Directory
Parts taxonomy
The component layer the systems graph draws its edges into. Every part page lists the systems using it, so a question like “what is installed underneath Slingshot” has an answer you can click to rather than assemble by hand.
Accelerators Cooling CPUs Facility Interconnect Memory Networking Operating systems Power Racks Storage
Accelerators
| Part | Supplier | Generation | Released | Systems | Rmax underneath |
|---|---|---|---|---|---|
| AMD Instinct MI355X | AMD | CDNA 4 | 2026 | 2 | 34.77 PFlop/s |
| Google TPU 8i TPU (systolic array) | Alphabet | TPU 8 | 2026 | 1 | - |
| Google TPU 8t TPU (systolic array) | Alphabet | TPU 8 | 2026 | 1 | - |
| Microsoft Maia 200 Maia · TSMC 3nm | Microsoft | Maia | 2026 | - | - |
| AWS Trainium3 NeuronCore-v4 | Amazon | Trainium | 2025 | 1 | - |
| Google TPU7x (Ironwood) TPU (systolic array) | Alphabet | TPU7x | 2025 | 4 | - |
| Huawei CloudMatrix 384 Da Vinci | Huawei | CloudMatrix | 2025 | 2 | - |
| NVIDIA B300 SXM CUDA (Blackwell Ultra) · TSMC 4NP | NVIDIA | Blackwell Ultra | 2025 | 1 | - |
| NVIDIA GB300 NVL72 | NVIDIA | Blackwell Ultra | 2025 | 13 | - |
| AMD Instinct MI325X ROCm (CDNA 3) · TSMC N5/N6 | AMD | CDNA 3 | 2024 | 1 | - |
| AWS Trainium2 NeuronCore-v3 · TSMC 5nm | Amazon | Trainium | 2024 | 4 | - |
| Cerebras WSE-3 Cerebras (wafer-scale) · 900000C · TSMC N5 | Cerebras | Wafer Scale Engine | 2024 | 4 | - |
| Google TPU v6e (Trillium) TPU (systolic array) | Alphabet | v6e | 2024 | 4 | - |
| Intel Gaudi 3 Gaudi · TSMC N5 | Habana Labs | Gaudi | 2024 | - | - |
| Meta MTIA v2 MTIA | Meta Platforms | MTIA | 2024 | - | - |
| NVIDIA B200 SXM CUDA (Blackwell) · TSMC 4NP | NVIDIA | Blackwell | 2024 | 36 | 534.0 PFlop/s |
| NVIDIA GB200 NVL72 Arm CPU + CUDA GPU · TSMC 4NP | NVIDIA | Blackwell | 2024 | 14 | 27.37 PFlop/s |
| NVIDIA H200 SXM CUDA (Hopper GH100) · 16896C · 1.98 GHz · TSMC 4N | NVIDIA | Hopper | 2024 | 53 | 1.21 EFlop/s |
| AMD Instinct MI300A ROCm (CDNA 3) · TSMC N5/N6 | AMD | CDNA 3 | 2023 | 15 | 2.81 EFlop/s |
| AMD Instinct MI300X ROCm (CDNA 3) · TSMC N5/N6 | AMD | CDNA 3 | 2023 | 3 | 117.6 PFlop/s |
| Google TPU v5e TPU (systolic array) | Alphabet | v5e | 2023 | 1 | - |
| Google TPU v5p TPU (systolic array) · 1.75 GHz | Alphabet | TPU v5 | 2023 | 5 | - |
| Groq LPU Groq TSP · 14 nm | Groq | LPU | 2023 | 3 | - |
| Intel Data Center GPU Max 1550 oneAPI (Xe-HPC) · Intel 7 / TSMC N5 | Intel | Ponte Vecchio | 2023 | 5 | 1.05 EFlop/s |
| Microsoft Maia 100 Maia · TSMC N5 | Microsoft | Maia | 2023 | - | - |
| NVIDIA GH200 Grace Hopper Superchip Arm CPU + CUDA GPU · TSMC 4N | NVIDIA | Grace Hopper | 2023 | 28 | 2.49 EFlop/s |
| NVIDIA H100 SXM5 94GB CUDA (Hopper GH100) · 16896C · 1.98 GHz · TSMC 4N | NVIDIA | Hopper | 2023 | 11 | 220.6 PFlop/s |
| NVIDIA L40S CUDA (Ada Lovelace) | NVIDIA | Ada Lovelace | 2023 | 20 | - |
| AMD Instinct MI210 ROCm (CDNA 2) · TSMC N6 | AMD | CDNA 2 | 2022 | 5 | 63.21 PFlop/s |
| NVIDIA H100 80GB PCIe CUDA (Hopper) | NVIDIA | Hopper | 2022 | 4 | 9.17 PFlop/s |
| NVIDIA H100 SXM5 CUDA (Hopper GH100) · 16896C · 1.98 GHz · TSMC 4N | NVIDIA | Hopper | 2022 | 81 | 2.21 EFlop/s |
| AMD Instinct MI250X ROCm (CDNA 2) · 1.7 GHz · TSMC N6 | AMD | CDNA 2 | 2021 | 15 | 2.39 EFlop/s |
| NUDT Matrix-3000 NUDT custom | NUDT | Matrix | 2021 | - | - |
| SambaNova RDU Reconfigurable dataflow · TSMC 5nm | SambaNova | RDU | 2021 | 1 | - |
| Tesla Dojo D1 Custom Dojo ISA (64-bit scalar + 64-byte SIMD vector) · 354C · 2 GHz · 7 nm | Tesla | D1 (Dojo v1) | 2021 | 1 | - |
| Anton 3 ASIC 528C · 7 nm | D. E. Shaw Research | Anton 3 | 2020 | 1 | - |
| Google TPU v4 | Alphabet | TPU v4 | 2020 | 2 | - |
| Graphcore IPU Graphcore IPU · TSMC 7nm | Graphcore | Colossus | 2020 | 1 | 7.94 PFlop/s |
| Hygon DCU ROCm-compatible | Hygon | DCU | 2020 | 1 | - |
| NVIDIA A100 SXM4 40GB CUDA (Ampere GA100) · 6912C · 1.41 GHz · 7 nm | NVIDIA | Ampere | 2020 | 45 | 482.9 PFlop/s |
| NVIDIA A100 SXM4 64GB CUDA (Ampere GA100) · 6912C · 1.41 GHz · 7 nm | NVIDIA | Ampere | 2020 | 2 | 244.7 PFlop/s |
