Friday, June 23rd 2023
Radeon RX 7800 XT Based on New ASIC with Navi 31 GCD on Navi 32 Package?
AMD Radeon RX 7800 XT will be a much-needed performance-segment addition to the company's Radeon RX 7000-series, which has a massive performance gap between the enthusiast-class RX 7900 series, and the mainstream RX 7600. A report by "Moore's Law is Dead" makes a sensational claim that it is based on a whole new ASIC that's neither the "Navi 31" powering the RX 7900 series, nor the "Navi 32" designed for lower performance tiers, but something in between. This GPU will be AMD's answer to the "AD103." Apparently, the GPU features the same exact 350 mm² graphics compute die (GCD) as the "Navi 31," but on a smaller package resembling that of the "Navi 32." This large GCD is surrounded by four MCDs (memory cache dies), which amount to a 256-bit wide GDDR6 memory interface, and 64 MB of 2nd Gen Infinity Cache memory.
The GCD physically features 96 RDNA3 compute units, but AMD's product managers now have the ability to give the RX 7800 XT a much higher CU count than that of the "Navi 32," while being lower than that of the RX 7900 XT (which is configured with 84). It's rumored that the smaller "Navi 32" GCD tops out at 60 CU (3,840 stream processors), so the new ASIC will enable the RX 7800 XT to have a CU count anywhere between 60 to 84. The resulting RX 7800 XT could have an ASIC with a lower manufacturing cost than that of a theoretical Navi 31 with two disabled MCDs (>60 mm² of wasted 6 nm dies), and even if it ends up performing within 10% of the RX 7900 XT (and matching the GeForce RTX 4070 Ti in the process), it would do so with better pricing headroom. The same ASIC could even power mobile RX 7900 series, where the smaller package and narrower memory bus will conserve precious PCB footprint.
Source:
Moore's Law is Dead (YouTube)
The GCD physically features 96 RDNA3 compute units, but AMD's product managers now have the ability to give the RX 7800 XT a much higher CU count than that of the "Navi 32," while being lower than that of the RX 7900 XT (which is configured with 84). It's rumored that the smaller "Navi 32" GCD tops out at 60 CU (3,840 stream processors), so the new ASIC will enable the RX 7800 XT to have a CU count anywhere between 60 to 84. The resulting RX 7800 XT could have an ASIC with a lower manufacturing cost than that of a theoretical Navi 31 with two disabled MCDs (>60 mm² of wasted 6 nm dies), and even if it ends up performing within 10% of the RX 7900 XT (and matching the GeForce RTX 4070 Ti in the process), it would do so with better pricing headroom. The same ASIC could even power mobile RX 7900 series, where the smaller package and narrower memory bus will conserve precious PCB footprint.
169 Comments on Radeon RX 7800 XT Based on New ASIC with Navi 31 GCD on Navi 32 Package?
I mean really, I'm looking at an 7900XT and I have yet to find fault with it in actual usage. Its virtually the same experience as an Nvidia card, and in terms of settings/GUI its a bit better. No Geforce Experience nagging you, but a proper functional and complete thing instead containing all the stuff you want. In games, I see a GPU that boosts and clocks dynamically, isn't freaking out over random stuff thrown at it, and just does what it must do, while murdering any game I throw at it.
And where is a "bug" in the first place?
In your mind? I have a reference model of 7900XTX and play on LG 4K/120Hz OLED TV with VRR over HDMI port.
Gameplay is smooth and 10-bit images are fantastic.
Is there anything wrong with my card? I keep hearing that Navi 31 has 'issues'. I can't see them. Is anyone able to enlighten me? "Fail to offer better performance over RDNA2..." What kind of nonsense is this?
7900XTX is 50% faster in 4K than 6900XT. To be 50% faster it needed 20% more CUs. All that for exactly the same price as in November 2020, no inflation included. So, the card is effectively even cheaper despite the same nominal value.
Where is alleged "failure" in any of this?
First time in a forum?
Besides, anyone can try to fit this at home by lowering refresh rate by 1Hz.
You jump in this thread that is now in 5th page with probably ABSOLUTELY NONE knowledge of what is written after page 1 and instead of hitting the brakes and tell to yourself "wait, let's just see how things progressed in this thread" you keep playing the same song.
This is beyond boring. Find someone else to vent your nerves.
