quickstart →

EC2 storage: NVMe vs gp3 vs tmpfs.

RunsOn launches the instance type you ask for, so the workspace disk is yours to pick: EBS gp3 by default, local NVMe on i and d families, or RAM with tmpfs: the workspace lives in memory, so the trade-off is capacity shared with the build, not durability.

Provisioned gp3 2.5× 401 MiB/s → 1,001 MiB/s seq read, same m8a.2xlarge; latency flat
Local NVMe by size 8.1× 503 MiB/s on i7i.large → 4.0 GiB/s on i7i.4xlarge
tmpfs for databases 8.8× 28,457 vs 3,234 TPS, pgbench, same m8a.2xlarge
  • Provisioned gp3 lifted sequential reads 2.5× on the same m8a.2xlarge (401 MiB/s to 1,001 MiB/s), but single-queue reads (1,587 to 1,776 IOPS) and synced writes (540 to 540 IOPS) barely moved: it buys bandwidth, not latency.
  • Local NVMe bandwidth grows with the instance: i7i.large reads 503 MiB/s, i7i.4xlarge 4.0 GiB/s (8.1×).
  • On the same m8a.2xlarge, tmpfs ran pgbench's read/write mix at 28,457 TPS against 3,234 on default gp3: 8.8× faster.
  • r8a.large with tmpfs ($0.0009/min at spot) answers 1,106,090 QD1 reads and 428,067 synced writes per second: 90× and 33× the best local NVMe at 2 vCPU (i7i.large, $0.0010) and 611× and 791× default EBS gp3 (c8a.large, $0.0008); its workspace is 14 GiB, from RAM.

Sequential read by instance size

Local NVMe grows with the instance; EBS gp3 grows with what you provision. Right: random-read IOPS.

Sequential read1M blocks · MiB/s, higher is better

i7i · local NVMe · Xeon 8559C

  1. i7i.large 2 vCPU 503 MiB/s , random read 75k IOPS
  2. i7i.xlarge 4 vCPU 1,010 MiB/s , random read 150k IOPS
  3. i7i.2xlarge 8 vCPU 2.0 GiB/s , random read 300k IOPS
  4. i7i.4xlarge 16 vCPU 4.0 GiB/s , random read 576k IOPS

m8id · local NVMe · Xeon 6975P-C

  1. m8id.large 2 vCPU 226 MiB/s , random read 34k IOPS
  2. m8id.xlarge 4 vCPU 451 MiB/s , random read 67k IOPS
  3. m8id.2xlarge 8 vCPU 905 MiB/s , random read 134k IOPS
  4. m8id.4xlarge 16 vCPU 1.8 GiB/s , random read 269k IOPS

m9gd · local NVMe · Neoverse-V3

  1. m9gd.large 2 vCPU 293 MiB/s , random read 44k IOPS
  2. m9gd.xlarge 4 vCPU 585 MiB/s , random read 87k IOPS
  3. m9gd.2xlarge 8 vCPU 1.1 GiB/s , random read 175k IOPS

EBS gp3 · default and provisioned

  1. m8a.2xlarge default gp3 401 MiB/s , random read 3.0k IOPS
  2. m8a.2xlarge gp3 1000 MiB/s 1,001 MiB/s , random read 16k IOPS
  3. m8azn.xlarge default gp3 401 MiB/s , random read 3.0k IOPS
  4. m8azn.xlarge gp3 1000 MiB/s 1,001 MiB/s , random read 16k IOPS

tmpfs is off this scale: 18.5 GiB/s on m8a.2xlarge, from RAM.

Every storage option measured

fio inside the job workspace; random reads span three orders of magnitude (log bars). Other providers for context.