| NVIDIA A100 SXM4 80GB | NVIDIA | Ampere | 2020 | 14 | 59.26 PFlop/s |
| Huawei Ascend 910 Da Vinci · 7 nm | Huawei | Da Vinci | 2019 | 3 | - |
| NEC SX-Aurora TSUBASA Vector Engine NEC vector · 8C · 1.6 GHz · 16 nm | NEC | SX-Aurora | 2018 | 5 | 51.28 PFlop/s |
| Preferred Networks MN-Core (1st generation) TSMC 12nm | Preferred Networks | MN-Core 1 | 2018 | 1 | 2.18 PFlop/s |
| NUDT Matrix-2000 NUDT custom · 128C · 1.2 GHz · 28 nm | NUDT | Matrix | 2017 | 1 | 61.44 PFlop/s |
| NVIDIA Tesla V100 CUDA (Volta GV100) · 5120C · 1.53 GHz · 12 nm | NVIDIA | Volta | 2017 | 91 | 529.0 PFlop/s |
| PEZY-SC2 MIPS64 (PE array) · 2048C · 0.7 GHz | PEZY Computing | SC2 | 2017 | 5 | 23.12 PFlop/s |
| Intel Xeon Phi 7250 (Knights Landing) x86-64 (Knights Landing) · 68C · 1.4 GHz · 14 nm | Intel | Knights Landing | 2016 | 6 | 58.29 PFlop/s |
| NVIDIA Tesla P100 CUDA (Pascal GP100) · 3584C · 1.48 GHz · 16 nm | NVIDIA | Pascal | 2016 | 28 | 70.56 PFlop/s |
| AMD FirePro S9150 TSMC 28nm | AMD | GCN 2.0 (Hawaii) | 2014 | 1 | 316.70 TFlop/s |
| Anton 2 ASIC 66C · 40 nm | D. E. Shaw Research | Anton 2 | 2014 | 1 | - |
| Intel Xeon Phi 31S1P x86-64 (Knights Corner) · 57C · 1.1 GHz · 22 nm | Intel | Knights Corner | 2013 | 2 | 63.51 PFlop/s |
| Intel Xeon Phi SE10P x86-64 (Knights Corner) · 61C | Intel | Knights Corner | 2012 | 1 | 5.17 PFlop/s |
| NVIDIA Tesla K20X CUDA (Kepler GK110) · 2688C · 0.732 GHz · 28 nm | NVIDIA | Kepler | 2012 | 11 | 24.90 PFlop/s |
| NVIDIA Tesla C2050 CUDA (Fermi) · 448C · 1.15 GHz · 40 nm | NVIDIA | Fermi | 2010 | 1 | 1.27 PFlop/s |
| NVIDIA Tesla M2050 CUDA (Fermi) · 448C · 1.15 GHz · 40 nm | NVIDIA | Fermi | 2010 | 4 | 3.98 PFlop/s |
| ATI Radeon HD 4870 X2 55 nm | AMD | TeraScale (RV770) | 2009 | 1 | 563.10 TFlop/s |
| IBM PowerXCell 8i Cell BE (PPE + 8 SPE) · 9C · 3.2 GHz · 65 nm | IBM | Cell BE | 2008 | 5 | 2.08 PFlop/s |
| AMD Instinct MI430X | AMD | Instinct MI400 | - | 4 | - |
| AMD Instinct MI455X CDNA | AMD | Instinct MI400 | - | - | - |
| Cerebras CS-2 (WSE-2) Cerebras (wafer-scale) | Cerebras | Wafer Scale Engine 2 | - | 5 | - |
| Huawei Ascend 910C Da Vinci | Huawei | Da Vinci | - | 1 | - |
| NVIDIA A100 (in DGX A100) | NVIDIA | Ampere | - | 13 | 95.65 PFlop/s |
| NVIDIA Vera Rubin | NVIDIA | Rubin | - | 10 | - |
Cooling
| Part | Supplier | Generation | Released | Systems | Rmax underneath |
|---|---|---|---|---|---|
| Direct-to-chip cold plate Cold plate on the package, air for the remainder · Removes roughly 50-80 percent of node heat to liquid | open standard | DLC | 2018 | 11 | 273.6 PFlop/s |
| Chilled-water indirect liquid cooling Closed-coupled liquid cooling, chiller-backed · 28 MW cooling capacity, Climaveneta chillers at Sunway TaihuLight | open standard | - | 2016 | 1 | 93.01 PFlop/s |
| Single-phase immersion Full submersion in a dielectric bath · 100 kW per tank and above | open standard | Immersion | 2015 | 5 | 43.39 PFlop/s |
| Direct liquid cooling, warm water Direct-to-chip cold plates, facility water loop · Removes 90+ percent of node heat to liquid | open standard | DLC | 2012 | 64 | 9.18 EFlop/s |
| Direct water-cooled LSI (pre-cold-plate era) Direct water cooling at the chip package | open standard | - | 2011 | 1 | 10.51 PFlop/s |
| Rear-door heat exchanger Liquid coil in the cabinet door · Roughly 30-60 kW per rack | open standard | RDHx | 2010 | - | - |
| Liquid-assisted air (in-cabinet coil) In-cabinet air, water-cooled coil · Roughly 20-45 kW per cabinet | open standard | Liquid-assisted air | 2007 | 6 | 74.75 PFlop/s |
| Conduction-cooled liquid (primary/secondary loop) Sealed primary loop at the package, secondary loop rejecting outdoors | open standard | - | 1993 | 1 | 124.0 GFlop/s |
| Air cooling (CRAC/CRAH) Room-level air handling · Practical ceiling around 15-20 kW per rack | open standard | Air | 1990 | 14 | 198.8 PFlop/s |
| ETA10 liquid nitrogen cooling | open standard | ETA10 | 1987 | 1 | - |
| Cray-2 Fluorinert immersion cooling | open standard | Cray-2 | 1985 | 1 | - |