The multi monitor VRAM fluctuation is no longer on the latest release bug list, but also isn't listed as being fixed either, so not sure if they forgot to write it or only add it under issues later once it's verified to still exist.
AMD clearly cares and can achieve some excellent power optimization (Rembrandt and Phoenix), but struggle when chiplets are involved. It's no easy task, to be clear, but they did advertise fanout link as being extremely efficient and capable of aggressively power management when Navi31 came out. Frustrating, because GCD power management is clearly outstanding, only to be squandered twice over by the MCDs and fanout link.
RDNA3 and Ada are closer in many aspects than quite a few previous generations. Nvidia went for 5nm-class manufacturing process for Ada, RDAN3 is using the same, RDNA3 doubled up on compute resources similarly to what Nvidia did in Turing/Ampere. Nvidia went with large LLC and narrower memory buses similarly to what AMD did in RDNA2.
Fermi vs TeraScale was different times. And DX11/DX12 transition on top of that. Off the top of my head these 480 shaders ion GTX480 were running at twice the clock rate of rest of the GPU. Wasn't TeraScale plagued by occupancy problems due to VLIW approach making those perform much slower than theoretical compute capabilities? The drawbacks present back then have been figured out for quite a while now.
Transistor counts by and large still follow the shader count and in recent times the large cache. Other parts - even if significant like Tensor cores - are comparatively smaller.
RDNA3 chiplet is not a radically different approach. It is clever, should be good for yields (=cost) but there really was no radical breakthrough here. What would that case be? RT has become (much) more relevant and RDNA3 competitor is not Ampere.
Also, while 7900XTX offers 3090Ti RT performance its raster performance is a good 20% faster...
AMD literally has a Jupiter sized window wide open to adjust their products and drivers, yet we only see some performance improvements, nothing special, like AMD is thinking they're Intel of 2012 when they had no one to compete with so making a dozen percent better product than the one from three years ago is completely fine.
It's not. Neither party deserves a cake. The only way I'm buying their BS SKUs is major discounts because they openly disrespect customers, me included. And AMD is worse because they don't use massive blunders by nVidia. This is only because we're a decade, maybe a couple decades too early for this technology. Nothing powerful enough to push this art to its beaut. RT is the answer to complete mirrors and reflections deficite in games. And the only stopping factor is hardware. It can't process it fast enough as of yet. When my post will be as old as me now RT will be a default feature, not even gonna doubt that (unless TWW3 puts us back to the iron age) nGreedia being nGreedia became obvious when they released their first RTX cards. It doesn't nullify my point though.
That being said, I'm just completely pessimistic about GPU market of the nearest couple generations. GPUs are ridiculously expensive and games are coming in "quality" so terrible it's them who must pay us to play it, not the opposite.
Again, this doesn't change the fact that the 7600 is, at the very least, 20% more efficient than the 6600XT. The node change for the 7600, by the way, was mostly insignificant. TSMC 7nm to "6nm" was a density increase with little-to-no efficiency increase. This is more clearly seen on the watt-per-frame graph, where the 7900XT & 7900XTX, both 5nm chips, have frames-per-watt equal to that of Nvidia's 4000 series, and better efficiency than the 7600.
Nvidia on the other hand, went from a mediocre Samsung 8nm node to a much superior TSMC 4nm node with the transition from the 3000 series to the 4000 series. It's not surprising to see a boost in efficiency.
At this point, there's not that much separating Nvidia and AMD GPU's in terms of hardware. Everything disappointing with the current GPU generation is the price with the ridiculous naming "upsell" of all cards by both AMD & Nvidia.
Diablo 4 on RDNA2 makes your desktop flicker when you alt tab to it, happens on my g14 laptop.
Driver installation requires you to disconnect from the internet. Which I guess is okay if you already know about it, I didn't, took me couple of hours to figure out why my laptop isn't working. Interesting. Are you of the same opinion with CPU's? Cause amd cpus consume 30-40w sitting there playing videos while intel drops to 5 watts.
You do realize the above shows that you don't understand efficiency either. You just throw away the power consumption numbers in the previous page of the review and keep that efficiency number because supports your opinion.
What I see is that efficiency is measured under a very specific scenario, which is Cyberpunk 2077. So, if 7600 enjoys an updated optimized driver in that game the result is in it's favor.
The problem with this efficiency result is that overall relative performance on both RX7600 and RX6650XT at 2160p is basically the same. Similarly, power consumption for these two is also basically the same.