Instance Rand read Rand write QD1 read p99 Sync write Seq read Seq write $/min
i7i · local NVMe instance storeXeon 8559C
i7i.4xlarge 16 vCPU 128 GB local NVMe 575,787 339,548 11,162 313 µs 11,234 4.0 GiB/s 3.1 GiB/s $0.0074 $0.0255 OD
i7i.2xlarge 8 vCPU 64 GB local NVMe 300,466 165,105 10,218 342 µs 11,228 2.0 GiB/s 1.5 GiB/s $0.0055 $0.0129 OD
i7i.xlarge 4 vCPU 32 GB local NVMe 150,054 82,479 12,059 210 µs 12,740 1,010 MiB/s 788 MiB/s $0.0018 $0.0066 OD
i7i.large 2 vCPU 16 GB local NVMe 75,023 41,222 12,284 241 µs 12,999 503 MiB/s 395 MiB/s $0.0010 $0.0035 OD
m8id · local NVMe instance storeXeon 6975P-C
m8id.4xlarge 16 vCPU 64 GB local NVMe 268,806 134,244 7,562 449 µs 6,642 1.8 GiB/s 867 MiB/s $0.0032 $0.0177 OD
m8id.2xlarge 8 vCPU 32 GB local NVMe 134,258 67,072 10,703 264 µs 7,569 905 MiB/s 432 MiB/s $0.0033 $0.0090 OD
m8id.xlarge 4 vCPU 16 GB local NVMe 67,090 33,526 10,866 249 µs 7,903 451 MiB/s 216 MiB/s $0.0016 $0.0047 OD
m8id.large 2 vCPU 8 GB local NVMe 33,542 16,749 11,212 105 µs 5,506 226 MiB/s 108 MiB/s $0.0008 $0.0025 OD
m9gd · local NVMe instance storeNeoverse-V3
m9gd.2xlarge 8 vCPU 32 GB local NVMe 174,572 87,209 14,984 71 µs 15,902 1.1 GiB/s 561 MiB/s $0.0031 $0.0087 OD
m9gd.xlarge 4 vCPU 16 GB local NVMe 87,221 43,591 13,375 111 µs 14,860 585 MiB/s 281 MiB/s $0.0014 $0.0045 OD
m9gd.large 2 vCPU 8 GB local NVMe 43,606 21,773 13,590 80 µs 7,267 293 MiB/s 141 MiB/s $0.0008 $0.0024 OD
EBS gp3 · RunsOn default volume, and provisioned
m8a.2xlarge 8 vCPU 32 GB EBS gp3 2,995 2,990 1,587 864 µs 540 401 MiB/s 401 MiB/s $0.0026 $0.0084 OD
m8a.2xlarge 8 vCPU 32 GB gp3 1000 MiB/s 15,995 15,989 1,776 766 µs 540 1,001 MiB/s 1,001 MiB/s $0.0048 $0.0105 OD
c8a.large 2 vCPU 4 GB EBS gp3 2,994 2,992 1,811 754 µs 541 401 MiB/s 401 MiB/s $0.0008 $0.0021 OD
c8a.xlarge 4 vCPU 8 GB EBS gp3 2,994 2,991 1,773 897 µs 540 401 MiB/s 401 MiB/s $0.0015 $0.0039 OD
c8a.2xlarge 8 vCPU 16 GB EBS gp3 2,995 2,991 1,821 897 µs 540 401 MiB/s 401 MiB/s $0.0029 $0.0075 OD
c8g.large 2 vCPU 4 GB EBS gp3 2,993 2,993 1,803 836 µs 537 401 MiB/s 401 MiB/s $0.0007 $0.0016 OD
c8g.xlarge 4 vCPU 8 GB EBS gp3 2,992 2,994 1,803 860 µs 538 401 MiB/s 401 MiB/s $0.0016 $0.0030 OD
c8g.2xlarge 8 vCPU 16 GB EBS gp3 2,994 2,993 1,783 856 µs 539 401 MiB/s 401 MiB/s $0.0029 $0.0056 OD
m8a.large 2 vCPU 8 GB EBS gp3 2,994 2,992 1,734 872 µs 541 401 MiB/s 401 MiB/s $0.0008 $0.0023 OD
m8a.xlarge 4 vCPU 16 GB EBS gp3 2,994 2,991 1,825 807 µs 540 401 MiB/s 401 MiB/s $0.0013 $0.0044 OD
m8azn.large 2 vCPU 8 GB EBS gp3 2,994 2,992 1,778 692 µs 541 401 MiB/s 401 MiB/s $0.0013 $0.0037 OD
m8azn.xlarge 4 vCPU 16 GB EBS gp3 2,994 2,992 1,583 918 µs 541 401 MiB/s 401 MiB/s $0.0024 $0.0072 OD