CPUs
| Part | Supplier | Generation | Released | Systems | Rmax underneath |
|---|---|---|---|---|---|
| AMD EPYC 9575F 64C · 3.3 GHz | AMD | Zen 5 (Turin) | 2026 | 5 | 34.77 PFlop/s |
| LingKun LX2 Armv9.2 SVE2/SME · 304C · 1.55 GHz | open standard | LX2 | 2026 | 1 | 2.20 EFlop/s |
| Intel Xeon Platinum 8570 x86-64 · 56C · 2.1 GHz · Intel 7 | Intel | Emerald Rapids | 2025 | 13 | 489.5 PFlop/s |
| Loongson 3C6000 LoongArch · 64C · 2.2 GHz · 12 nm | Loongson Technology | LA664 | 2025 | - | - |
| SiPearl Rhea1 Armv8.4 (Neoverse V1) · 80C · TSMC N6 | SiPearl | Neoverse V1 | 2025 | - | - |
| AMD EPYC 9005 (Turin) x86-64 · 192C · 2.7 GHz · TSMC N3E | AMD | Zen 5 | 2024 | 14 | 374.5 PFlop/s |
| AWS Graviton4 Armv9 (Neoverse V2) · 96C · 2.8 GHz · TSMC 4nm | Amazon | Neoverse V2 | 2024 | - | - |
| Intel Xeon 6 6700E (Sierra Forest) x86-64 · 144C · 2.4 GHz · Intel 3 | Intel | Sierra Forest | 2024 | 1 | 4.53 PFlop/s |
| Intel Xeon 6 6900P (Granite Rapids) x86-64 · 128C · 2 GHz · Intel 3 | Intel | Granite Rapids | 2024 | 7 | 150.3 PFlop/s |
| Intel Xeon Gold 6548Y+ x86-64 · 32C · 2.5 GHz · Intel 7 | Intel | Emerald Rapids | 2024 | 3 | 44.37 PFlop/s |
| Intel Xeon Platinum 8558 x86-64 · 48C · 2.1 GHz · Intel 7 | Intel | Emerald Rapids | 2024 | 6 | 405.5 PFlop/s |
| AMD 4th Gen EPYC 24C (in MI300A) x86-64 · 24C · 1.8 GHz · TSMC N5 | AMD | Zen 4 | 2023 | 10 | 2.75 EFlop/s |
| Intel Xeon CPU Max 9470 x86-64 · 52C · 2 GHz · Intel 7 | Intel | Sapphire Rapids HBM | 2023 | 2 | 1.04 EFlop/s |
| Intel Xeon Gold 6430 x86-64 · 32C · 2.1 GHz · Intel 7 | Intel | Sapphire Rapids | 2023 | 1 | 106.2 PFlop/s |
| Intel Xeon Platinum 8460Y+ x86-64 · 40C · 2 GHz · Intel 7 | Intel | Sapphire Rapids | 2023 | 4 | 143.4 PFlop/s |
| Intel Xeon Platinum 8462Y+ x86-64 · 32C · 2.8 GHz · Intel 7 | Intel | Sapphire Rapids | 2023 | 10 | 246.2 PFlop/s |
| Intel Xeon Platinum 8468 x86-64 · 48C · 2.1 GHz · Intel 7 | Intel | Sapphire Rapids | 2023 | 11 | 346.1 PFlop/s |
| Intel Xeon Platinum 8480C x86-64 · 56C · 2 GHz · Intel 7 | Intel | Sapphire Rapids | 2023 | 27 | 1.53 EFlop/s |
| NVIDIA Grace Armv9 (Neoverse V2) · 72C · 3.1 GHz · TSMC 4N | NVIDIA | Grace | 2023 | 32 | 2.49 EFlop/s |
| AMD EPYC 9004 (Genoa) x86-64 · 96C · 2.4 GHz · TSMC N5 | AMD | Zen 4 | 2022 | 47 | 456.9 PFlop/s |
| AMD EPYC 7513 (Milan) x86-64 · 32C · 2.6 GHz · 7 nm | AMD | Zen 3 | 2021 | 5 | 21.79 PFlop/s |
| AMD EPYC 7713 (Milan) x86-64 · 64C · 2 GHz · 7 nm | AMD | Zen 3 | 2021 | 19 | 33.80 PFlop/s |
| AMD EPYC 7763 (Milan) x86-64 · 64C · 2.45 GHz · 7 nm | AMD | Zen 3 | 2021 | 27 | 216.8 PFlop/s |
| AMD EPYC 7A53 (Trento) x86-64 · 64C · 2 GHz · 7 nm | AMD | Zen 3 | 2021 | 12 | 2.32 EFlop/s |
| Ampere Altra Max Armv8.2 (Neoverse N1) · 128C · 3 GHz · 7 nm | Ampere Computing | Altra | 2021 | - | - |
| Intel Xeon Platinum 8358 x86-64 · 32C · 2.6 GHz · 10 nm | Intel | Ice Lake-SP | 2021 | 14 | 269.2 PFlop/s |
| Intel Xeon Platinum 8362 (Ice Lake) x86-64 · 32C · 2.8 GHz · Intel 10nm SuperFin | Intel | Ice Lake | 2021 | 3 | 4.65 PFlop/s |
| Sunway SW39000 Sunway (custom RISC) · 390C · 2.25 GHz | NRCPC | Sunway | 2021 | - | - |
| AMD EPYC 7V12 x86-64 · 48C · 2.45 GHz · 7 nm | AMD | Zen 2 (Rome) | 2020 | 6 | 150.4 PFlop/s |
| AMD EPYC 7402 (Rome) x86-64 · 24C · 2.8 GHz · 7 nm | AMD | Zen 2 | 2019 | 3 | 50.14 PFlop/s |
| AMD EPYC 7532 x86-64 · 32C · 2.4 GHz · 7 nm | AMD | Zen 2 (Rome) | 2019 | 2 | 25.81 PFlop/s |
| AMD EPYC 7742 (Rome) x86-64 · 64C · 2.25 GHz · 7 nm | AMD | Zen 2 | 2019 | 38 | 298.7 PFlop/s |
| AMD EPYC 7H12 (Rome, high frequency) x86-64 · 64C · 2.6 GHz · 7 nm | AMD | Zen 2 | 2019 | 10 | 53.14 PFlop/s |
| Fujitsu A64FX Armv8.2-A SVE · 48C · 2.2 GHz · 7 nm | Fujitsu | Armv8.2-A SVE | 2019 | 9 | 517.8 PFlop/s |
| Huawei Kunpeng 920 Armv8.2 · 64C · 2.6 GHz · 7 nm | Huawei | Kunpeng | 2019 | - | - |
| Intel Xeon Platinum 8260M (Cascade Lake) x86-64 · 24C · 2.4 GHz | Intel | Cascade Lake | 2019 | 1 | 2.18 PFlop/s |
| Intel Xeon Platinum 8280 x86-64 · 28C · 2.7 GHz · 14 nm | Intel | Cascade Lake | 2019 | 4 | 30.11 PFlop/s |
