compare

Every chip, one schema

ASICs and GPUs side by side. Sort by any column. Every number links to the vendor's own documentation.


TPU7x Ironwoodgoogle	2025	192 GiB	7380	2307 TFLOPS	4614	n/a	9216 pod	224	960	Intel Emerald Rapids	n/a	cloud-only
NVIDIA B200 SXMnvidia	2024	180 GiB	7750	2250 TFLOPS	4500	1000	8 HGX B200 8-GPU	n/a	n/a	n/a	$5.98	cloud-and-purchasable
AWS Trainium3aws	2025	144 GiB	4900	671 TFLOPS	2517	n/a	144 server	n/a	n/a	n/a	n/a	cloud-only
NVIDIA H200 SXMnvidia	2023	141 GiB	4800	989.5 TFLOPS	1979	700	8 HGX/DGX 8-GPU	n/a	n/a	n/a	$3.59	cloud-and-purchasable
Intel Gaudi 3 (HL-325)intel	2024	128 GiB	3700	1835 TFLOPS	1835	900	8 card	160	1792	Intel Xeon	$125000	cloud-and-purchasable
AWS Trainium2aws	2024	96 GiB	2900	667 TFLOPS	1299	n/a	16 Trn2.48xlarge	192	2048	Intel Xeon Sapphire Rapids	n/a	cloud-only
Intel Gaudi 2 (HL-225)intel	2022	96 GiB	2450	432 TFLOPS	865	600	8 card	n/a	n/a	n/a	$65000	cloud-and-purchasable
NVIDIA RTX PRO 6000 Blackwellnvidia	2025	96 GiB	1792	n/a	n/a	600	1 card	n/a	n/a	n/a	$1.8	purchasable
TPU v5pgoogle	2024	95 GiB	2765	459 TFLOPS	459	n/a	8960 pod	208	448	Intel Sapphire Rapids	$4.2	cloud-only
NVIDIA H100 SXMnvidia	2022	80 GiB	3350	989.5 TFLOPS	1979	700	8 HGX/DGX 8-GPU	n/a	n/a	n/a	$2.69	cloud-and-purchasable
NVIDIA RTX 6000 Ada Generationnvidia	2023	48 GiB	960	n/a	n/a	300	1 card	n/a	n/a	n/a	$0.77	purchasable
TPU v3google	2018	32 GiB	900	123 TFLOPS	n/a	220	1024 pod	n/a	n/a	n/a	n/a	cloud-only
TPU v4google	2022	32 GiB	1200	275 TFLOPS	n/a	170	4096 pod	n/a	n/a	n/a	$3.22	cloud-only
TPU v6e Trilliumgoogle	2024	32 GiB	1638	918 TFLOPS	n/a	n/a	256 pod	44	176	AMD EPYC Genoa	n/a	cloud-only
AWS Trainiumaws	2022	32 GiB	820	190 TFLOPS	190	n/a	16 trn1.32xlarge	8	32	Intel Xeon Ice Lake 8375C	$1.344	cloud-only
AWS Inferentia2aws	2023	32 GiB	820	190 TFLOPS	n/a	n/a	12 inf2.48xlarge	4	16	AMD EPYC 7R13	$0.76	cloud-only
Intel Gaudi (HL-2000)intel	2019	32 GiB	1000	n/a	n/a	300	8 card	96	768	Intel Xeon Platinum 8275CL	$13.10904	cloud-and-purchasable
GeForce RTX 5090nvidia	2025	32 GiB	1792	n/a	n/a	575	1 card	n/a	n/a	n/a	$0.99	purchasable
GeForce RTX 4090nvidia	2022	24 GiB	1008	n/a	n/a	450	1 card	n/a	n/a	n/a	$0.69	purchasable
TPU v5egoogle	2023	16 GiB	800	197 TFLOPS	n/a	n/a	256 pod	24	48	n/a	$1.2	cloud-only
TPU v2google	2017	8 GiB	600	45 TFLOPS	n/a	n/a	512 pod	n/a	n/a	n/a	n/a	cloud-only
TPU v1google	2016	24 MiB	n/a	n/a	n/a	40	1 card	n/a	n/a	n/a	n/a	internal-only

Underlined values link to their primary source. Scale unit is the largest tightly-coupled group of chips, a single card, a GPU node, or a full TPU pod. Compute is stored dense: NVIDIA headline tensor numbers include 2:4 sparsity, so we halve them to the dense value for a fair comparison.