m8azn.xlarge 4 vCPU 16 GB gp3 1000 MiB/s 15,994 15,991 1,772 799 µs 541 1,001 MiB/s 1,001 MiB/s $0.0045 $0.0093 OD
m8azn.3xlarge 12 vCPU 48 GB EBS gp3 2,997 2,994 1,802 811 µs 538 401 MiB/s 401 MiB/s $0.0081 $0.0210 OD
m8i.large 2 vCPU 8 GB EBS gp3 2,993 2,994 1,625 987 µs 535 401 MiB/s 401 MiB/s $0.0009 $0.0021 OD
m8i.xlarge 4 vCPU 16 GB EBS gp3 2,993 2,994 1,666 954 µs 518 401 MiB/s 401 MiB/s $0.0016 $0.0038 OD
m8i.2xlarge 8 vCPU 32 GB EBS gp3 2,993 2,994 1,565 995 µs 486 401 MiB/s 401 MiB/s $0.0030 $0.0074 OD
m8i-flex.large 2 vCPU 8 GB EBS gp3 2,993 2,994 1,771 881 µs 530 401 MiB/s 401 MiB/s $0.0008 $0.0020 OD
m8i-flex.xlarge 4 vCPU 16 GB EBS gp3 2,993 2,994 1,753 918 µs 525 401 MiB/s 401 MiB/s $0.0012 $0.0037 OD
m8i-flex.2xlarge 8 vCPU 32 GB EBS gp3 2,993 2,995 1,780 954 µs 518 401 MiB/s 401 MiB/s $0.0025 $0.0070 OD
m9g.large 2 vCPU 8 GB EBS gp3 2,993 2,994 1,702 905 µs 542 401 MiB/s 401 MiB/s $0.0007 $0.0019 OD
m9g.xlarge 4 vCPU 16 GB EBS gp3 2,994 2,992 1,837 868 µs 543 401 MiB/s 401 MiB/s $0.0016 $0.0036 OD
m9g.2xlarge 8 vCPU 32 GB EBS gp3 2,995 2,992 1,768 901 µs 543 401 MiB/s 401 MiB/s $0.0023 $0.0068 OD
r8a.large 2 vCPU 16 GB EBS gp3 2,995 2,992 1,703 860 µs 541 401 MiB/s 401 MiB/s $0.0009 $0.0030 OD
t8i.medium 2 vCPU 4 GB EBS gp3 3,001 2,998 1,737 827 µs 523 401 MiB/s 401 MiB/s $0.0006 $0.0011 OD
tmpfs · workspace in RAM
m8a.2xlarge 8 vCPU 32 GB tmpfs 4,357,799 3,821,571 1,177,079 <1 µs 429,009 18.5 GiB/s 10.1 GiB/s $0.0026 $0.0084 OD
m9g.2xlarge 8 vCPU 32 GB tmpfs 2,734,029 2,529,692 689,402 <1 µs 303,107 15.8 GiB/s 10.0 GiB/s $0.0023 $0.0068 OD
r8a.2xlarge 8 vCPU 64 GB tmpfs 4,310,049 3,831,509 1,153,982 <1 µs 430,608 18.4 GiB/s 10.0 GiB/s $0.0033 $0.0110 OD
m9g.xlarge 4 vCPU 16 GB tmpfs 2,714,301 2,564,752 677,554 <1 µs 304,157 14.8 GiB/s 9.2 GiB/s $0.0016 $0.0036 OD
r8a.xlarge 4 vCPU 32 GB tmpfs 4,201,526 3,800,170 1,102,306 <1 µs 433,841 17.8 GiB/s 9.9 GiB/s $0.0016 $0.0056 OD
r9g.xlarge 4 vCPU 32 GB tmpfs 2,605,268 2,485,222 643,948 <1 µs 301,871 14.0 GiB/s 8.9 GiB/s $0.0023 $0.0046 OD
r8a.large 2 vCPU 16 GB tmpfs 2,147,299 1,930,574 1,106,090 <1 µs 428,067 18.0 GiB/s 10.0 GiB/s $0.0009 $0.0030 OD
r9g.large 2 vCPU 16 GB tmpfs 1,277,889 1,234,447 652,948 <1 µs 297,441 13.8 GiB/s 9.2 GiB/s $0.0009 $0.0024 OD
Other providers, 4/8/2 vCPU x64 (hardware suite)
Namespace 4x8 4 vCPU 8 GB VM disk EPYC 1,209,187 607,540 105,744 10 µs 30,836 5.0 GiB/s 3.4 GiB/s $0.0060
Namespace 4x16 4 vCPU 16 GB VM disk EPYC 1,202,685 564,373 104,443 11 µs 28,561 5.0 GiB/s 3.4 GiB/s $0.0120
Namespace 8x32 8 vCPU 32 GB VM disk EPYC 1,095,705 569,755 104,850 10 µs 29,050 5.0 GiB/s 3.4 GiB/s $0.0240