| Hygon Dhyana x86-64 · 32C · 2 GHz · 14 nm | Hygon | Zen (licensed) | 2018 | 1 | 4.33 PFlop/s |
| Intel Xeon Platinum 8174 x86-64 · 24C · 3.1 GHz · 14 nm | Intel | Skylake-SP | 2018 | 4 | 34.70 PFlop/s |
| IBM POWER9 22C POWER · 22C · 3.07 GHz · 14 nm | IBM | POWER9 | 2017 | 8 | 303.2 PFlop/s |
| Intel Xeon Gold 6148 x86-64 · 20C · 2.4 GHz · 14 nm | Intel | Skylake-SP | 2017 | 31 | 103.6 PFlop/s |
| Intel Xeon Platinum 8160 x86-64 · 24C · 2.1 GHz · 14 nm | Intel | Skylake-SP | 2017 | 5 | 37.27 PFlop/s |
| Phytium FT-2000+/64 Armv8 · 64C · 2.4 GHz · 16 nm | Phytium | FT-2000 | 2017 | - | - |
| Intel Xeon E5-2650 v4 x86-64 · 12C · 2.2 GHz · 14 nm | Intel | Broadwell | 2016 | 5 | 5.00 PFlop/s |
| Intel Xeon E5-2680 v4 x86-64 · 14C · 2.4 GHz · 14 nm | Intel | Broadwell-EP | 2016 | 12 | 26.92 PFlop/s |
| Intel Xeon E5-2683 v4 x86-64 · 16C · 2.1 GHz · 14 nm | Intel | Broadwell-EP | 2016 | 2 | 2.18 PFlop/s |
| Intel Xeon E5-2695 v4 x86-64 · 18C · 2.1 GHz · 14 nm | Intel | Broadwell-EP | 2016 | 23 | 38.82 PFlop/s |
| Sunway SW26010 Sunway (custom RISC) · 260C · 1.45 GHz · 28 nm | NRCPC | Sunway | 2016 | 1 | 93.01 PFlop/s |
| Intel Xeon D-1571 x86-64 · 16C · 1.3 GHz · 14 nm | Intel | Broadwell-DE | 2015 | 4 | 22.30 PFlop/s |
| Intel Xeon E5-2690 v3 x86-64 · 12C · 2.6 GHz · 22 nm | Intel | Haswell-EP | 2014 | 16 | 38.75 PFlop/s |
| Intel Xeon E5-2697 v3 x86-64 · 14C · 2.6 GHz · 22 nm | Intel | Haswell-EP | 2014 | 8 | 8.81 PFlop/s |
| Intel Xeon E5-2698 v3 x86-64 · 16C · 2.3 GHz · 22 nm | Intel | Haswell-EP | 2014 | 11 | 40.31 PFlop/s |
| Intel Xeon E5-2620 v2 (Ivy Bridge EP) x86-64 · 6C · 2.1 GHz | Intel | Ivy Bridge EP | 2013 | 1 | 125.20 TFlop/s |
| Intel Xeon E5-2690 v2 x86-64 · 10C · 3 GHz · 22 nm | Intel | Ivy Bridge-EP | 2013 | 3 | 914.16 TFlop/s |
| Intel Xeon E5-2692 v2 x86-64 · 12C · 2.2 GHz · 22 nm | Intel | Ivy Bridge-EP | 2013 | 2 | 63.51 PFlop/s |
| Intel Xeon E5-2697 v2 x86-64 · 12C · 2.7 GHz · 22 nm | Intel | Ivy Bridge-EP | 2013 | 12 | 8.76 PFlop/s |
| Intel Xeon E5-2670 x86-64 · 8C · 2.6 GHz · 32 nm | Intel | Sandy Bridge-EP | 2012 | 48 | 19.46 PFlop/s |
| Intel Xeon E5-2680 x86-64 · 8C · 2.7 GHz · 32 nm | Intel | Sandy Bridge-EP | 2012 | 5 | 11.01 PFlop/s |
| AMD Opteron 6274 (Interlagos) x86-64 · 16C · 2.2 GHz · 32 nm | AMD | Bulldozer | 2011 | 2 | 17.59 PFlop/s |
| AMD Opteron 6276 (Interlagos) x86-64 · 16C · 2.3 GHz · 32 nm | AMD | Bulldozer | 2011 | 6 | 2.25 PFlop/s |
| Fujitsu SPARC64 IXfx SPARC64 · 16C · 1.848 GHz | Fujitsu | SPARC64 | 2011 | 2 | 1.21 PFlop/s |
| IBM PowerPC A2 (Power BQC) PowerPC A2 · 18C · 1.6 GHz · 45 nm | IBM | BlueGene/Q | 2011 | 15 | 43.81 PFlop/s |
| ShenWei SW1600 ShenWei (custom RISC) · 16C · 0.975 GHz | NRCPC | ShenWei | 2011 | 1 | 795.90 TFlop/s |
| Fujitsu SPARC64 VIIIfx SPARC64 · 8C · 2 GHz · 45 nm | Fujitsu | SPARC64 | 2010 | 1 | 10.51 PFlop/s |
| Intel Xeon X5650 x86-64 · 6C · 2.66 GHz · 32 nm | Intel | Westmere-EP | 2010 | 8 | 2.00 PFlop/s |
| Intel Xeon X5670 x86-64 · 6C · 2.93 GHz · 32 nm | Intel | Westmere-EP | 2010 | 10 | 6.38 PFlop/s |
| AMD Opteron 2435 (Istanbul) x86-64 · 6C · 2.6 GHz · 45 nm | AMD | K10 | 2009 | 1 | 1.76 PFlop/s |
| Intel Xeon E5540 x86-64 · 4C · 2.53 GHz · 45 nm | Intel | Nehalem-EP | 2009 | 1 | 563.10 TFlop/s |
| AMD Opteron Quad-Core 1.9GHz x86-64 · 4C · 1.9 GHz | AMD | K10 (Barcelona/Shanghai) | 2008 | 1 | 180.60 TFlop/s |
| AMD Opteron (Barcelona) 2.0GHz x86-64 · 4C · 2 GHz · 65 nm | AMD | K10 Barcelona | 2007 | 1 | 326.00 TFlop/s |
| IBM PowerPC 450 850MHz PowerPC · 4C · 0.85 GHz · 90 nm | IBM | BlueGene/P | 2007 | 8 | 1.01 PFlop/s |
| Intel Xeon E5450 x86-64 · 4C · 3 GHz · 45 nm | Intel | Harpertown | 2007 | 6 | 875.78 TFlop/s |
| AMD Opteron 2210 (Santa Rosa) x86-64 · 2C · 1.8 GHz · 90 nm | AMD | K8 | 2006 | 1 | 1.03 PFlop/s |
| Intel Xeon 5160 x86-64 · 2C · 3 GHz · 65 nm | Intel | Woodcrest | 2006 | 1 | 9.60 TFlop/s |
| IBM PowerPC 440 700MHz PowerPC · 2C · 0.7 GHz · 130 nm | IBM | BlueGene/L | 2004 | 11 | 767.28 TFlop/s |
