I’m not sure the bulk of datacenter cards count as ‘discrete GPUs’ anymore, and they aren’t counted in that survey. They’re generally sold socketed into 8P servers with crazy interconnects, hyper specialized to what they do. Nvidia does sell some repurposed gaming silicon as a ‘low end’ PCIe server card, but these don’t get a ton of use compared to the big silicon sales.
I wouldn’t be surprised in the slightest if they are included in the list. I dunno, I’m not the statistician who crunched the numbers here. I didn’t collect the data, and that source material is not available for me to examine.
What I can say is that the article defines “discrete” GPUs instead of just “GPUs” to eliminate all the iGPUs. Because Intel dominates that space with AMD, but it’s hard to make an iGPU when you don’t make CPUs, and the two largest CPU manufacturers make their own iGPUs.
The overall landscape of the GPU market is very different than what this data implies.
AMD’s RX 9070 XT and RX 9070 represent AMD’s new RDNA 4 architecture, competing with Nvidia’s midrange offerings. Nvidia introduced two new Blackwell-series AIBs: the GeForce RTX 5080 Super and the RTX 5070. The company also announced the RTX 500 workstation AIB. Rumors have persisted about two new AIBs from Intel, including a dual-GPU model.
It is including workstation cards like the Blackwell Pro. But this is clearly not including server silicon like the B200, H200, MI325X and so on, otherwise they would have mentioned updates. They are not AIBs.
I hate to obsess over such a distinction, but it’s important: server sales are not skewing this data, and workstation sales volumes are pretty low. It’s probably a accurate chart for gaming GPUs.
The original post is about Nvidia’s domination of discrete GPUs, not consumer GPUs.
So I’m not limiting myself to people running an LLM on their personal desktop.
That’s what I was trying to get across.
And it’s right on point for the original material.
I’m not sure the bulk of datacenter cards count as ‘discrete GPUs’ anymore, and they aren’t counted in that survey. They’re generally sold socketed into 8P servers with crazy interconnects, hyper specialized to what they do. Nvidia does sell some repurposed gaming silicon as a ‘low end’ PCIe server card, but these don’t get a ton of use compared to the big silicon sales.
I wouldn’t be surprised in the slightest if they are included in the list. I dunno, I’m not the statistician who crunched the numbers here. I didn’t collect the data, and that source material is not available for me to examine.
What I can say is that the article defines “discrete” GPUs instead of just “GPUs” to eliminate all the iGPUs. Because Intel dominates that space with AMD, but it’s hard to make an iGPU when you don’t make CPUs, and the two largest CPU manufacturers make their own iGPUs.
The overall landscape of the GPU market is very different than what this data implies.
Well, it’s no mystery:
https://www.jonpeddie.com/news/q225-pc-graphics-add-in-board-shipments-increased-27-0-from-last-quarter/
It’s specifically desktop addin boards:
It is including workstation cards like the Blackwell Pro. But this is clearly not including server silicon like the B200, H200, MI325X and so on, otherwise they would have mentioned updates. They are not AIBs.
I hate to obsess over such a distinction, but it’s important: server sales are not skewing this data, and workstation sales volumes are pretty low. It’s probably a accurate chart for gaming GPUs.