NVIDIA RTX 40-series "Ada" GPUs to Stick to PCI-Express Gen 4

[btarunr]

NVIDIA's next-generation GeForce "Ada" graphics architecture may stick to PCI-Express 4.0 as its system bus interface, according to kopite7kimi, a reliable source with NVIDIA leaks. This is unlike Ada's sister-architecture for compute, "Hopper," which leverages PCI-Express 5.0 in its AIC form-factor cards, for its shared memory pools and other resource-sharing features similar to CXL. This would make Ada the second graphics architecture from NVIDIA to use PCIe Gen 4, after the current-gen "Ampere." The previous-gen "Turing" used PCIe Gen 3. PCI-Express 4.0 x16 offers 32 GB/s per-direction bandwidth, and NVIDIA has implemented the Resizable-BAR feature with "Ampere," which lets the system see the entire dedicated video memory as one addressable block, rather than through tiny 256 MB apertures.

Despite using PCI-Express 4.0 for its host interface, GeForce "Ada" graphics cards are expected to extensively use the ATX 3.0 spec 16-pin power connector that the company debuted with the RTX 3090 Ti, particularly with higher-end GPUs that have typical board power above 225 W. The 16-pin connector is being marketed as a "PCIe Gen 5" generation standard, particularly by PSU manufacturers cashing in on early-adopter demand. All eyes are now on AMD's RDNA3 graphics architecture, on whether it's first to market with PCI-Express Gen 5, the way RDNA (RX 5000 series) was with PCIe Gen 4. The decision to stick with PCIe Gen 4 is particularly interesting given that Microsoft DirectStorage may gain use in the coming years, something that is expected to strain the system bus for the GPU, as SSD I/O transfer-rates increase with M.2 PCIe Gen 5 SSDs.



View at TechPowerUp Main Site | Source

 

[Flydommo]

PCIe Gen 4 should be enough in nearly all use cases at this point. Only a few users need even more speed for file/data transfers on an everyday basis. Maybe with the RTX 50-series we'll see PCIe Gen 5 in like 2023/24.

 

[Crackong]

Day1 PCI-E 5 device = 0
Day180 PCI-E 5 device still= 0

Poor Intel

 

[R0H1T]

PCIe Gen 4 should be enough in nearly all use cases at this point. Only a few users need even more speed for file/data transfers on an everyday basis. Maybe with the RTX 50-series we'll see PCIe Gen 5 in like 2023/24.

I doubt that, it will increase the cost of GPU's as well. AMD/NVidia will avoid it for as long as they can, it's really not needed now with SLI or CrossFire a thing of the past.

Earliest I predict it being a part of mainstream is 2025, weirdly Intel may push it because it's the new entrant & this could be a differentiating factor they'd obviously wanna promote!

 

[agent_x007]

They can try to push it for low end segment to chop PCI-e lanes to "x1 5.0/6.0".

 

[AusWolf]

Good. Hopefully it will bring costs and prices down a notch. Not that we need gen 5 so soon anyway.

 

[Broken Processor]

I really hope we don't see any own goals next gen like pcie4 x4 cards AMD released scuppering pcie 3 owners.

 

[Xajel]

While I wish it has PCIe 5.0 just for the sake of it. I know it won't matter even when MS DirectStorage becomes a thing.

I mean, PCIe 4.0 x16 will have the same bandwidth as PCIe 5.0 x8, meaning you need 2x PCIe 5.0 NVMe at full speed and on RAID 0 to really make things closer to a bottle neck.
But again, I wish it will have PCIe 5.0 just for the sake of it and coz other things will have it in it's lifetime.

 

[Dr. Dro]

Day1 PCI-E 5 device = 0
Day180 PCI-E 5 device still= 0

Poor Intel

I don't see how having forward-support for an interconnect as important as PCI Express is a bad thing

 

[TheinsanegamerN]

PCIe Gen 4 should be enough in nearly all use cases at this point. Only a few users need even more speed for file/data transfers on an everyday basis. Maybe with the RTX 50-series we'll see PCIe Gen 5 in like 2023/24.

