Wednesday, August 28th 2024
AMD RDNA 4 GPU Memory and Infinity Cache Configurations Surface
AMD's next generation RDNA 4 graphics architecture will see the company focus on the performance segment of the market. The company is rumored to not be making a successor to the enthusiast-segment "Navi 21" and "Navi 31" chips based on RDNA 4, and will instead focus on improving performance and efficiency in the most high-volume segments, just like the original RDNA-powered generation, the Radeon RX 5000 series. There are two chips in the new RDNA 4 generation that have hit the rumor mill, the "Navi 48" and the "Navi 44." The "Navi 48" is the faster of the two, powering the top SKUs in this generation, while the "Navi 44" is expected to be the mid-tier chip.
According to Kepler_L2, a reliable source with GPU leaks, and VideoCardz, which connected the tweet to the RDNA 4 generation, the top "Navi 48" silicon is expected to feature a 256-bit wide GDDR6 memory interface—so there's no upgrade to GDDR7. The top SKU based on this chip, the "Navi 48 XTX," will feature a memory speed of 20 Gbps, for 640 GB/s of memory bandwidth. The next-best SKU, codenamed "Navi 48 XT," will feature a slightly lower 18 Gbps memory speed at the same bus-width, for 576 GB/s of memory bandwidth. The "Navi 44" chip has a respectable 192-bit wide memory bus, and its top SKU will feature a 19 Gbps speed, for 456 GB/s of bandwidth on tap.
Another set of rumors from the same sources also point to the Infinity Cache sizes of these chips. "Navi 48" comes with 64 MB of it, which will be available on both the "Navi 48 XTX" and "Navi 48 XT," while the "Navi 44" silicon comes with 48 MB of it. We are hearing from multiple sources that the "Navi 4x" GPU family will stick to traditional monolithic silicon designs, and not venture out into chiplet disaggregation like the company did with the "Navi 31" and the "Navi 32."
Yet another set of rumors, these from Moore's Law is Dead, talk about how AMD's design focus with RDNA 4 will be to ace performance, performance-per-Watt, and performance cost of ray tracing, in the segments of the market that NVIDIA makes the most volumes in, if not the most margins in. MLID points to the likelihood of the ray tracing performance improvements riding on there being not one, but two ray accelerators per compute unit, with a greater degree of fixed-function acceleration for the ray tracing workflow (i.e. less of it will be delegated to the programmable shaders).
Sources:
Kepler_L2 (memory speeds), Wccftech, VideoCardz (memory speeds), Kepler_L2 (cache size), VideoCardz (cache size), Moore's Law is Dead (YouTube)
According to Kepler_L2, a reliable source with GPU leaks, and VideoCardz, which connected the tweet to the RDNA 4 generation, the top "Navi 48" silicon is expected to feature a 256-bit wide GDDR6 memory interface—so there's no upgrade to GDDR7. The top SKU based on this chip, the "Navi 48 XTX," will feature a memory speed of 20 Gbps, for 640 GB/s of memory bandwidth. The next-best SKU, codenamed "Navi 48 XT," will feature a slightly lower 18 Gbps memory speed at the same bus-width, for 576 GB/s of memory bandwidth. The "Navi 44" chip has a respectable 192-bit wide memory bus, and its top SKU will feature a 19 Gbps speed, for 456 GB/s of bandwidth on tap.
Yet another set of rumors, these from Moore's Law is Dead, talk about how AMD's design focus with RDNA 4 will be to ace performance, performance-per-Watt, and performance cost of ray tracing, in the segments of the market that NVIDIA makes the most volumes in, if not the most margins in. MLID points to the likelihood of the ray tracing performance improvements riding on there being not one, but two ray accelerators per compute unit, with a greater degree of fixed-function acceleration for the ray tracing workflow (i.e. less of it will be delegated to the programmable shaders).
104 Comments on AMD RDNA 4 GPU Memory and Infinity Cache Configurations Surface
overclocking/comments/15yo3ng
www.techpowerup.com/review/cyberpunk-2077-phantom-liberty-benchmark-test-performance-analysis/6.html
You sound like a true fanboy, with 500 dollars ready to buy RDNA4 on release.
Sadly it will be another joke release from AMD.
Nothing from AMD will be worth buying till maybe RDNA5, completely new arch, on 3nm or better in late 2025 or early 2026.
RDNA4 is nothing but RDNA3 refined with slightly better RT performance. No-one really cares.
Also if that's correct then why did monolithic 7800 XT not outperform this "failed" 7900XT?
Their performance difference is 30%. If your theory is correct then should not 7800 XT perform as well as 7900 XT because it's monolithic?Wow you dug up a five year old card. AMD must be doing well if you had to go back five years to find one example.