| NEC SX-6 vector processor NEC vector · 1C · 0.5 GHz · 150 nm | NEC | SX | 2002 | 2 | 36.11 TFlop/s |
| IBM POWER3-II POWER · 1C · 0.375 GHz · 220 nm | IBM | POWER3 | 1999 | 3 | 8.92 TFlop/s |
| DEC Alpha 21264 (EV68) Alpha · 1C · 1.25 GHz · 180 nm | Digital Equipment Corporation | Alpha | 1998 | 3 | 22.32 TFlop/s |
| IBM PowerPC 604e PowerPC · 1C · 0.332 GHz · 350 nm | IBM | PowerPC 604e | 1998 | 1 | 2.14 TFlop/s |
| SGI Origin 2000 (MIPS R10000) MIPS IV · 1C · 0.25 GHz · 350 nm | SGI | MIPS R10000 | 1998 | 1 | 1.61 TFlop/s |
| Hitachi SR2201 processing unit PA-RISC 1.1 · 1C · 0.15 GHz | Hitachi | SR2201 | 1996 | 3 | 610.8 GFlop/s |
| Intel Pentium Pro 200 MHz x86 (P6) · 1C · 0.2 GHz · 350 nm | Intel | P6 | 1996 | 1 | 1.07 TFlop/s |
| Fujitsu VPP500 vector processing element Fujitsu vector · 1C · 0.105 GHz · GaAs | Fujitsu | VPP500 | 1993 | 2 | 222.9 GFlop/s |
| Thinking Machines CM-5 vector node SPARC + custom vector · 1C · 0.032 GHz | Thinking Machines | Connection Machine | 1993 | 4 | 120.5 GFlop/s |
| Intel i860/XP i860 RISC · 1C · 0.05 GHz · 1000 nm | Intel | i860 | 1991 | 3 | 285.7 GFlop/s |
| Cray Y-MP CPU Cray Y-MP (custom) · 1C · 0.1667 GHz | Cray | Cray Y-MP | 1988 | 1 | - |
| ETA10 CPU ETA10 (custom CMOS) · 1C · 0.095 GHz | ETA Systems | ETA10 | 1987 | 1 | - |
| Thinking Machines CM-2 processor array 1-bit SIMD + Weitek FPU · 65536C | Thinking Machines | Connection Machine CM-2 | 1987 | 1 | - |
| Cray-2 CPU Cray-2 (custom) · 1C · 0.244 GHz | Cray | Cray-2 | 1985 | 1 | - |
| Cray X-MP CPU Cray X-MP (custom) · 1C · 0.105 GHz | Cray | Cray X-MP | 1982 | 1 | - |
| Cray-1 CPU Cray-1 (custom) · 1C · 0.08 GHz | Cray | Cray-1 | 1976 | 3 | - |
| AMD 3rd Gen EPYC 64C (Andromeda hosts) | AMD | Milan | - | 1 | - |
Facility
| Part | Supplier | Generation | Released | Systems | Rmax underneath |
|---|---|---|---|---|---|
| Dry cooler Closed-loop air-side heat rejection | open standard | - | 2005 | - | - |
| Medium-voltage switchgear Utility-to-building electrical boundary · Typically tens of MW at a large campus | open standard | - | 2000 | - | - |
| Evaporative cooling tower Open or closed evaporative heat rejection | open standard | - | 1990 | - | - |
Interconnect
| Part | Supplier | Generation | Released | Systems | Rmax underneath |
|---|---|---|---|---|---|
| LingQi LQLink Dual-plane multi-rail fat-tree · 1.6 Tb/s per node · ~1 microsecond | open standard | LQLink | 2026 | 1 | 2.20 EFlop/s |
| NVIDIA InfiniBand XDR (Quantum-X800) Fat tree or dragonfly+ · 800 Gb/s per port | NVIDIA | XDR | 2024 | 9 | 19.80 PFlop/s |
| NVIDIA NVLink 5 / NVSwitch All-to-all within an NVLink domain · 1.8 TB/s per GPU | NVIDIA | NVLink 5 | 2024 | 3 | 5.16 PFlop/s |
| Broadcom RoCE Ethernet | Broadcom | Tomahawk/NetXtreme | 2023 | 5 | 205.7 PFlop/s |
| NVIDIA InfiniBand NDR (Quantum-2) Fat tree or dragonfly+ · 400 Gb/s per port · Under 0.6 microseconds | NVIDIA | NDR | 2022 | 149 | 5.03 EFlop/s |
| HPE Slingshot-11 Dragonfly · 200 Gb/s per port · Around 2 microseconds | Hewlett Packard Enterprise | Slingshot | 2021 | 56 | 7.67 EFlop/s |
| HPE Slingshot-10 Dragonfly · 100 Gb/s per port | Hewlett Packard Enterprise | Slingshot | 2020 | 11 | 102.4 PFlop/s |
| MN-Core DirectConnect 112 Gb/s per link | Preferred Networks | DirectConnect | 2020 | 1 | 2.18 PFlop/s |
| AWS Elastic Fabric Adapter Cloud fabric · Up to 3.2 Tb/s per instance · Low, but not deterministic | Amazon | EFA | 2019 | 3 | - |
| Fujitsu Tofu Interconnect D 6D mesh/torus · 6.8 GB/s per link per direction | Fujitsu | Tofu D | 2019 | 7 | 513.9 PFlop/s |
| Huawei HCCS Mesh within a node | Huawei | HCCS | 2019 | 1 | - |
| NUDT TH Express-3 Fat tree | NUDT | TH Express | 2019 | - | - |
| NVIDIA InfiniBand HDR Fat tree or dragonfly+ · 200 Gb/s per port · About 0.6 microseconds | NVIDIA | HDR | 2019 | 171 | 1.34 EFlop/s |
| Eviden BXI v2 Fat tree or dragonfly+ · 100 Gb/s per port | Eviden | BXI | 2016 | 2 | 114.0 PFlop/s |
| Sunway interconnect Two-level fat tree · 16 GB/s per node · About 1 microsecond | NRCPC | Sunway | 2016 | 1 | 93.01 PFlop/s |