Namespace 8x16 8 vCPU 16 GB VM disk EPYC 931,530 546,288 106,191 10 µs 27,450 4.5 GiB/s 2.9 GiB/s $0.0120
Namespace 2x8 2 vCPU 8 GB VM disk EPYC 820,729 574,093 104,675 10 µs 30,063 4.5 GiB/s 3.0 GiB/s $0.0060
Blacksmith 4 vCPU 4 vCPU 16 GB VM disk EPYC 728,035 218,915 71,728 16 µs 17,925 7.0 GiB/s 1.7 GiB/s $0.0080
Blacksmith 2 vCPU 2 vCPU 8 GB VM disk EPYC 703,229 166,801 73,779 17 µs 17,755 5.5 GiB/s 2.0 GiB/s $0.0040
Avrea 2 vCPU 2 vCPU 8 GB VM disk EPYC 4585PX 564,218 245,269 53,018 14 µs 12,665 16.0 GiB/s 10.6 GiB/s $0.0040
Blacksmith 8 vCPU 8 vCPU 32 GB VM disk EPYC 511,769 231,934 70,703 17 µs 17,670 6.4 GiB/s 1.9 GiB/s $0.0160
boat 2 vCPU 2 vCPU 4 GB VM disk Ryzen 7 9700X 460,252 120,989 10,663 218 µs 2,270 10.2 GiB/s 832 MiB/s $0.0003
Avrea 4 vCPU 4 vCPU 16 GB VM disk EPYC 4585PX 453,992 236,446 49,955 21 µs 12,305 15.9 GiB/s 11.0 GiB/s $0.0080
boat 4 vCPU 4 vCPU 8 GB VM disk Ryzen 9 9950X 385,820 154,383 8,814 297 µs 2,235 7.8 GiB/s 1.3 GiB/s $0.0006
Avrea 8 vCPU 8 vCPU 32 GB VM disk EPYC 4585PX 343,658 200,865 52,383 18 µs 11,105 16.8 GiB/s 11.2 GiB/s $0.0160
boat 8 vCPU 8 vCPU 16 GB VM disk EPYC 4585PX 309,424 85,110 8,181 545 µs 1,390 8.1 GiB/s 1,020 MiB/s $0.0012
Ubicloud premium 8 vCPU 8 vCPU 32 GB VM disk Ryzen 9 7950X3D 289,856 19,284 8,638 515 µs 2,698 2.9 GiB/s 993 MiB/s $0.0080
Ubicloud standard 8 vCPU 8 vCPU 32 GB VM disk Ryzen 9 7950X3D 258,356 20,072 8,526 587 µs 2,039 2.6 GiB/s 989 MiB/s $0.0050
Ubicloud premium 4 vCPU 4 vCPU 16 GB VM disk Ryzen 9 7950X3D 224,242 23,696 7,621 860 µs 2,661 2.3 GiB/s 902 MiB/s $0.0040
Warpbuild 2 vCPU 2 vCPU 8 GB VM disk Ryzen 9 9950X 199,592 90,570 51,593 37 µs 3,078 4.2 GiB/s 2.7 GiB/s $0.0040
Warpbuild 4 vCPU 4 vCPU 16 GB VM disk Ryzen 9 9950X 198,918 75,037 47,423 31 µs 2,989 4.2 GiB/s 2.9 GiB/s $0.0080
Warpbuild 8 vCPU 8 vCPU 32 GB VM disk Ryzen 9 9950X 189,921 75,793 38,684 27 µs 2,993 4.4 GiB/s 3.1 GiB/s $0.0160
Ubicloud standard 4 vCPU 4 vCPU 16 GB VM disk EPYC 9454P 176,445 24,130 7,775 713 µs 1,622 2.4 GiB/s 620 MiB/s $0.0025
Ubicloud standard 2 vCPU 2 vCPU 8 GB VM disk EPYC 9454P 128,623 23,218 6,828 763 µs 749 2.4 GiB/s 710 MiB/s $0.0013
Ubicloud premium 2 vCPU 2 vCPU 8 GB VM disk Ryzen 9 7950X3D 116,749 23,514 4,832 1,733 µs 1,444 2.5 GiB/s 832 MiB/s $0.0020
GitHub 8-core 8 vCPU 32 GB VM disk EPYC 9V45 38,976 18,540 7,529 231 µs 4,104 784 MiB/s 540 MiB/s $0.0220
GitHub 4-core 4 vCPU 16 GB network disk Xeon 8573C 19,575 11,216 7,286 241 µs 4,163 394 MiB/s 244 MiB/s $0.0120
GitHub 2-core 2 vCPU 8 GB VM disk Xeon 8573C 9,377 8,367 6,395 262 µs 3,665 199 MiB/s 166 MiB/s $0.0060
AWS CodeBuild medium 4 vCPU 8 GB EBS Xeon 8488C 2,992 2,974 1,796 786 µs 566 241 MiB/s 251 MiB/s $0.0100
