guru3d 9800X3D Specs leaked

[UltraMega]

Quote

Advanced Micro Devices (AMD) has announced the upcoming release of its Ryzen 9000X3D desktop processor, scheduled for November 7. 

https://www.guru3d.com/story/amd-set-to-launch-ryzen-9000x3d-desktop-processor-on-november-7/

 

 

 

Quote

Advanced Micro Devices (AMD) has officially implemented price reductions across its Ryzen 9000 "Granite Ridge" desktop processor lineup. The adjustments apply to all four current models, making these high-performance CPUs more accessible to consumers through retail channels. Notably, the flagship Ryzen 9 9950X, featuring 16 cores and 32 threads, has seen a price decrease of up to $50 from its initial launch cost, now available for approximately $600. Similarly, the Ryzen 9 9900X with 12 cores and 24 threads has been reduced by up to $30, bringing its price down to around $470. The Ryzen 7 9700X (8 cores/16 threads) and Ryzen 5 9600X (6 cores/12 threads) have also experienced price cuts of up to $30, with new prices set at approximately $330 and $250, respectively. These reductions are effective immediately, allowing consumers to benefit from the updated pricing without delay.

https://www.guru3d.com/story/amd-reduces-prices-for-ryzen-9000-series-desktop-processors/

 

 

 

Definitely curious to see how the X3D chips shake out.

 

Edited October 28 by UltraMega

[UltraMega]

Update:

 

Quote

AMD’s upcoming CPU will reportedly utilize eight Zen 5 cores with a 4.70GHz base clock speed, which can be boosted up to 5.20GHz - a 0.20GHz jump over the AMD Ryzen 7 7800X3D, and with the CPU multiplier unlocked, users will be able to manually overclock this new processor. This wasn’t possible with previous AMD CPUs that utilized 3D V-Cache, which could be a significant boon for overclocking enthusiasts.

 

The leaked specifications reference 96MB of 3D V-Cache, identical to the 7800X3D (which is among AMD’s best processors), but the aforementioned marketing leak suggested that the 9800X3D may have an 8% performance boost over AMD's current high-end gaming CPU. This is reportedly powered by ‘Next-Gen 3D V-Cache’, but the Geizhals leak doesn’t appear to add any more information.

 

AMD’s Ryzen 7 9800X3D full specifications have leaked, and it’s great news for gamers and overclockers alike

 

[Sir Beregond]

On 28/10/2024 at 14:21, UltraMega said:

Update:

 

 

AMD’s Ryzen 7 9800X3D full specifications have leaked, and it’s great news for gamers and overclockers alike

 

The biggest interesting rumor to me with this is the unlocked multiplier. Definitely interested to see what it's actually capable of when pushed if they have figured out a way to deal with thermals on stacked cache.

[UltraMega]

https://www.guru3d.com/story/amd-officially-announces-nextgeneration-amd-ryzen-7-9800x3d-at/

 

 

$479

 

[mouacyk]

On 29/10/2024 at 21:02, Sir Beregond said:

The biggest interesting rumor to me with this is the unlocked multiplier. Definitely interested to see what it's actually capable of when pushed if they have figured out a way to deal with thermals on stacked cache.

Hopefully that comes with an unlocked core voltage beyond 1.2v and vsoc beyond 1.3v also

[Sir Beregond]

21 hours ago, mouacyk said:

Hopefully that comes with an unlocked core voltage beyond 1.2v and vsoc beyond 1.3v also

Good call out, and agreed.

[Fluxmaven]

Whenever I can walk into Microcenter and grab one for under $400, I'll do so. I'm not really in a rush since I already have a 7800X3D rig. I also have a 7700X on the test bench I could run as a placeholder and go ahead and rebuild the 5900X rig soon-ish. Then I could toss the 9800X3D in the Gene and bench it before it gets swapped into the main rig. 

 

I guess it'll depend on how unlocked it is, if I want to set aside time to bench it first. Just haven't been as interested in PC stuff the past year or so. 

[UltraMega]

30 minutes ago, Fluxmaven said:

Just haven't been as interested in PC stuff the past year or so. 

 

I'm just going to take this as an opportunity to go on a mini-rant:

 

We're at the point where Moore's Law is dead. 

 

I think Moore's law actually died at 16nm. It's my understanding that there are parts of a CPU that just cannot get smaller than 16nm, where as others can. Individual parts of a CPU have different limits to how small they can be. While we're not quite at the limit for how small transistors can be, we're close and we are at limits for a lot of other parts of modern chips. Ever since we started hitting these limits, there is less and less to gain and less and less to be excited about with new hardware. 

 

That's why efficiency gains are a bigger deal now than ever before. Partly because increasing efficiency is one way to reduce heat and thus push chips harder, and also because there are not big wins left for performance. 

 

 

But then there's NPUs. 

 

I honestly think Nvidia did not have a plan to create DLSS when it launched the first NPU/Tensor enabled graphics cards. I don't know this to be true and it's pure speculation, but just following the development of this stuff as it went on I felt like DLSS was Nvidia trying to create something to justify the extra cost of the tensor cores. I think Nvidia knew they would be useful for something, and they wanted to create a foundation of hardware that could support a market for things like LLMs and all the other stuff NPUs can do. Nvidia was lucky that DLSS worked out so well, much better than I think they had any idea of when they first released the 2000 series. 

 

Fun fact about DLSS

Spoiler

At one point I did a deep dive into DLSS and how it compared to other stuff like FSR and XeSS to try to answer the question of "Could DLSS work on AMD GPUs that don't have tensor cores?" And the answer is definitely no.

