02Hardware

The chips doing the thinking

Every accelerator that matters — specs normalized to dense FLOPS so vendors’ sparsity-inflated headlines compare honestly, plus what a FLOP actually rents for.

Accelerator database

Dense tensor throughput (vendor “sparse” headline figures are 2× these). Open a chip’s context note for source-specific details.

AI accelerator specifications

Instinct MI400, AMDest — estimated / secondary-source figure Context for Instinct MI400 AMD · CDNA-Next with HBM4; AMD quotes 40 PF FP4 with sparsity. OpenAI is the anchor customer (6 GW agreement).	Dec 2026	—	10 PF	20 PF	432 GB	19.6 TB/s	—	Announced
Rubin (VR200), NVIDIAest — estimated / secondary-source figure Context for Rubin (VR200) NVIDIA · HBM4 + Vera CPU pairing. NVIDIA quotes ~50 PF NVFP4 per package (~25 PF dense-equivalent). “In full production” Jan 2026; volume H2 2026.	Sep 2026	—	—	25 PF	288 GB	20 TB/s	1800 W	Announced
Maia 200, Microsoftest — estimated / secondary-source figure Context for Maia 200 Microsoft · Microsoft’s inference chip, live in Azure Jan 2026; serves GPT-5.x-class models at claimed 30% better perf/$.	Jan 2026	—	5 PF	10 PF	216 GB	7 TB/s	750 W	Shipping
Trainium3, AWSest — estimated / secondary-source figure Context for Trainium3 AWS · GA Dec 2025; 4.4× the performance of a Trn2 UltraServer at 4× perf/W. Trainium4 already announced.	Dec 2025	1.26 PF	2.52 PF	—	144 GB	4.9 TB/s	—	Shipping
TPU v7 (Ironwood), Googleest — estimated / secondary-source figure Context for TPU v7 (Ironwood) Google · GA on Google Cloud Mar 2026; no public list price — capacity heavily pre-committed (Anthropic: up to 1M chips).	Nov 2025	—	4.61 PF	—	192 GB	7.37 TB/s	—	Ramping
Instinct MI355X, AMDest — estimated / secondary-source figure Context for Instinct MI355X AMD · CDNA 4, liquid-cooled; AMD quotes 20.1 PF FP4/FP6 — on paper the FLOPS-per-dollar leader of 2026.	Oct 2025	5 PF	10.1 PF	20.1 PF	288 GB	8 TB/s	1400 W	Shipping
B300 (Blackwell Ultra), NVIDIAest — estimated / secondary-source figure Context for B300 (Blackwell Ultra) NVIDIA · 1.5× the FP4 inference of B200 with 288 GB HBM3e; first GB300 NVL72 racks deployed Jul 2025, volume through 2026.	Jul 2025	2.25 PF	4.5 PF	15 PF	288 GB	8 TB/s	1400 W	Shipping
GeForce RTX 5090, NVIDIA Context for GeForce RTX 5090 NVIDIA · Consumer Blackwell with FP4 support.	Jan 2025	419 TF	838 TF	1.68 PF	32 GB	1.79 TB/s	575 W	Shipping
B200 (HGX), NVIDIAest — estimated / secondary-source figure Context for B200 (HGX) NVIDIA · Dual-die Blackwell; first NVIDIA part with FP4 tensor formats.	Dec 2024	2.25 PF	4.5 PF	9 PF	192 GB	8 TB/s	1000 W	Shipping
Instinct MI325X, AMDest — estimated / secondary-source figure Context for Instinct MI325X AMD · MI300X refresh with HBM3e.	Dec 2024	1.31 PF	2.61 PF	—	256 GB	6 TB/s	1000 W	Shipping
Trainium2, AWSest — estimated / secondary-source figure Context for Trainium2 AWS · AWS in-house silicon; >500K deployed for Anthropic’s Project Rainier.	Dec 2024	650 TF	1.3 PF	—	96 GB	2.9 TB/s	500 W	Shipping
TPU v6e (Trillium), Google Context for TPU v6e (Trillium) Google · 4.7× the peak compute of v5e.	Oct 2024	918 TF	—	—	32 GB	1.64 TB/s	—	Shipping
Gaudi 3, Intel Context for Gaudi 3 Intel · Priced aggressively; FP8 and BF16 share the same MME throughput.	Sep 2024	1.84 PF	1.84 PF	—	128 GB	3.7 TB/s	900 W	Shipping
H200 SXM, NVIDIAest — estimated / secondary-source figure Context for H200 SXM NVIDIA · Same compute as H100; HBM3e lifts capacity 1.8× and bandwidth 1.4×.	Jun 2024	989 TF	1.98 PF	—	141 GB	4.8 TB/s	700 W	Shipping
Instinct MI300X, AMDest — estimated / secondary-source figure Context for Instinct MI300X AMD · CDNA 3 chiplet design; led on memory capacity at launch.	Dec 2023	1.31 PF	2.61 PF	—	192 GB	5.3 TB/s	750 W	Shipping
TPU v5p, Google Context for TPU v5p Google · Trained Gemini 1.0.	Dec 2023	459 TF	—	—	95 GB	2.77 TB/s	—	Shipping
TPU v5e, Google Context for TPU v5e Google · Efficiency-tier TPU; INT8 394 TOPS.	Aug 2023	197 TF	—	—	16 GB	0.82 TB/s	—	Shipping
H100 SXM, NVIDIA Context for H100 SXM NVIDIA · Defined the 2023–25 training buildout; the market’s reference chip.	Oct 2022	989 TF	1.98 PF	—	80 GB	3.35 TB/s	700 W	Shipping
GeForce RTX 4090, NVIDIA Context for GeForce RTX 4090 NVIDIA · Consumer anchor: the cheapest FLOPS money can buy, minus interconnect.	Oct 2022	330 TF	661 TF	—	24 GB	1.01 TB/s	450 W	Shipping
A100 SXM 80GB, NVIDIA Context for A100 SXM 80GB NVIDIA · Workhorse of the GPT-3→GPT-4 era.	Nov 2020	312 TF	—	—	80 GB	2.04 TB/s	400 W	Legacy
V100 SXM2, NVIDIA Context for V100 SXM2 NVIDIA · First tensor-core GPU; FP16 only (no BF16).	Jun 2017	125 TF	—	—	32 GB	0.9 TB/s	300 W	Legacy

What a PFLOP-hour rents for

Cheapest specialist-cloud on-demand ÷ dense BF16 throughput

Static chart frame · exact values in table view

Rental quotes as of Jul 2026; dense BF16 from vendor specs. GB200 row uses per-GPU rack pricing.

The memory-bandwidth race

HBM bandwidth by launch date — hollow points are announced parts

Static chart frame · exact values in table view

Vendor spec sheets; consumer cards excluded. Rubin/MI400 figures are announced targets.