I do not believe it is the driver. But RDNA3 has tweaks in the architecture and setup that are likely to benefit a cutting edge game like Cyberpunk 2077.
Now, going to the power consumption page we see
Idle : 7600 2W, 6600XT 2W
Multi monitor : 7600 18W, 6600XT 18W
Video Playback: 7600 27W, 6600XT 10W
Gaming: 7600 152W, 6600XT 159W
Ray Tracing: 7600 142W, 6600XT 122W
Maximum: 7600 153W, 6600XT 172W
VSync 60Hz: 7600 76W, 6600XT 112W
Spikes: 7600 186W, 6600XT 207W
From the above I see some optimizations in gaming and some odd problems. VSync 60Hz probably (IF I understand it correctly) shows that RDNA3 is way better than RDNA2 when asked to do some job that doesn't needs to push the chip's performance at maximum. Raster, maximum and spikes, that 7600 is more optimized than 6600XT. And optimizations there could be on the rest of the PCB and it's components, not the GPU itself. Idle at 2W can't help, multi monitor at 6600XT's levels, that probably AMD didn't improved that area over 6000 series. Raytracing is extremely odd and Video playback problematic. Best case scenario AMD to do what it did with 7900 series and bring the power consumption of video playback at around 10-15W.
In any case a new arch on a new node should be showing greens everywhere but idle, where it was already low. And not at almost equal average performance. 7600 should had the performance of 6700 and a power consumption equal or better in everything compared to 6600XT. We don't see this, especially the performance.
I don't feel duty-bound to defend the indefensible, these GPUs are hot garbage, not that Nvidia's are any better below the 4090 and I've pointed out my bone with the 4090 being heavily cutdown more than once, this is a lost generation and I only hope the next one is better. I want a battle royale with Battlemage, RDNA 4 and Blackwell in the midrange, and a competent AMD solution at the high end. I literally want to give AMD my money, but they don't make it easy! Every. Single. Generation. there's some bloody tradeoff, some but or if, some feature that doesn't work or some caveat that you have to keep in mind. This is why I bought my RTX 3090 after 4 generations of being a Radeon faithful. I no longer have the time or desire to spend hours on end debugging problems, working around limitations, or missing out on new features because AMD deems them "not important" or they "can't afford to allocate resources to that right now" or "we'll eventually make an open-source equivalent" that is either worse than the competition (such as FSR) or never gets adopted. I want a graphics card I can enjoy now, not potentially some day down the road. The day AMD understands this they will have gone half way through the road to glory.
It's an RDNA2 refresh with no USPs - look at that efficiency gap. That's not typical of RDNA3.
My 'you' is always a royal you unless specified otherwise :)
But yeah I share some of your pessimism, OTOH, its been worse. During Turing and mining for example. What an absolute shitshow we've been having. Perhaps we're still on the road to recovery altogether. You have been looking at the wrong numbers, as I also pointed out somewhere earlier or somewhere else - RDNA3's efficiency is very close to Ada, and is chart topping altogether. See above. The worst case scenario gap between RDNA3 and Ada is 15% in efficiency; if you take the 4080 at 4.0W versus the 4.6 of 7900. Also take note of the fact that monolithic and pretty linearly scaled Ada is itself showing gaps of 15% between cards in its own stack just the same. Reading this being said of the company that presented us with the shoddy and totally unnecessary 12VHWPR is... ironic. The same thing goes for all those 3 slot GPUs that have no business being a 3 slotter given their TDPs. Ada is fucking lazy, and much like RDNA3 it barely moves forward, the similarities are staggering. There is a power efficiency jump from the past gen, and that's really all she wrote. The rest is marketing BS. Nvidia just presents its thumb up ass story in a better way, that is really all it is. Geforce is a hand-me-down from enterprise-first technology, now more than ever.
The Vapor Chamber Fail on AMD side was worse though, I agree.
But Nvidia polished? That was three generations ago. GTX was polished. RTX is open beta, and featureset isn't even backwards compatible within its own short history. DLSS3 not being supported pre Ada is absolutely not polish and great product and caring about customers. Ampere was a complete shitshow and Ada is now positioned primarily to push you into a gen-to-gen upgrade path due to lack of VRAM on the entire stack below the 4080. I think you need a reality check, and fast.
So, despite nGreedia deserving their greedy men status, we all have to consider AMD even worse because they don't have no reason to be greedy.