 

DLSS hammers the tensor cores with data that gets processed extremely fast. The tensor cores can appear to be doing almost nothing during a DLSS workload because they process the data so quickly that it basically has a lot of idle time between loads. The data has to be process super-fast in order for it to work at the level DLSS does it, with temporal data and "AI" driven inferencing and all that. 

 

Another fun fact, XeSS on an Intel GPU actually processes way more data than DLSS, like ten times as much. On other GPUs it uses a lower quality fall back method, but on an Intel GPU it's doing a ton of work. 

 

AMD GPUs just don't have that capability right now. They can do AI workloads, but not real time 3D rendering AI-driven workloads because they just don't have a component to push through huge amounts of inferencing data in real time. 

 

I think NPUs are a very interesting future direction for hardware, much more so in relation to gaming than anything else going on at the moment. We're still in the early days with what NPUs can bring to the table for 3D rendering. NPUs can do a lot to speed up the actually rendering process, and so far we are just using it for image enhancement, so it's basically a post processing effect in a lot of ways (though not entirely). 

 

Nvidia was basically able to get an 8X multiplier of rendering power in the sense that DLSS3 with frame gen can render 1 out of ever 8 pixels you actually see and get roughly comparable quality to native rendering, though with some obvious trade-offs. I don't think most people realize how impressive that is right now because the effects of Moore's law ending make it feel like there are just normal gains, but they're actually much more fascinating gains based on new hardware that is still really early yet extremely capable. 

[Sir Beregond]

28 minutes ago, UltraMega said:

 

I'm just going to take this as an opportunity to go on a mini-rant:

 

We're at the point where Moore's Law is dead. 

 

I think Moore's law actually died at 16nm. It's my understanding that there are parts of a CPU that just cannot get smaller than 16nm, where as others can. Individual parts of a CPU have different limits to how small they can be. While we're not quite at the limit for how small transistors can be, we're close and we are at limits for a lot of other parts of modern chips. Ever since we started hitting these limits, there is less and less to gain and less and less to be excited about with new hardware. 

Yeah I think in many way yes, in some ways no.

 

Analog components (your I/O, etc.) stopped scaling with new nodes way back in 28nm and maybe even before then. Cache was scaling well until around 7nm and has been leveling off since. Logic (your CPU / GPU cores, etc) on the other hand, continues to scale with each new process shrink, and actually looks like it might start scaling even better as we continue.

 

Seems to suggest scaling is favoring logic these days while cache and analog transistors are already so much smaller than logic transistors, that they are probably already at their physical limits and in that sense, yeah Moore's Law is dead.

 

And then yeah you mention the efficiency part which is interesting because on one hand companies are really trying to market themselves as efficient. Meanwhile the actual base power draw is so much more these days that they used to be. Top end GPUs used to max out at 250W. Now you have a 450W (default) 4090 that is arguably one of the most efficient power to performance GPUs Nvidia has released to the consumer market, yet its still drawing far more than we used to.

 

Anyway, I'm interested in hardware still to an extent, but likewise hard to get excited about stuff when large market segments have stagnated and only gotten more expensive.

 

 

Based on this chart from AMD, one of the reasons chiplets made sense. Logic on the most advanced node they want to use, meanwhile offload I/O component to older nodes because there is no reason to use the more advanced/expensive ones.

Edited November 1 by Sir Beregond

[UltraMega]

46 minutes ago, Sir Beregond said:

Yeah I think in many way yes, in some ways no.

 

Analog components (your I/O, etc.) stopped scaling with new nodes way back in 28nm and maybe even before then. Cache was scaling well until around 7nm and has been leveling off since. Logic (your CPU / GPU cores, etc) on the other hand, continues to scale with each new process shrink, and actually looks like it might start scaling even better as we continue.

 

Seems to suggest scaling is favoring logic these days while cache and analog transistors are already so much smaller than logic transistors, that they are probably already at their physical limits and in that sense, yeah Moore's Law is dead.

 

And then yeah you mention the efficiency part which is interesting because on one hand companies are really trying to market themselves as efficient. Meanwhile the actual base power draw is so much more these days that they used to be. Top end GPUs used to max out at 250W. Now you have a 450W (default) 4090 that is arguably one of the most efficient power to performance GPUs Nvidia has released to the consumer market, yet its still drawing far more than we used to.

 

Anyway, I'm interested in hardware still to an extent, but likewise hard to get excited about stuff when large market segments have stagnated and only gotten more expensive.

 

 

Based on this chart from AMD, one of the reasons chiplets made sense. Logic on the most advanced node they want to use, meanwhile offload I/O component to older nodes because there is no reason to use the more advanced/expensive ones.

 

 

The gains are getting smaller and smaller, and further apart. Even if smaller nodes scale we'll, we're still almost at the end of the road. We're at 3nm right now, right? Can't get any smaller than 1nm or even if we can it will be just barely. Once we get there, that's it. The only gains left will be architecture improvements until and unless we find some new ways to make chips. 

Edited November 1 by UltraMega

[GanjaSMK]

I thought predictions on compute chips (CPU / FPU {then} | NPU / APU {now} / etc) some 25 years ago (1999 computer mags) envisioned chips would become independent rather than combined once proper interconnects (then theory was possibly graphene, photonics, etc) could handle the data sets at speeds required for the 'future' of computing to continue. 

 

We're seeing this beginning to happen now, no?