PCIe gen 3 is plenty enough. Technically PCIe gen 2 is plenty, the penalty for single GPU at 2.0 really only applies to 3090 tier GPUs.

 

[AusWolf]

PCIe gen 3 is plenty enough. Technically PCIe gen 2 is plenty, the penalty for single GPU at 2.0 really only applies to 3090 tier GPUs.

... on 16 lanes, I might add. You take 12 of those 16 away, like AMD did with the 6400 / 6500 XT, and you have a bit of a hit-and-miss situation.

 

[dinmaster]

thats fine if motherboards get pci-e v5. gpu's don't need to pay a premium for a standard they don't even come close to in terms of bandwidth usage. The real reason for pci-e v5 is going to be storage speed in m.2 ssd sticks and to have more channels open as well. I'm waiting for a cable standard to replace sata3 cables so we can have harddrives/ssd's runing on pci-e (cabled) instead of the dusty sata3 standard. we may be slowly going in that direction...

 

[defaultluser]

Hopefully, these second-gen PCIE4 slots can be more efficient - the fact 5that the Rtx 3050 is a 150w card is mostly due to this!

I mean, you can certainly have a bus-powered card, (if you castrate the connection to x4, and the bus to 64-bit - 6400 or you take the 3060, then under-clock to 3050-performance)

 

[AusWolf]

Hopefully, these second-gen PCIE4 slots can be more efficient - the fact 5that the Rtx 3050 is a 150w card is mostly due to this!

I mean, you can certainly have a bus-powered card, (if you castrate the connection to x4, and the bus to 64-bit - 6400 or you take the 3060, then under-clock to 3050-performance)

It's not the bus that consumes power, but the GPU and the settings it runs at. The reason why modern GPUs aren't nearly as efficient as they could be is due to the fact that both nvidia and AMD run them at the peak of their efficiency curves by default. No one asked the 6500 XT to run at 2800+ MHz and consume 100-120 Watts. It could have easily been a no-power-connector 75 W card at 2500-2600 MHz, but nooo! Performance is king nowadays, even in the lower segments.

 

[Assimilator]

Considering that Hopper already has PCIe 5.0, I'm gonna go out on a limb and guess that Ada also has that capability, but NVIDIA has simply fused it off. Then when AMD launches its next series of GPUs, NVIDIA will launch a minor refresh of Ada that is PCIe 5.0-enabled to steal their thunder.

 

[noel_fs]

pcie5 on motherboard is enough for storage purposes

 

[Raiden85]

We haven't maxed out 16x PCI-E 3.0 just yet let alone 4.0, 5.0 at least on GPU's for now is pointless.

 

[ModEl4]

I don't know about Nvidia, but AMD might use PCI-E 5.0.
If the rumours are true and the frequency of Navi 31 is 3GHz, then it will have logically double the pixel-fillrate of Navi 21 (Navi 33 64RBs half of Navi 21, Navi 32 128RBs, Navi 31 192RBs) while the memory bus will still be 256bit and with the GDDR6 being at 6950XT level or slightly more (20Gbps?, the Samsung 24Gbps option probably shouldn't be ready for launch) it will need to throw the kitchen sink in order not to be bandwidth limited, so at least 256MB infinity cache (rumour is for 512MB which is an insane amount of transistors/mm2 on 6nm, without the additional logic that the module will incorporate, we are talking for at least 24 billion transistors and more than 250mm² just for the 512MB cache portion of the chiplet) , PCI-E 5.0, better compression etc in order to help with all the memory access related issues of the memory stack.