| Omni-Path Architecture Fat tree · 100 Gb/s per port · About 1 microsecond | Cornelis Networks | OPA | 2015 | 79 | 250.8 PFlop/s |
| Mellanox InfiniBand EDR Fat tree · 100 Gb/s per 4x port · About 0.9 microseconds | NVIDIA | EDR | 2014 | 83 | 543.6 PFlop/s |
| NUDT TH Express-2 Fat tree · 160 Gb/s per link · About 1.6 microseconds | NUDT | TH Express | 2013 | 2 | 63.51 PFlop/s |
| Cray Aries Dragonfly · About 10 GB/s per direction · Around 1 microsecond | Cray | Aries | 2012 | 51 | 154.4 PFlop/s |
| IBM BlueGene/Q 5D torus 5D torus · 2 GB/s per link per direction | IBM | BlueGene/Q | 2011 | 14 | 43.09 PFlop/s |
| Mellanox InfiniBand FDR Fat tree · 56 Gb/s per 4x port · About 1 microsecond | NVIDIA | FDR | 2011 | 125 | 105.0 PFlop/s |
| Cray Gemini 3D torus · About 6 GB/s per direction per link · Around 1 microsecond | Cray | Gemini | 2010 | 16 | 25.44 PFlop/s |
| Fujitsu Tofu Interconnect 6D mesh/torus · 5 GB/s per link per direction | Fujitsu | Tofu | 2010 | 3 | 11.72 PFlop/s |
| NUDT Arch interconnect Hybrid fat tree, optoelectronic · 160 Gb/s bidirectional per link · 1.57 microseconds per hop | NUDT | Arch | 2010 | 1 | 2.57 PFlop/s |
| Mellanox InfiniBand QDR Fat tree · 40 Gb/s per 4x port | NVIDIA | QDR | 2008 | 75 | 22.54 PFlop/s |
| Cray SeaStar2+ 3D torus · 9.6 GB/s per direction | Cray | SeaStar | 2007 | 1 | 1.76 PFlop/s |
| Sun Magnum InfiniBand DDR Switch full non-blocking fat tree · DDR 4x, 20Gb/s per port | Sun Microsystems | - | 2007 | 1 | 326.00 TFlop/s |
| Mellanox InfiniBand DDR Fat tree · 20 Gb/s per 4x port · About 2 microseconds | NVIDIA | DDR | 2006 | 22 | 2.28 PFlop/s |
| IBM BlueGene 3D torus 3D torus, plus separate collective and barrier networks · 175 MB/s per link per direction | IBM | BlueGene | 2004 | 13 | 1.70 PFlop/s |
| NEC Earth Simulator crossbar Single-stage full crossbar · 12.3 GB/s per direction per node | NEC | SX | 2002 | 1 | 35.86 TFlop/s |
| IBM SP Switch2 Multi-stage (omega) switch · 500 MB/s per direction bidirectional | IBM | SP | 2000 | 1 | 4.94 TFlop/s |
| Quadrics QsNet Fat tree · 340 MB/s per direction · About 2 microseconds | Quadrics | QsNet | 1999 | 8 | 116.53 TFlop/s |
| HIPPI 800 Mbit/s per HIPPI-800 link | SGI | - | 1998 | 1 | 1.61 TFlop/s |
| IBM SP Switch multi-stage Omega network · 150 MB/s per node | IBM | SP Switch (TB3) | 1998 | 3 | 6.13 TFlop/s |
| Hitachi 3D hyper-crossbar 3D hyper-crossbar · 300 MB/s point to point | Hitachi | SR2201 | 1996 | 3 | 610.8 GFlop/s |
| Intel ASCI Red mesh 2D mesh · 400 MB/s per direction | Intel | Paragon | 1996 | 1 | 1.07 TFlop/s |
| Fujitsu VPP500 crossbar Full crossbar · 400 MB/s per direction | Fujitsu | VPP500 | 1993 | 2 | 222.9 GFlop/s |
| Thinking Machines CM-5 fat tree Fat tree · 20 MB/s per link | Thinking Machines | Connection Machine | 1993 | 1 | 59.7 GFlop/s |
| Intel Paragon 2D mesh 2D mesh with torus wraparound · 200 MB/s per link | Intel | Paragon | 1992 | 3 | 285.7 GFlop/s |
| Thinking Machines CM-2 hypercube network Hypercube | Thinking Machines | Connection Machine CM-2 | 1987 | 1 | - |
| Google TPU inter-chip interconnect (3D torus) 3D torus · 1,200 GB/s bidirectional per chip (v5p and TPU7x) | Alphabet | ICI | - | 4 | - |
Memory
| Part | Supplier | Generation | Released | Systems | Rmax underneath |
|---|---|---|---|---|---|
| HBM3E JEDEC HBM3E (extended) · Up to 36 GB per stack · Over 1.2 TB/s per stack | open standard | HBM3E | 2024 | - | - |
| Micron HBM3E 8-High JEDEC HBM3E (extended) · 24 GB per stack · About 1.2 TB/s per stack | Micron | HBM3E | 2024 | - | - |
| SK hynix HBM3E 12-High JEDEC HBM3E (extended) · 36 GB per stack · Over 1.2 TB/s per stack | SK hynix | HBM3E | 2024 | - | - |
| HBM3 JEDEC HBM3 · Up to 24 GB per stack · About 819 GB/s per stack | open standard | HBM3 | 2022 | - | - |
| DDR5 RDIMM JEDEC DDR5 · Typically 32-128 GB per module · Up to about 51 GB/s per module at DDR5-6400 | open standard | DDR5 | 2021 | - | - |
