market. Just some kind of straight upgrade from a 3060 12GB since both the 40 series and 7000 series refused to give us that. Don’t care if it’s the 5060 or 8600XT. Just +30% performance uplift at $300 and at least maintaining perf/watt… though preferably more efficient too. I don’t think I’d be able to buy both if they both happened but I’d be very happy if either happened.

JaRail

For me, the 4090 I've got does everything I need in terms of gaming. Well, maybe a little slow on the path traced extreme examples. What I care about is vram. I do a lot of AI and 24gb is just not enough for LLMs. Even with a 4090, I end up renting cloud instances with beefier cards. I'd upgrade for 32gb but it'd only be 48 that would get me excited. That said, for LLMs, nvidia is now competing against macbook pros. The unified memory means you can get one of those with 128gb and use that for GPU inference. nVidia *really* needs to offer higher vram solutions soon if they want to stay competitive with Apple in those enthusiast/developer use cases. I'd say $1600 again for 24GB, $2000 for 32GB, $2500 for 48GB.

TechnoLadz

I would say the 8800 XT would end up competing against the 5060 Ti, not the 5070. Best case for all: Also performance, VRAM and pricing wise, I would say the following: NVIDIA: 5090: 48GB VRAM, 70% greater than 4090, $1999 - 500W 5080 Ti: 36GB VRAM, 50% greater than 4090, $1499 - 450W 5080: 24GB VRAM, 20% faster than 4090, $1049 - 350W 5070 Ti High RAM: 24GB RAM, 5% slower than 4090, $849 - 285W 5070 Ti Low RAM: 18GB RAM, 5% slower than 4090, $799 - 285W 5070: 18GB RAM, 7% faster than 4080S, $649 - 230W 5060 Ti High RAM: 18GB RAM, 10% slower than 4080, $499 - 165W 5060 Ti Low RAM: 12GB RAM, 15% slower than 4080, $449 - 165W 5060: 12GB RAM, equal with 4070 Ti, $349 - 135W 5050: 12GB RAM, equal with 7700 XT, $249 - 75W AMD: 8800 XTX: 32GB RAM, equal with the 4080 - launch at $649 - reduction to $599 at 5060 Ti launch - 225W 8800 XT: 16GB RAM, 5% slower than the 4080 - launch at $549 - reduction to $449 at 5060 Ti launch - 190W 8700 XT: 16GB RAM, between 4070 Ti and 4070 Ti Super launch at $449 - reduction to $349 at 5060 launch - 180W 8600 XT: 16GB RAM, between 7800 XT and 4070 Super - launch at $349 - reduction to $279 at 5060 launch - 160W 8600: 16GB RAM, between 7700 XT and 7800 XT launch at $299 - reduction to $229 at 5060 launch - 140W 8500 XT: 8GB RAM, equal with the 6600 XT - launches at $179 - reduction to $149 at 5050 launch - 75W Worst case: Nvidia: RTX 5090: 24GB VRAM, 40% faster than 4090, $1999 - 650W RTX 5080: 16GB VRAM, 10% slower than 4090, $1099 - 350W RTX 5070 Ti: 16GB VRAM, 5% slower than 4080, $799 - 285W RTX 5070: 18GB VRAM, equal with 4070 Ti Super, $649 - 240W RTX 5060 Ti: 12GB VRAM, equal with 4070, $449 - 220W RTX 5060: 8GB VRAM, equal with 7700 XT, $349 - 180W RTX 5050: 6GB VRAM, equal with 6700 XT, $279 - 140W AMD: RX 8800 XTX: 32GB VRAM, equal with 4080, $999 - 350W RX 8800 XT: 16GB VRAM, 10% slower than a 4080, $649 - 280W RX 8700 XT: 12GB VRAM, equal with 4070 Super, $449 - 240W RX 8600 XT: 16GB VRAM, equal with 4070, $399 - 180W RX 8600: 8GB VRAM, equal with 4060 Ti, $349 - 150W RX 8500, equal with 3050, 8GB VRAM, $199 - 75W

QuickJumper

AMD should focus on value. Let's be honest this is what AMD can do pretty well. Even if they are bit behind in software as long as they focus on great value they will be successful with RDNA4.

