Inspecting Gaming-pro
GB203 16 GB GDDR7GeForce RTX 5080 16GB Blackwell, PCIe dual-slot
16 GB GDDR7 at 360 W. The lot-of-20 card for render farms and small inference fleets.
Bandwidth 960 GB/s
FP8 0.9 PF
Per GB $74.94 $1,199
Inspecting Gaming-pro
GB202 32 GB GDDR7GeForce RTX 5090 32GB Blackwell, PCIe triple-slot
The fastest card you can buy at retail. 32 GB GDDR7, 1.8 TB/s, no ECC, no excuses.
Bandwidth 1.792 TB/s
FP8 1.68 PF
Per GB $78.09 $2,499
Inspecting Workstation
GB203 32 GB GDDR7RTX PRO 4500 Blackwell 32GB Blackwell, PCIe dual-slot
32 GB with ECC at 200 W. The quiet dev box card.
Bandwidth 896 GB/s
FP8 1.3 PF
Per GB $81.25 $2,600
Inspecting Workstation
GB202 96 GB GDDR7RTX PRO 6000 Blackwell 96GB Blackwell, PCIe dual-slot
The biggest VRAM you can put under a desk. 96 GB GDDR7, 24k CUDA cores, runs 70B models in a tower.
Bandwidth 1.8 TB/s
FP8 3.8 PF
Per GB $88.54 $8,500
Inspecting Datacenter
GB202 96 GB GDDR7RTX PRO 6000 Blackwell Server Edition 96GB Blackwell, PCIe dual-slot
96 GB GDDR7 per slot, passive. Eight in a 4U is 768 GB of VRAM for $69,184 at the 8-unit tier.
Bandwidth 1.6 TB/s
FP8 3.6 PF
Per GB $95.83 $9,200
Inspecting Workstation
GB202 48 GB GDDR7RTX PRO 5000 Blackwell 48GB Blackwell, PCIe dual-slot
48 GB at 300 W. Two of these beat one 96 GB card on throughput for most inference.
Bandwidth 1.344 TB/s
FP8 2.1 PF
Per GB $95.83 $4,600
Inspecting Datacenter
2,304 GB HBM3eHGX B300 8-GPU Blackwell Ultra, HGX baseboard
Eight B300s on one baseboard. 2.3 TB of HBM3e behind NVLink 5 for the largest single-node models.
Bandwidth 64 TB/s
FP8 72 PF
Per GB $182.29 $420,000
Inspecting Datacenter
AD102 48 GB GDDR6L40S 48GB Ada Lovelace, PCIe dual-slot
The universal datacenter card. 48 GB, FP8, video encode, runs in any 2U with 350 W to spare.
Bandwidth 864 GB/s
FP8 1.466 PF
Per GB $185.42 $8,900
Inspecting Datacenter
20,736 GB HBM3eGB300 NVL72 Blackwell Ultra, Rack
Seventy-two Blackwell Ultra GPUs, one NVLink domain, 20 TB of HBM3e in a single liquid-cooled rack.
Bandwidth 576 TB/s
FP8 720 PF
Per GB $188.08 $3,900,000
Inspecting Datacenter
GH100 141 GB HBM3eH200 NVL 141GB Hopper, PCIe dual-slot
141 GB of HBM3e on a PCIe card. Drops into a 4U chassis, bridges to 4-way NVLink. The practical path to 564 GB per box.
Bandwidth 4.8 TB/s
FP8 3.34 PF
Per GB $219.86 $31,000
Inspecting Datacenter
GB100 GB100 192 GB HBM3eB200 SXM 192GB Blackwell, SXM
The Blackwell training part. 192 GB HBM3e, 8 TB/s, NVLink 5 at 1.8 TB/s per GPU.
Bandwidth 8 TB/s
FP8 9 PF
Per GB $234.38 $45,000
Inspecting Datacenter
GH100 141 GB HBM3eH200 SXM 141GB Hopper, SXM
Hopper with 141 GB. The inference workhorse of 2026 and the easiest SXM part to actually get.
Bandwidth 4.8 TB/s
FP8 3.96 PF
Per GB $248.23 $35,000
Inspecting Datacenter
GH100 94 GB HBM3H100 NVL 94GB Hopper, PCIe dual-slot
94 GB HBM3 on PCIe at 400 W. Eight in a 4U gets you 752 GB without liquid cooling.
Bandwidth 3.9 TB/s
FP8 3.34 PF
Per GB $276.60 $26,000
Inspecting Datacenter
GH100 80 GB HBM3H100 SXM 80GB Hopper, SXM
The part the whole industry trained on. Still the best price per FLOP in SXM, allocated in 1 to 2 weeks.
Bandwidth 3.35 TB/s
FP8 3.96 PF
Per GB $343.75 $27,500