And if we're talking about old cards then there was the FX 5800 "leaf blower" and the GTX 480 "Jensen's Grill" too.Ah yes. Every Nvidia's fanboy favorite tech demo. We'll here's one that is not custom made for Nvidia's hardware:
4% faster than 3090. 6% worse than 3090 Ti on average. Not bad for a card that supposedly cant even do RT.
You also forgot to mention that this destroys 4090 itself. 50fps at 1440p? 40fps with PT at 1440p? Unplayable slideshow on a $1700+ card.
But as long as AMD is below 30fps and 10fps respectively in this test it doesn't really matter for a fanboy does it?
Despite their GPUs not selling in as great numbers as Nvidia's, AMD is still a profitable company, mainly due to CPUs.
Let's also not forget the fact that a smaller company needs to sell lower quantities to stay profitable. Please don't tell me that your burger van has to compete with McDonald's in terms of sales numbers to stay competitive. :laugh:Better performance per watt is due to the architecture, not to MCM. Otherwise we wouldn't see the 8000G series being as efficient as they are.Check your idle power consumption. ;)What's wrong with that? Why do you think no one cares? I had a 7800 XT which is a fine card. My only issue was the video playback power consumption, on which if AMD can improve, then I'll be interested.Yes, because the 7600 is clearly 30% faster than the 6600 XT. Oh wait... :slap:
Lets have a look at their two recent ones:
www.techpowerup.com/review/star-wars-outlaws-fps-performance-benchmark/5.html
www.techpowerup.com/review/black-myth-wukong-fps-performance-benchmark/5.html
4090 absolutely wrecks 7900XTX.
More than 50% faster in pure raster, way more when adding RT testing + DLSS/DLAA destroys FSR with ease and Nvidia Frame Gen is highly superior to AMD Frame Gen too.
In a nutshell, you get what you pay for.
4080 uses 300 watts on average in gaming. 7900XTX uses 360 watts with custom cards peaking at 400+ which is the same as 4090, that performs way way better.
Atleast AMD fixed the massive power spikes Radeon 6800/6900 series suffered from
AMD GPUs have much higher powerdraw than Nvidia in multiple scenarios -> Idle, multi monitor, video playback, and more.
Also AMD GPU generally sucks in alot of games, especially competitive -
AMD GPU also sucks for emulation, in betas, in early access titles and just lesser popular games in general. AMD often don't have drivers ready for new games launching. Nvidia always have gameready drivers on day one, often many days before.
So, in the end, you save absolutely nothing buying an AMD GPU, when you consider the much lower resell value and higher powerdraw.
RDNA4 will change nothing
RDNA5 might, yet its not even close, 2026 probably
News, RDNA4 looks to be even more disappointing, 8700XT is going to be the top card
videocardz.com/newz/amd-rdna4-radeon-gpus-rumored-to-mirror-rdna1-in-product-positioning
And launch is like half a year away still. Reveal at CES 2025.
It's only MCM CPUs that suck a lot of power at idle (because the IO die and the infinity fabric eat up to 30 W), the 8000G series don't have trouble with it.
No it doesn't, because it's fine as an architecture. Your statement was RDNA3 provides no uplift over RDNA2, which is false. Also 7700xt is 25% faster than 6700XT, but the point is you can't choose and pick a single model and extrapolate that to the whole architecture.MCM was for cost savings due to higher yields, they took a hit to performance in the process and not the other way around. I forgot who did the analysis but they'd easily get another 10% performance with the same power if they didn't go MCM.
HBM for Fury was a necessity due to the power consumption of GCN when scaled to the max. But it's interesting you mention the 9xx generation, because in that generation the 970 was supposed to be the one that crushed 290x/390x. I have both, and the 290x aged far, far better than 970 . What was once supposed to be GTX780's competitor was competing with 780, then 780ti, then 970 and then GTX980. Point being, it's not always the case that there's no point in buying AMD it just depends on pricing and deals and regardless of the resale value, you can save a good chunk at times. Not to mention, the 970 specifically sucked when it came to longevity.
What AMD 'thinks' is what they've tested and come up with with regard to hotspot temperatures. If a chip runs hotter than another it makes 0 difference because the heat it dumps is the same whether it's 40'c or 100'c as long as the power is the same. If a chip is validated for a certain temperature there's no issue with that really as long as it's stable.
Where are you getting your sources about RDNA4 being a bugfix? What bug was there to fix? It's clear you have no idea about RDNA4 and are spouting stuff that don't really make sense but stating them anyway.
RDNA3 is neither. Added nothing, regressed mid range performance and pricing, and is in no way technically superior to the architecture it replaced.
