Flopper：面向 AI 基础设施的硬件浏览器，可对比 221 款 GPU 与 82 套系统
Show HN: Flopper is a hardware explorer for AI infrastructure and beyond
Flopper 是一个 AI 基础设施硬件浏览器，把 GPU 规格、性能与云租赁价格集中到一处对比，数据库收录 221 款 GPU、82 套多卡 AI 系统和 786 个 GPU 数据中心。所有厂商数据被统一归一化到同一 TFLOPS 基线，覆盖 FP64 到 FP4 各精度，每个数据点均链接回官方 datasheet、白皮书或经验证的 benchmark。
Compare the hardware powering AI
Explore GPUs, specifications, performance and cloud pricing in one place.
Database totals as of 10 October 2026
Popular GPUs for AI
The most researched GPUs you can buy or rent today for training or hosting AI models.
# GPU Memory Bandwidth FP8 FP8 per $ Lowest Median Providers Availability 30d trend Rent
Rental prices Explore all 221 GPUs
Where to rent
A few of the clouds we track, with the number of GPUs priced live in their catalogue right now.
Popular AI Systems
The most popular multi-GPU AI systems you can buy or rent today, from individual servers to full rack-scale platforms.
# System Accelerators Total Memory FP16 From $/hr Power Rent
† Aggregate theoretical peak: GPU count times the accelerator's per-chip FP16 figure, not a vendor-published system total.
‡ From $/hr is the accelerator's cheapest live on-demand rate times the GPU count, not a quoted node price.
Analysis
Original reporting from the Flopper desk: roadmaps decoded, claims checked against the database.
Skynet Ran at 60 Teraflops. That Is One Gaming GPU Now.
31 Oct 2026 In Terminator 3 (2003), Skynet runs at "60 teraflops per second", which beat the fastest computer on Earth at the time. In 2026 that is one RTX 4090, or 1/18,000th of a GB300 NVL72 rack at FP4. Count in FP64, the precision a 2003 supercomputer meant, and the whole rack is only 1.7 Skynets. A Halloween look at why a TFLOPS figure without a precision is a plot device, not a spec. 17 Sept 2026 NVIDIA is bringing native Rust support to CUDA with two projects, cuda-oxide and cutile-rs. The bigger story may be CUDA Tile, a programming model where developers describe operations on tiles of data and the compiler decides how they map onto the GPU. The hardware differences between an A100, an H100 and a B200 are not going anywhere, but software may need to know less about them. 17 Sept 2026 A new Blackwell workstation card appeared on NVIDIA's site with no press release: 84 GB of GDDR7 with ECC, 1,398 GB/s, up to 600 W and the RTX 5090's 21,760 CUDA cores, aimed at rack-mounted workstation deployment. NVIDIA has published no TFLOPS, price or date, and the 84 GB does not fit the listed 416-bit bus. We separate what is published from what is not, and the card is now in the database.
NVIDIA Wants GPU Code to Care Less About the GPU
NVIDIA Quietly Lists the RTX PRO 5500 Blackwell: 84 GB, Built for the Rack
One Database. Every Precision.
Vendor datasheets use different formats, units, and conditions. We normalize everything so you can compare what actually matters.
Normalized Across Vendors
NVIDIA, AMD, Intel, Google, and more — all converted to the same TFLOPS baseline at every precision level. No more cross-referencing datasheets.
Every Precision Tracked
FP64 through FP4, including BF16, TF32, and INT8. Filter and sort by the precision your workload actually uses — not just the headline number.
Specs Tied to Sources
Every data point links back to its origin: official datasheets, vendor whitepapers, or verified benchmarks. No unattributed claims.
Your AI Infrastructure Toolkit
Search, compare, calculate, and map — free tools for evaluating datacenter GPUs and AI accelerators.
Compare 221+ datacenter GPUs across FP64, FP32, FP16, BF16, FP8, and more.
Browse GPUs Build custom GPU configurations and calculate aggregate TFLOPS, power draw, and memory. DGX, HGX, NVL, and multi-GPU clusters from NVIDIA, AMD, and more with aggregate
performance. Track 786 GPU datacenters worldwide from xAI, Meta, Microsoft, OpenAI,
and more.
Most Compared Accelerators
The datacenter GPUs drawing the most head-to-head comparisons from AI and HPC teams right now.
Memory 432 GB FP16 5,033 TFLOPS FP8 20,133 TFLOPS Rubin · 2026 Memory 288 GB FP16 4,000 TFLOPS FP8 17,500 TFLOPS CDNA 4 · 2025 Memory 288 GB TDP 1400W FP16 2,517 TFLOPS FP8 5,033 TFLOPS Blackwell Ultra · 2025 Memory 288 GB FP16 2,500 TFLOPS FP8 5,000 TFLOPS Blackwell · 2024 Memory 186 GB FP16 2,500 TFLOPS FP8 5,000 TFLOPS CDNA 4 · 2025 Memory 288 GB TDP 1000W FP16 2,310 TFLOPS FP8 4,614 TFLOPSNVIDIA
Rubin SXM
AMD
Instinct MI355X OAM
NVIDIA
GB300 NVL72 GPU 288GB
NVIDIA
GB200 NVL72 GPU 186GB
AMD
Instinct MI350X OAM
Go Deeper
Architecture breakdowns, generation comparisons, and selection guides for infrastructure buyers.
A $3,999 Strix Halo desktop with 128GB unified memory takes on Nvidia's DGX Spark. It
holds huge models — but a 256 GB/s bus caps how fast it serves them. Guide The M5 Max matches a V100 at 7x the power efficiency. Unified memory, M1–M5 specs,
and how SoC GPUs compare to datacenter accelerators. H200 vs B200 architecture comparison. Performance gains, memory improvements, and which to
choose for training vs inference. From Pascal to Blackwell. Understand superchips, Tensor Cores, multi-die designs, and how
GPU architectures have evolved. TFLOPS, memory, TDP, precision types explained. Use case recommendations for LLM training,
inference, and HPC workloads.AMD Ryzen AI Halo Review
Apple Silicon Explained
Hopper vs Blackwell
NVIDIA Architectures Explained
How to Choose a Datacenter GPU
Ready to Compare?
Search 221+ datacenter GPUs. Filter by precision, vendor, architecture, and form factor. Compare specs side by side.
© 2026 Flopper.io - Compare the GPUs Powering AI
来源：Hacker News · flopper.io