AWS CodeBuild large 8 vCPU 16 GB EBS Xeon 8488C 2,989 2,977 1,758 741 µs 566 243 MiB/s 251 MiB/s $0.0200
AWS CodeBuild small 2 vCPU 3 GB EBS Xeon 8488C 2,988 2,972 1,762 803 µs 571 251 MiB/s 251 MiB/s $0.0050
  • IOPS are 4k blocks · QD32 keeps 32 requests in flight, QD1 one
  • $/min: EC2 spot + EBS, on-demand (OD) under it; a dash means no price yet. The RunsOn license is a fixed cost, not included
More findings11: every NVMe family, provisioned gp3, tmpfs
  • i7i: random read climbs from 75,023 IOPS on i7i.large to 575,787 on i7i.4xlarge (7.7×), while single-request (QD1) reads stay between 10,218 and 12,284 IOPS.
  • m8id: random read climbs from 33,542 IOPS on m8id.large to 268,806 on m8id.4xlarge (8.0×), while single-request (QD1) reads range between 7,562 and 11,212 IOPS.
  • m9gd: random read climbs from 43,606 IOPS on m9gd.large to 174,572 on m9gd.2xlarge (4.0×), while single-request (QD1) reads stay between 13,375 and 14,984 IOPS.
  • On m8azn.xlarge, gp3 provisioned at 1,000 MiB/s and 16,000 IOPS delivers it: 1,001 MiB/s and 15,994 random-read IOPS, against 401 MiB/s and 2,994 on the default volume. QD1 reads go from 1,583 to 1,772 IOPS.
  • tmpfs on r9g.large answers 1,277,889 random-read IOPS and 297,441 synced writes per second; m9g.large (same CPU, 8 vs 16 GiB RAM) on default gp3 manages 2,993 and 542. The price is capacity: the workspace is 14 GiB of RAM.
  • tmpfs on m9g.xlarge answers 2,714,301 random-read IOPS and 304,157 synced writes per second; the same m9g.xlarge on default gp3 manages 2,994 and 543. The price is capacity: the workspace is 14 GiB of RAM.
  • tmpfs on r8a.xlarge answers 4,201,526 random-read IOPS and 433,841 synced writes per second; c8a.xlarge (same CPU, 8 vs 32 GiB RAM) on default gp3 manages 2,994 and 540. The price is capacity: the workspace is 29 GiB of RAM.
  • tmpfs on r9g.xlarge answers 2,605,268 random-read IOPS and 301,871 synced writes per second; m9g.xlarge (same CPU, 16 vs 32 GiB RAM) on default gp3 manages 2,994 and 543. The price is capacity: the workspace is 29 GiB of RAM.
  • tmpfs on m8a.2xlarge answers 4,357,799 random-read IOPS and 429,009 synced writes per second; the same m8a.2xlarge on default gp3 manages 2,995 and 540. The price is capacity: the workspace is 29 GiB of RAM.
  • tmpfs on m9g.2xlarge answers 2,734,029 random-read IOPS and 303,107 synced writes per second; the same m9g.2xlarge on default gp3 manages 2,995 and 543. The price is capacity: the workspace is 29 GiB of RAM.
  • tmpfs on r8a.2xlarge answers 4,310,049 random-read IOPS and 430,608 synced writes per second; c8a.2xlarge (same CPU, 16 vs 64 GiB RAM) on default gp3 manages 2,995 and 540. The price is capacity: the workspace is 59 GiB of RAM.

