Nvidia Confirms GeForce Cards Lack P2P, Pushing Expensive Pro Cards

RTX 6000 Ada Generation
RTX 6000 Ada Generation (Image credit: Nvidia)

The GeForce RTX 4090 is one of the best graphics cards for gaming, but consumers have found other uses for the Ada Lovelace flagship, such as content creation, machine learning, and scientific workloads. However, the GeForce RTX 4090 may be losing its appeal now that an Nvidia employee has confirmed that the graphics card doesn't support P2P (peer-to-peer) functionality.

In short, P2P is a neat technology that debuted on Nvidia graphics cards several years ago. The feature fundamentally acts like a highway between two Nvidia graphics cards. P2P enables data transmission between one graphics card's memory to the other, bypassing the memory on the system. It's a great feature for users working with CUDA programs because P2P accelerates memory access and transfer as opposed to the data having to pass through the system memory.

Puget Systems recently ran different multi-GPU benchmarks on AMD and Intel systems and discovered that P2P is partially broken on the GeForce RTX 4090. The publication pointed out that P2P-related workloads either failed or were corrupted. When P2P is enabled, the simpleP2P test fails. One Nvidia owner documented the problem in November last year, detailing that standard Nvidia tests failed when having two GeForce RTX 4090 graphics cards on a system. Months later, an Nvidia employee replied to the thread and confirmed that the GeForce RTX 4090 doesn't support P2P.

"Hi all. Apologies for the delay. Feedback from Engineering is that Peer to Peer is not supported on 4090. The applications/driver should not report this configuration as peer to peer capable. The reporting is being fixed and future drivers will report the following instead," wrote the Nvidia representative.

P2P works over PCIe or Nvidia's NVLink. The upside to using NVLink is that you enjoy lower latency and substantially more bandwidth. The funny part is that even the GeForce RTX 10-series (Pascal) offerings supported P2P over PCIe. Sadly, the GeForce RTX 20-series (Turing) graphics cards were the last generation to support P2P. However, P2P support is only available through the NVLink bridge, effectively limiting the number of total graphics cards per system to two. The prior GeForce RTX 30-series (Ampere) graphics cards lack P2P support, too, as evidenced by Puget Systems' tests.

If you meditate on the issue, it's easy to comprehend why Nvidia gradually dropped P2P support on GeForce graphics cards. The chipmaker has always disapproved of consumers using its mainstream products for anything other than gaming. However, during the Ampere days, word had gotten out that some system integrators started using blower versions of the GeForce RTX 3090 over the higher-priced Quadro alternatives to offer cost-effective server products. As a result, it didn't take long before the GeForce RTX 3090 blower graphics cards dematerialized from the market.

When you have a $1,599 GeForce RTX 4090, besting a $6,800 RTX 6000 Ada Generation on its home turf, it's certainly not good for business. Since you don't want to gimp the performance of your product because it would let your rival win, you have to find other ways to differentiate your mainstream offerings from your professional and workstation parts, such as playing with the feature set. For example, the RTX 6000 Ada Generation, which also lacks NVLink, has P2P support, according to Puget Systems. However, the GeForce RTX 4090 and the RTX 6000 Ada Generation are on the same Ada Lovelace architecture and use an identical AD102 silicon, so there's no good reason why one does P2P and the other can't if it's not a physical limitation on behalf of the manufacturer.

Nvidia's A-series offerings, previously Quadro, aren't all evil; they have some selling points, such as drivers certified for certain professional applications, better software compatibility, and more abundant VRAM. However, you do end up paying a massive premium for those features. If P2P is important in your daily work life, the RTX 6000 Ada Generation is the way to go if you have the funds. The previous RTX A6000, which costs $4,650, is still available and also supports NVLink as well as P2P.

Zhiye Liu
News Editor and Memory Reviewer

Zhiye Liu is a news editor and memory reviewer at Tom’s Hardware. Although he loves everything that’s hardware, he has a soft spot for CPUs, GPUs, and RAM.

  • derekullo
    "If you meditate on the issue, it's easy to comprehend why Nvidia gradually dropped P2P support on GeForce graphics cards "

    Here I am meditating on the issue

    https://i.imgur.com/mrxvTS7.png



    AI generated on a 4090
    Reply
  • DavidLejdar
    When someone wants to run 3 displays at 4K each with 3x 4090 for gaming, I suppose they are not gojng to try that with 8 GB of DDR4. But still... even if hardly a basic setup, not that far out that some may not be interested to have it - and somewhat curious that Nvidia doesn't want to give it some support, isn't it?
    Reply
  • edzieba
    Should be of no surprise whatsoever: P2P was only supported on 10xx to 30xx series cards with an NVLink connector, and no 40xx series card has an NVLink connector.
    Reply
  • eye4bear
    The article states at one point that the 20XX cards were the last to support P2P, meaning no one noticed this when the 30XX cards were tested, which would seem to me to point out this is not really used by many. Or did I read it wrong? If so, this would seem to be making a mountain out of a mole hill.
    Reply
  • Eximo
    I think the article mentioned that Puget was aware of the lack in the 30 series cards. But they also had 24GB cards, so I don't think it was that big a deal. Also had the Titan in the 20 series with 24GB. Which meant that if you wanted that much VRAM there was an option that wasn't too crazy in price.

    Basically separating the plebs at 24GB with the experts at 48GB who can afford to spend 4 times as much.
    Reply
  • mac_angel
    DavidLejdar said:
    When someone wants to run 3 displays at 4K each with 3x 4090 for gaming, I suppose they are not gojng to try that with 8 GB of DDR4. But still... even if hardly a basic setup, not that far out that some may not be interested to have it - and somewhat curious that Nvidia doesn't want to give it some support, isn't it?
    I'm not sure I follow. SLI for gaming is gone, the 40 Series doesn't even have the SLI/NVLink bridge. You can game with 3 monitors/displays (I am currently), but as far as I know, you can only use a single GPU with 3 displays plugged into it to enable NVidia Surround. The only other way I know of that might be possible is their A series with a Quadro Sync II card.
    Reply
  • Jensen should change his name to Mo Money
    Reply
  • Kyl3wyld
    Um, 3090s and 3090ti's do support NV link though. There's even 4 slot NV link bridges. That's something you should have already known before you ever wrote A word of this article. If You don't know the ins and outs of the products you cover how can you expect people to have any faith in your publication?
    Reply
  • Eximo
    But they still don't support P2P. So you can have the same memory contents on both GPUs, not use all of the memory as a single pool.
    Reply