Roman Matúš Labuš (edited)

Comment edits

2026-09-05 20:00 Tom Jen-Sun Huang says that the R&D for Blackwell was around 10 billion. How the hell did he get to that number? Can it be really so expensive?
2024-03-21 21:44

Woody Chang

Hi Tom and Dan, Since the RX 8000 series won't have top end dies, they need to compete against nVidia with features. For example, it won't matter that much what rDNA4 competes against in terms of raster if FSR quality continues to suck even compared to even XeSS. I mean, what does it say that Sony is making a whole new upscaling technology if not to say that FSR isn't good enough? Pie in the sky dream- AMD finally finishes the machine learning FSR with dedicated hardware to back it up. Worst case, they will do nothing to improve FSR and hope that people won't notice all the artifacts.

Woody Chang

We also need to separate out ray tracing performance from overall performance. For example, let's say the RX 8800XT is equivalent to the RTX 5070 Ti in raster. Historically, the ray tracing performance for AMD seem to be equivalent to 1-2 tiers below where their nVidia raster equivalent is, so RX 8800XT ray tracing will perform more like a 5070 or 5060 Ti. Perhaps if they can make it so that the ray tracing in AMD only causes performance to fall to about a half tier below the comparable raster equivalent nVidia card, then I would count it as a win.

Chris Rijk

The conventional wisdom is that we need AMD to put pressure on Nvidia to reduce prices. If anything I think it’s the other way around. Nvidia only seems to care about selling enough at a high margin and I suspect that they’d rather lose a lot of market share than reduce their margins. AMD responds much more conventionally to competitive pressure and will definitely reduce prices in response. In short, I don’t expect Navi 48 GPUs to be aggressively priced at launch due to a lack of competition and instead AMD will reduce prices over time only as much as they need to and if Nvidia does show up with truly competitive prices on paper I suspect we’ll be back to fake MSRPs.

Backcut

Let's revisit the MLID Radeon 30% Rule For Success as it applies to N48. Given the expectations for Blackwell rollout, do you think it's likely that RDNA4 cards will be competing against their Lovelace peers well into 2025? If N48 matches a 4080 in raster and comes close in RT, it's hard to imagine AMD charging less than $700 as long as the 4080S remains at $1k. If N48 comes out closer to 7900XT in raster with less-than-amazing RT improvements, $600 still seems like a reasonable prediction. I'm struggling to imagine N48 pricing anywhere under $600 unless performance is just disappointing or 4080s price drops (not happening).

Backcut

It seems like Nvidia's lesson from Lovelace was that if the flagship card is uncontested you can charge whatever you want for it, but nobody wants to pay over $1000 for an 80 series card. I don't think any of us will be surprised to see a 5090 that kills it priced at $2500 and it'll probably sell, but do you think Nvidia will take the lesson to keep the 5080 at or under $1k, or will they try for $1200 again (or more)?

Chris Rijk

I fear Nvidia will be stingy on improving cost-per-frame - perhaps around 10%. I think they can get away with that. However, I don’t think they’ll be able to get away with being stingy with VRAM on lower cost models.

zombykilr777

I am hoping that the 8800XT will have higher than 16gb of vram to curb the performance difference from the 5070. depending on vram for 5070, this could be the better buy for future-proofing with the current trend of higher detailed graphics

OmegaDoge

If the RX 8800XT is projected to compete with the 5070, should it be called the RX 8700XT instead so AMD can be seen as being better competition tier-for-tier?

Justin John

Yes but for the price point of 599 which the 4070 debut at. Let’s call it a range of 550-650. Don’t think someone in that price range needs more than 16 gbs for resolution they’d be playing at. Rather have it cheaper with 16 than cost more Obvs nvidia wants to skimp on ram.

Jen-Hsun Huang, AI is not circular

It probably includes all the R&D that went into all the parts of their “GB200 NVL72” Aka, Blackwell, Grace, NVLINK, other connectivity stuff, and more. Plus non-AI Blackwell R&D probably got lumped in because “it’s all Blackwell!”

Dave Scholze

5090: at least 32GB of ram and 40% faster. But I dunno if we're gonna get that below 2000$.

SaltyInternetPirate

I just want raytracing that's worth turning on in this upcoming generation. That's what I expected from the RTX 30 and the RX 6000, but even one generation after those, it's still such a big performance nose-dive, that there's no benefit. Also devs need to mark more materials as non-glossy. Most of the world doesn't need to shine reflections at us.

Bran Muffin

Will RTX 5090 need PCIe 5.0 to reach its full potential, or will 4.0 be enough? Debating whether to build a new rig for Blackwell.