What storage does to real builds

The same three builds on RunsOn runners that differ by storage, one 2 vCPU runner per kind first.

Runner Rust TypeScript bun install Docker $/Rust build
RunsOn x6429 runners
m8azn.large 2 vCPU 8 GB EBS gp3 EPYC 9R05 2m 39s1m 43s 21s n/a: Not run: the PostHog Docker build needs more than 8 GB of memory; it ran out of memory on 8 GB runners. $0.0040
m8a.large 2 vCPU 8 GB EBS gp3 EPYC 9R45 2m 57s1m 49s 21s n/a: Not run: the PostHog Docker build needs more than 8 GB of memory; it ran out of memory on 8 GB runners. $0.0030
r8a.large 2 vCPU 16 GB EBS gp3 EPYC 9R45 2m 58s1m 42s 23s 11m 32s $0.0032
c8a.large 2 vCPU 4 GB EBS gp3 EPYC 9R45 3m 12sfailed: out of memory: 4 GB was not enough on this runner for the opencode typecheck, even with turbo limited to 4 tasks: TypeScript processes are killed out of memory (exit 137). 16 GB passes. It has 3.7 GiB of swap (zram). — n/a: Not run: the PostHog Docker build needs more than 8 GB of memory; it ran out of memory on 8 GB runners. $0.0032
m8i.large 2 vCPU 8 GB EBS gp3 Xeon 6975P-C 4m 54s2m 57s 23s n/a: Not run: the PostHog Docker build needs more than 8 GB of memory; it ran out of memory on 8 GB runners. $0.0053
m8i-flex.large 2 vCPU 8 GB EBS gp3 Xeon 6975P-C 4m 56s2m 54s 22s n/a: Not run: the PostHog Docker build needs more than 8 GB of memory; it ran out of memory on 8 GB runners. $0.0054
t8i.medium 2 vCPU 4 GB EBS gp3 Xeon 6975P-C 5m 30sfailed: out of memory: 4 GB was not enough on this runner for the opencode typecheck, even with turbo limited to 4 tasks: TypeScript processes are killed out of memory (exit 137). 16 GB passes. It has 3.7 GiB of swap (zram). — n/a: Not run: the PostHog Docker build needs more than 8 GB of memory; it ran out of memory on 8 GB runners. $0.0037
m8azn.3xlarge 12 vCPU 48 GB EBS gp3 EPYC 9R05 57s36s 20s 4m 55s $0.0133
c8a.2xlarge 8 vCPU 16 GB EBS gp3 EPYC 9R45 1m 15s48s 24s 7m 49s $0.0049
m8a.2xlarge 8 vCPU 32 GB EBS gp3 EPYC 9R45 1m 16s45s 23s 5m 39s $0.0046
m8azn.xlarge 4 vCPU 16 GB EBS gp3 EPYC 9R05 1m 42s57s 19s 8m 48s $0.0061
m8i.2xlarge 8 vCPU 32 GB EBS gp3 Xeon 6975P-C 1m 42s57s 23s 7m 11s $0.0072
m8i-flex.2xlarge 8 vCPU 32 GB EBS gp3 Xeon 6975P-C 1m 46s1m 02s 24s 7m 29s $0.0072
m8a.xlarge 4 vCPU 16 GB EBS gp3 EPYC 9R45 1m 49s1m 03s 20s 9m 08s $0.0034
c8a.xlarge 4 vCPU 8 GB EBS gp3 EPYC 9R45 1m 50s1m 12s 22s n/a: Not run: the PostHog Docker build needs more than 8 GB of memory; it ran out of memory on 8 GB runners. $0.0036
m8i.xlarge 4 vCPU 16 GB EBS gp3 Xeon 6975P-C 2m 47s1m 27s 23s 11m 56s $0.0054
m8i-flex.xlarge 4 vCPU 16 GB EBS gp3 Xeon 6975P-C 2m 49s1m 31s 23s 11m 46s $0.0054
m8a.2xlarge no 2 vCPU EBS gp3, provisioned runner 8 vCPU 32 GB gp3 1000 MiB/s EPYC 9R45 1m 13s35s 12s 4m 39s $0.0079
m8azn.xlarge 4 vCPU 16 GB gp3 1000 MiB/s EPYC 9R05 1m 41s47s 10s 7m 05s $0.0104
m8id.large 2 vCPU 8 GB local NVMe Xeon 6975P-C 4m 47s2m 53s 36s n/a: Not run: the PostHog Docker build needs more than 8 GB of memory; it ran out of memory on 8 GB runners. $0.0049
i7i.large 2 vCPU 16 GB local NVMe Xeon 8559C 5m 10s2m 37s 19s 16m 59s $0.0060