| LPDDR5X JEDEC LPDDR5X · Up to 480 GB per Grace socket · Up to 512 GB/s per Grace socket | open standard | LPDDR5X | 2021 | - | - |
| HBM2E JEDEC HBM2E · 16 GB per stack · About 460 GB/s per stack | open standard | HBM2E | 2020 | - | - |
| HBM2 JEDEC HBM2 · Up to 8 GB per stack · About 256 GB/s per stack | open standard | HBM2 | 2016 | - | - |
| GDDR5 JEDEC GDDR5, board-mounted, not stacked · 3-8 GB per accelerator in this era · 150-320 GB/s per accelerator | open standard | GDDR5 | 2008 | - | - |
Networking
| Part | Supplier | Generation | Released | Systems | Rmax underneath |
|---|---|---|---|---|---|
| Arista 7800R4 AI Modular Ethernet chassis · Up to 800G per port | Arista Networks | 7800R | 2024 | 1 | - |
| Broadcom Tomahawk 5 Ethernet switch ASIC · 51.2 Tb/s aggregate, 64 ports at 800G | Broadcom | Tomahawk | 2022 | - | - |
| NVIDIA Spectrum-4 Ethernet switch ASIC · 51.2 Tb/s aggregate, 64 ports at 800G | NVIDIA | Spectrum | 2022 | 3 | 41.50 PFlop/s |
Operating systems
| Part | Supplier | Generation | Released | Systems | Rmax underneath |
|---|---|---|---|---|---|
| HPE Cray Operating System SUSE-derived Linux, container-based management · Proprietary layer over open-source Linux | Hewlett Packard Enterprise | COS | 2019 | 27 | 4.05 EFlop/s |
| McKernel Lightweight multi-kernel (IHK-partitioned), co-resident with Linux · Research software, RIKEN-developed | RIKEN | McKernel | 2016 | 1 | 442.0 PFlop/s |
| Sunway RaiseOS Linux derivative for the Sunway architecture · Domestic distribution | NRCPC | RaiseOS | 2016 | 1 | 93.01 PFlop/s |
| Kylin Linux Linux distribution · State-endorsed Chinese distribution | KylinSoft | Kylin | 2010 | 5 | 2.26 EFlop/s |
| Windows HPC Server 2008 Windows Server kernel with HPC Pack job scheduler · Commercial license | Microsoft | Windows HPC Server | 2008 | 3 | 277.43 TFlop/s |
| Cray Linux Environment Linux, Cray-customised · Proprietary layer over open-source Linux | Cray | CLE | 2007 | 47 | 152.0 PFlop/s |
| Tri-Lab Operating System Stack Linux distribution, RHEL-derived, HPC-tuned · Open source, LLNL-maintained | Lawrence Livermore National Laboratory | TOSS | 2007 | 24 | 2.22 EFlop/s |
| CentOS Linux distribution, RHEL-compatible rebuild · Open source, community-maintained | The CentOS Project | CentOS | 2004 | 38 | 207.5 PFlop/s |
| IBM Compute Node Kernel Microkernel, no dynamic scheduling · Proprietary | IBM | CNK | 2004 | 11 | 31.48 PFlop/s |
| Ubuntu Server Linux distribution · Open source, free with paid support tiers | Canonical | Ubuntu LTS | 2004 | 52 | 1.77 EFlop/s |
| NEC SUPER-UX Unix (NEC SX-proprietary) · Proprietary | NEC | SUPER-UX | 2002 | 8 | 168.42 TFlop/s |
| Red Hat Enterprise Linux Linux distribution · Open source, commercial support subscription | Red Hat | RHEL | 2002 | 110 | 4.05 EFlop/s |
| SUSE Linux Enterprise Server Linux distribution · Open source, commercial support subscription | SUSE | SLES | 2000 | 34 | 123.9 PFlop/s |
| Compaq Tru64 UNIX Unix (OSF/1-derived) · Proprietary | Compaq | Tru64 UNIX | 1998 | 3 | 22.32 TFlop/s |
| IRIX | SGI | - | 1998 | 1 | 1.61 TFlop/s |
| Cougar Custom lightweight kernel, no Unix layer · Open source, government-authored | Sandia National Laboratories | Cougar | 1997 | 1 | 1.07 TFlop/s |
| Hitachi HI-UX/MPP Unix (Mach 3.0 microkernel), Hitachi-customised · Proprietary | Hitachi | HI-UX/MPP | 1996 | 8 | 5.72 TFlop/s |
| Fujitsu UXP Unix (SVR4-derived), Fujitsu-customised · Proprietary | Fujitsu | UXP | 1993 | 5 | 1.36 TFlop/s |
| SUNMOS Custom lightweight kernel, no Unix layer · Open source, government-authored | Sandia National Laboratories | SUNMOS | 1993 | 1 | 143.4 GFlop/s |
| Thinking Machines CMost Unix (TMC-proprietary variant) · Proprietary | Thinking Machines | CMost | 1991 | 4 | 120.5 GFlop/s |
| IBM AIX Unix (System V-derived) · Proprietary | IBM | AIX | 1986 | 20 | 586.99 TFlop/s |
| Rocky Linux RHEL-compatible Linux, community-maintained · Downstream rebuild of Red Hat Enterprise Linux source | open standard | - | - | 16 | 168.5 PFlop/s |
Power