 

[defaultluser]

I don't know about Nvidia, but AMD might use PCI-E 5.0.
If the rumours are true and the frequency of Navi 31 is 3GHz, then it will have logically double the pixel-fillrate of Navi 21 (Navi 33 64RBs half of Navi 21, Navi 32 128RBs, Navi 31 192RBs) while the memory bus will still be 256bit and with the GDDR6 being at 6950XT level or slightly more (20Gbps?, the Samsung 24Gbps option probably shouldn't be ready for launch) it will need to throw the kitchen sink in order not to be bandwidth limited, so at least 256MB infinity cache (rumour is for 512MB which is an insane amount of transistors/mm2 on 6nm, without the additional logic that the module will incorporate, we are talking for at least 24 billion transistors and more than 250mm² just for the 512MB cache portion of the chiplet) , PCI-E 5.0, better compression etc in order to help with all the memory access related issues of the memory stack.

We'll see - they were already bandwidth-limited on Big Navi!

Then there's the added overhead of getting two chiplets to talk to each other (all while sharing the same castrated 256-bit bus!) The only thing improved in this aspect is the doubling of infinity cache, plus a 10% bump in GDDR6 clock!

You'll be lucky if the performance is 40% faster at 4k

 

[R-T-B]

We haven't maxed out 16x PCI-E 3.0 just yet let alone 4.0, 5.0 at least on GPU's for now is pointless.

Actually with the 3090 and above we have, but barely. There is a consistent but present gain of around 2% between pcie 3.0 x16 and pcie 4.0 x16.

See TPUs review on this.

 

[ModEl4]

We'll see - they were already bandwidth-limited on Big Navi!

Then there's the added overhead of getting two chiplets to talk to each other (all while sharing the same castrated 256-bit bus!) The only thing improved in this aspect is the doubling of infinity cache, plus a 10% bump in GDDR6 clock!

You'll be lucky if the performance is 40% faster at 4k

lol 40%, what are you talking about?
Now seriously, the easy prediction i can offer is that if infinity cache is 512MB, AMD will try to label Navi 31 as an 8K capable card, forcing reviewers to examine 8K resolution and will try to change the narrative into how it can win in some titles in 8K res vs Nvidia AD102 with it's only 96MB cache, when in reality the comparison should have been at 4K where the 96MB for the Nvidia Architecture is just fine...

 

[lexluthermiester]

There is a consistent but present gain of around 2% between pcie 3.0 x16 and pcie 4.0 x16.

That has to do more with the improved latency of PCIe4.0 than the raw bandwidth. We're still a little ways off from fully saturating the PCIe3.0 16x bus. But in fairness, we are close.

See TPUs review on this.

W1zzard did some testing with a RX5700XT.

PCI-Express 4.0 Performance Scaling with Radeon RX 5700 XT

PCI-Express 4.0 has been one of the main new features prominently brandished on the product boxes of both the 3rd generation Ryzen desktop processors and Radeon RX 5700 series graphics cards. We examine the performance impact of running these cards on older generations of PCIe.

www.techpowerup.com

The difference between PCIe 2.0, 3.0 and 4.0 were 1% or 2%, depending on the resolution.

The next year he followed up with a more testing, but this time with a 3080.

NVIDIA GeForce RTX 3080 PCI-Express Scaling

NVIDIA Ampere finally brings PCI-Express 4.0 support to the high-end graphics market. The new interface promises twice the bandwidth of PCI-Express 3.0. We've setup an AMD Ryzen 3900XT system to test how various PCIe generations and lane widths affect gaming performance.

www.techpowerup.com

In this series of testing, he included PCIe1.1 spec in the 16x lanes as well as PCIe1.1 8x lanes. For PCIe 2.0 the difference where few percent more, but PCIe 3.0 and 4.0 were still at a 1% difference, not enough to be at all worried about.

While a 3090 or a 3090ti are faster than a 3080, they are not so much faster as to present serious PCIe bus bottlenecking situation for the PCIe 3.0 bus, and minimal bottlenecking for PCIe 2.0.

 

[R-T-B]

lol 40%, what are you talking about?
Now seriously, the easy prediction i can offer is that if infinity cache is 512MB, AMD will try to label Navi 31 as an 8K capable card, forcing reviewers to examine 8K resolution and will try to change the narrative into how it can win in some titles in 8K res vs Nvidia AD102 with it's only 96MB cache, when in reality the comparison should have been at 4K where the 96MB for the Nvidia Architecture is just fine...