m8id.2xlarge 8 vCPU 32 GB local NVMe Xeon 6975P-C 1m 46s45s 11s 6m 00s $0.0080
i7i.2xlarge 8 vCPU 64 GB local NVMe Xeon 8559C 1m 51s46s 10s 6m 16s $0.0135
m8id.xlarge 4 vCPU 16 GB local NVMe Xeon 6975P-C 2m 44s1m 20s 18s 9m 15s $0.0074
i7i.xlarge 4 vCPU 32 GB local NVMe Xeon 8559C 2m 54s1m 17s 12s 9m 15s $0.0069
r8a.large 2 vCPU 16 GB tmpfs EPYC 9R45 3m 06s1m 26s 12s n/a: Not run: needs more than 16 GB of RAM with tmpfs, where Docker's data root lives in RAM; the 16 GB m9g.xlarge tmpfs runner was shut down mid-build, even with zram swap. The 32 GB tmpfs runners build it. $0.0032
r8a.2xlarge 8 vCPU 64 GB tmpfs EPYC 9R45 1m 15s32s 9s 3m 42s $0.0062
m8a.2xlarge 8 vCPU 32 GB tmpfs EPYC 9R45 1m 15s32s 9s 3m 32s $0.0046
r8a.xlarge 4 vCPU 32 GB tmpfs EPYC 9R45 1m 52s49s 9s 4m 55s $0.0036
RunsOn arm6413 runners
m9g.large 2 vCPU 8 GB EBS gp3 Neoverse-V3 2m 58s1m 48s 21s n/a: Not run: the PostHog Docker build needs more than 8 GB of memory; it ran out of memory on 8 GB runners. $0.0027
c8g.large 2 vCPU 4 GB EBS gp3 Neoverse-V2 5m 05sfailed: out of memory: 4 GB was not enough on this runner for the opencode typecheck, even with turbo limited to 4 tasks: TypeScript processes are killed out of memory (exit 137). 16 GB passes. It has 3.7 GiB of swap (zram). — n/a: Not run: the PostHog Docker build needs more than 8 GB of memory; it ran out of memory on 8 GB runners. $0.0049
m9g.2xlarge 8 vCPU 32 GB EBS gp3 Neoverse-V3 1m 17s46s 25s 6m 03s $0.0039
c8g.2xlarge 8 vCPU 16 GB EBS gp3 Neoverse-V2 1m 38s56s 27s 8m 40s $0.0058
m9g.xlarge 4 vCPU 16 GB EBS gp3 Neoverse-V3 1m 49s58s 21s 9m 04s $0.0035
c8g.xlarge 4 vCPU 8 GB EBS gp3 Neoverse-V2 2m 29s1m 19s 22s n/a: Not run: the PostHog Docker build needs more than 8 GB of memory; it ran out of memory on 8 GB runners. $0.0048
m9gd.large 2 vCPU 8 GB local NVMe Neoverse-V3 2m 53s1m 43s 29s n/a: Not run: the PostHog Docker build needs more than 8 GB of memory; it ran out of memory on 8 GB runners. $0.0026
m9gd.2xlarge 8 vCPU 32 GB local NVMe Neoverse-V3 1m 13s30s 9s 4m 24s $0.0051
m9gd.xlarge 4 vCPU 16 GB local NVMe Neoverse-V3 1m 47s50s 14s 6m 19s $0.0029
r9g.large 2 vCPU 16 GB tmpfs Neoverse-V3 3m 04s1m 27s 12s n/a: Not run: needs more than 16 GB of RAM with tmpfs, where Docker's data root lives in RAM; the 16 GB m9g.xlarge tmpfs runner was shut down mid-build, even with zram swap. The 32 GB tmpfs runners build it. $0.0036
m9g.2xlarge 8 vCPU 32 GB tmpfs Neoverse-V3 1m 15s31s 8s 4m 04s $0.0039
r9g.xlarge 4 vCPU 32 GB tmpfs Neoverse-V3 1m 50s49s 9s 5m 22s $0.0055
m9g.xlarge 4 vCPU 16 GB tmpfs Neoverse-V3 1m 52s48s 9s n/a: Not run: needs more than 16 GB of RAM with tmpfs, where Docker's data root lives in RAM; even with zram swap the runner was shut down mid-build. The 32 GB tmpfs runners build it. $0.0035
Same CPU, different disk31 comparisons
  • m8a.2xlarge on EBS gp3 and on gp3 1000 MiB/s: Rust even within run-to-run spread (1m 13s–1m 21s on EBS gp3, 1m 12s–1m 21s on gp3 1000 MiB/s); TypeScript 35s on gp3 1000 MiB/s vs 45s; Docker even within run-to-run spread (5m 33s–6m 28s on EBS gp3, 4m 13s–7m 11s on gp3 1000 MiB/s).