If RDNA4 is more of the same, it has to be a Polaris move, else it will just not sell at all.
It might not have been as good as you'd hoped, but it's pointless to keep saying 'it adds nothing, not superior in any way' as that's just incorrect.
Only AMD can look at the 4060 and say "I'll do worse, hold my beer"
All improvements are nominal and whenever it's a concern, the architectural regressions are very much real. The fact that the 6900 XT is generally outperformed by the 7900 XTX is merely due to the scale of the 7900 XTX - which was clearly and painfully obviously designed to be a competitor to the RTX 4090. The 384-bit bus, the six MCDs, the high clocks - it was all obviously targeted at the 4090, until something went wrong along the way and they just didn't get the core and clock scaling they wanted out of the processor. Until that point, fine, just lower the TGP and release it anyway - the thing is that nobody was expecting the 4090 to be that powerful despite it being cutdown hardware. You should remember that the 4090 and 4080 launched first, in that situation all AMD could do was dig into their margins and release the product line with the flagship product targeting the RTX 4080, which is a smaller processor with a cheaper BoM.
Think, why were the MSRPs for 7900 XTX at 999 and 7900 XT at 900? Because they were both meant to be higher, and they just weren't sure how cheaper could they make them at the time. The 7900 XT at $900 was probably one of the worst value for money GPUs in recent history, and that's including the RTX 4080 at $1200. This is all before we even dig in at the ever problematic driver situation, the fact that Nvidia simply gives you more on that front for your investment, etc.
The entire point is that Nvidia's margins on the RTX 4090 are extreme - they're probably taking $1200+ per card sold. It's something previously unheard of in consumer-grade products, if there was any real pressure Nvidia could still make an insane amount of money by dropping the recommended price on the 4090 by $800. 1000 even. I sincerely don't think that it costs Nvidia more than $500 to manufacture an RTX 4090.
AMD probably spend sub 1% R&D funds developing it, DOA release
Even the AMD biased Youtubers are disappointed :laugh:
RDNA5 can't come soon enough.
Yes I know that there are things that didn't meet AMD's expectations internally but they absolutely weren't drastically off. But the 7900XTX wasn't meant to be a 4090 competitor. None of the things you said matter - 384bit bus was a necessity to feed the cores and deal with the reduced cache (as tested). High clocks weren't really much higher than RDNA2 at all, pretty close actually. These two things are standard architectural progressions and have nothing to do with the 4090. It's been a while and I can't remember which interview it was, but AMD knew they were incurring a performance penalty by not going monolithic (and reduction of cache because they're on the MCD now). This was done for yields but also as an experimentation or a 'tech demo' of sorts to see how far they can push the links between the dies and it wasn't an easy feat. The fact that they were intentionally taking a performance hit for yields shows they weren't really targeting the 4090 because they knew full well they need every % they can get in order to compete with it.
There were a few other reasons as well, like the dual issue SIMD's they introduced in RDNA3 will not be able to extract much performance because of quite a few bottlenecks present, but there's potential. Best way to improve it? Releasing hardware with that functionality, which is exactly what they've done.
I think this article outlines most of the changes in detail when going from RDNA2 to RDNA3. If you look closely, 7900XTX's GCD+MCD has about the same transistor count as the 4080 at slightly higher transistor density. I don't think AMD were thinking 'hey, let's compete with the 4090 with the same transistor budget as the 4080 while also taking a hit by going chiplet'. The conclusion sort of alludes to the same, have a read.I think at this point it would do well for you to stop reading those leaks and wait for release. None of these leaks are accurate and you have been told as such earlier yet you keep spamming the same thing every couple of days like you're a bot.
edit: Since you seem really into leaks, have a read at something that isn't one. You'll at least learn a thing or two unlike the silly leaks
Those silly leaks will be the official specs, you will see in a few months
I will be very impressed if they even come close to 7900XT (while charging 500 dollars tops). Way less cores on the top RDNA4 chip, smaller bus, I bet it will be around 7900 GRE only.
It will be a forgettable release. Low to mid-end focus. Meh.
RDNA5 = Brand new arch on 3nm TSMC or better. RDNA4 is a stop gap solution and nothing else. Maybe they can take back some marketshare if price is low enough. Nothing else is going to matter really. AMD is like sub 10% dGPU marketshare now.
RDNA4 will bring absolutely nothing new to the table. Lets hope for slightly better perf per watt and dollar.
RDNA5 is the next brand new arch and is not even close, expect late 2025 or even 2026. Probably won't have high-end SKUs either tho, since AMD officially left high-end market now. Lets hope they can compete with 5070, 5070 Ti and 5080 at least.