I mean nvidia already attempted to brand the GA102 chip as an 8k chip. No one took it seriously.

 

[ModEl4]

I mean nvidia already attempted to brand the GA102 chip as an 8k chip. No one took it seriously.

And rightfully so but this time if the performance is 2.5X vs 6900XT according to leaks (i doubt it will reach this level, probably 2.3X-2.4X at 4K) then it will offer the same experience at 8K as RX 6800 had at launch at 4K (AMD branded 6800 as an 4K card at launch).
The problem is that there aren't many people having 8K displays (and those aren't OLED based unless we are talking about 10.000€-30.000€ TVs) and the forecast is that it will take around 5 years for 8K OLEDs to reach mainstream prices (when Chinese manufacturers gradually ramp up production in the next 5 years - let's see how the war and its reverberations will escalate first
And honestly I'm sick and tired every time AMD tries to change the narrative however they fit them (for example they decided for 6650XT, a $400-500 SRP card depending on the brand/model to place it as an 1080p card, guiding the reviewers to take priority in their conclusions for the 1080p difference vs the competition due to the fact that infinity cache size kills the performance at higher resolutions (for example a MSI RTX 3060 Gaming X is 1-2% faster than a MSI RX 6650XT Gaming X at 4K)
Sure 4K isn't the intended resolution for these cards, but there are many games like Doom Eternal, Resident Evil 3, F1 2020 etc + older ones of course that the MSI RTX 3060 Gaming X averaging more than 60fps at 4K max settings, now add to that all the games that with slightly lower settings can hit 60fps at 4K with very minor visual differences and the catalog isn't small.
But set aside the 4K argument which rightfully shouldn't be the main criteria for 3060-6650 performance difference, why not 1440p, is it too much to ask for a 400-500 SRP card? (after all the average fps in TPU setup, a 6650XT is hitting at QHD is very similar with a reference 6900XT at 4K, around 82 vs 87 average fps, so if 6900XT is fine for 4K why not QHD for 6650XT?) Is it the fact that $399 SRP 3060Ti is around 20% faster in QHD and +37% in 4K for example? (reference vs reference or OC vs OC)

 

[AusWolf]

That has to do more with the improved latency of PCIe4.0 than the raw bandwidth. We're still a little ways off from fully saturating the PCIe3.0 16x bus. But in fairness, we are close.

W1zzard did some testing with a RX5700XT.

PCI-Express 4.0 Performance Scaling with Radeon RX 5700 XT

PCI-Express 4.0 has been one of the main new features prominently brandished on the product boxes of both the 3rd generation Ryzen desktop processors and Radeon RX 5700 series graphics cards. We examine the performance impact of running these cards on older generations of PCIe.

www.techpowerup.com

The difference between PCIe 2.0, 3.0 and 4.0 were 1% or 2%, depending on the resolution.

The next year he followed up with a more testing, but this time with a 3080.

NVIDIA GeForce RTX 3080 PCI-Express Scaling

NVIDIA Ampere finally brings PCI-Express 4.0 support to the high-end graphics market. The new interface promises twice the bandwidth of PCI-Express 3.0. We've setup an AMD Ryzen 3900XT system to test how various PCIe generations and lane widths affect gaming performance.

www.techpowerup.com

In this series of testing, he included PCIe1.1 spec in the 16x lanes as well as PCIe1.1 8x lanes. For PCIe 2.0 the difference where few percent more, but PCIe 3.0 and 4.0 were still at a 1% difference, not enough to be at all worried about.

While a 3090 or a 3090ti are faster than a 3080, they are not so much faster as to present serious PCIe bus bottlenecking situation for the PCIe 3.0 bus, and minimal bottlenecking for PCIe 2.0.

I think the difference between PCI-e versions is nothing to be concerned about with a normal x16 graphics card. It's more of an issue on fewer lanes, like the x4 of the Radeon RX 6400 and 6500 XT.