| Part | Supplier | Generation | Released | Systems | Rmax underneath |
|---|---|---|---|---|---|
| Schneider Galaxy VX UPS Modular three-phase UPS, N+1 typical · Up to about 3.75 MW per frame, expandable | Schneider Electric | Galaxy | 2016 | - | - |
| Vertiv Liebert UPS Large-frame UPS, typically N+1 or 2N · Up to several MW per frame | Vertiv | Liebert | 2015 | 1 | 241.2 PFlop/s |
| Eaton Power Xpert busway Overhead busway, tap-off boxes to each row · Typically 800A to 5000A per run | Eaton | Power Xpert | 2014 | - | - |
Racks
| Part | Supplier | Generation | Released | Systems | Rmax underneath |
|---|---|---|---|---|---|
| Meta Grand Teton Integrated GPU server on OCP OpenRack | Meta Platforms | Grand Teton | 2022 | 2 | - |
| OCP Open Rack v3 Open rack, 48V busbar, front and rear liquid manifolds · Rated up to roughly 100 kW, higher with liquid options | open standard | ORv3 | 2022 | - | - |
| HPE Cray EX cabinet 100% direct-liquid-cooled blade cabinet · Roughly 400 kW per cabinet | Hewlett Packard Enterprise | Cray EX | 2019 | - | - |
Storage
| Part | Supplier | Generation | Released | Systems | Rmax underneath |
|---|---|---|---|---|---|
| Meta Tectonic Distributed storage, FUSE client | Meta Platforms | Tectonic | 2021 | 2 | - |
| Intel DAOS Object store, no POSIX layer underneath · 230 PB on Aurora · About 31 TB/s | Intel | DAOS | 2020 | 3 | 1.03 EFlop/s |
| HPE Cray ClusterStor E1000 Lustre appliance, flash and disk tiers · 700 PB on Frontier (Orion) · About 10 TB/s peak | Hewlett Packard Enterprise | ClusterStor | 2019 | 17 | 4.98 EFlop/s |
| VAST Data Platform Disaggregated shared-everything | VAST Data | VAST | 2019 | 17 | 237.7 PFlop/s |
| DDN EXAScaler Lustre appliance | DDN | EXAScaler | 2018 | 18 | 474.0 PFlop/s |
| WEKA Data Platform Parallel filesystem, software-defined | WEKA | WEKA | 2018 | 7 | 79.46 PFlop/s |
| IBM Spectrum Scale ESS (GPFS) Parallel filesystem (GPFS), erasure-coded declustered RAID · 250 PB on Summit (Alpine) · About 2.5 TB/s | IBM | Spectrum Scale | 2017 | 18 | 1.55 EFlop/s |
| Sunway TaihuLight storage network Storage network plus disk array · 20 PB · 288 GB/s total I/O bandwidth | NRCPC | - | 2016 | 1 | 93.01 PFlop/s |
| Cray ClusterStor 300 Lustre appliance, scale-out storage server units (SSUs) | Hewlett Packard Enterprise | ClusterStor 300 | 2014 | 1 | 10.68 PFlop/s |
| H2FS Hybrid virtual namespace over local + shared tiers, Lustre backend | NUDT | H2FS | 2013 | 1 | 61.44 PFlop/s |
| Fujitsu FEFS Lustre-based parallel filesystem · Roughly 150 PB across multiple volumes on Fugaku | Fujitsu | FEFS | 2011 | 4 | 453.7 PFlop/s |
| Ceph Distributed object storage, erasure-coded pools | open standard | Ceph | 2010 | 1 | 6.92 PFlop/s |
| DDN S2A9900 Block-storage SAN, parallel filesystem layered on top | DDN | S2A9900 | 2008 | 1 | 458.61 TFlop/s |
| Panasas ActiveScale (PanFS) Object storage, blade shelves (1 director + 10 storage blades) | Panasas | ActiveScale | 2008 | 4 | 1.14 PFlop/s |
| Lustre Parallel filesystem, separate metadata and object servers | open standard | Lustre | 2003 | 25 | 294.9 PFlop/s |
| NEC Earth Simulator tape library Tape library, staged behind disk · More than 1.5 PB | NEC | - | 2002 | 1 | 35.86 TFlop/s |
| Distributed RAID 5 disk array Direct-attached, no separate metadata server · Hundreds of GB, era-dependent | open standard | - | 1996 | 1 | 368.2 GFlop/s |
| Paragon I/O node RAID storage Direct-attached RAID behind dedicated I/O nodes, diskless compute nodes | Intel | XP/S RAID | 1993 | 1 | 143.4 GFlop/s |
| Scalable Disk Array Direct-attached disk nodes on the CM-5 data network, NFS-mountable | Thinking Machines | SDA | 1993 | 2 | 59.7 GFlop/s |
| BeeGFS Distributed parallel filesystem, no dedicated storage-node tier | open standard | - | - | 1 | 334.80 TFlop/s |
| HPE Cray ClusterStor E2000 | Hewlett Packard Enterprise | ClusterStor E2000 | - | 3 | 603.2 PFlop/s |
| Pure Storage FlashArray | Pure Storage | - | - | 1 | - |
| Pure Storage FlashBlade | Pure Storage | - | - | 2 | 7.57 PFlop/s |