  • m8a.2xlarge on EBS gp3 and on tmpfs: Rust even within run-to-run spread (1m 13s–1m 21s on EBS gp3, 1m 13s–1m 18s on tmpfs); TypeScript 32s on tmpfs vs 45s; Docker 3m 32s on tmpfs vs 5m 39s.
  • m8azn.xlarge on EBS gp3 and on gp3 1000 MiB/s: Rust even within run-to-run spread (1m 36s–1m 47s on EBS gp3, 1m 37s–1m 48s on gp3 1000 MiB/s); TypeScript 47s on gp3 1000 MiB/s vs 57s; Docker even within run-to-run spread (7m 50s–21m 49s on EBS gp3, 6m 38s–9m 53s on gp3 1000 MiB/s).
  • m9g.xlarge on EBS gp3 and on tmpfs: Rust even within run-to-run spread (1m 47s–1m 52s on EBS gp3, 1m 46s–2m 14s on tmpfs); TypeScript 48s on tmpfs vs 58s.
  • m9g.2xlarge on EBS gp3 and on tmpfs: Rust even within run-to-run spread (1m 13s–1m 23s on EBS gp3, 1m 13s–1m 20s on tmpfs); TypeScript 31s on tmpfs vs 46s; Docker even within run-to-run spread (5m 44s–6m 13s on EBS gp3, 3m 57s–7m 21s on tmpfs).
  • r8a.large on EBS gp3 and on tmpfs: Rust even within run-to-run spread (2m 50s–3m 15s on EBS gp3, 2m 53s–3m 20s on tmpfs); TypeScript even within run-to-run spread (1m 35s–1m 47s on EBS gp3, 1m 25s–1m 38s on tmpfs).
  • m8a.2xlarge on gp3 1000 MiB/s and on tmpfs: Rust even within run-to-run spread (1m 12s–1m 21s on gp3 1000 MiB/s, 1m 13s–1m 18s on tmpfs); TypeScript even within run-to-run spread (34s–35s on gp3 1000 MiB/s, 29s–36s on tmpfs); Docker 3m 32s on tmpfs vs 4m 39s.
  • m8i.2xlarge (EBS gp3) and m8id.2xlarge (local NVMe) share the Xeon 6975P-C: Rust even within run-to-run spread (1m 38s–1m 55s on EBS gp3, 1m 40s–1m 53s on local NVMe); TypeScript 45s on local NVMe vs 57s; Docker even within run-to-run spread (6m 50s–9m 00s on EBS gp3, 5m 42s–8m 32s on local NVMe).
  • m8i-flex.2xlarge (EBS gp3) and m8id.2xlarge (local NVMe) share the Xeon 6975P-C: Rust even within run-to-run spread (1m 41s–2m 02s on EBS gp3, 1m 40s–1m 53s on local NVMe); TypeScript 45s on local NVMe vs 1m 02s; Docker even within run-to-run spread (7m 21s–9m 33s on EBS gp3, 5m 42s–8m 32s on local NVMe).
  • m8i-flex.large (EBS gp3) and m8id.large (local NVMe) share the Xeon 6975P-C: Rust even within run-to-run spread (4m 39s–5m 41s on EBS gp3, 4m 38s–5m 09s on local NVMe); TypeScript even within run-to-run spread (2m 48s–3m 25s on EBS gp3, 2m 46s–3m 34s on local NVMe).
  • m8i-flex.xlarge (EBS gp3) and m8id.xlarge (local NVMe) share the Xeon 6975P-C: Rust even within run-to-run spread (2m 40s–2m 57s on EBS gp3, 2m 37s–2m 49s on local NVMe); TypeScript 1m 20s on local NVMe vs 1m 31s; Docker 9m 15s on local NVMe vs 11m 46s.
  • m8i.large (EBS gp3) and m8id.large (local NVMe) share the Xeon 6975P-C: Rust even within run-to-run spread (4m 38s–5m 37s on EBS gp3, 4m 38s–5m 09s on local NVMe); TypeScript even within run-to-run spread (2m 32s–3m 17s on EBS gp3, 2m 46s–3m 34s on local NVMe).
  • m8i.xlarge (EBS gp3) and m8id.xlarge (local NVMe) share the Xeon 6975P-C: Rust even within run-to-run spread (2m 41s–2m 51s on EBS gp3, 2m 37s–2m 49s on local NVMe); TypeScript even within run-to-run spread (1m 24s–1m 34s on EBS gp3, 1m 19s–1m 24s on local NVMe); Docker 9m 15s on local NVMe vs 11m 56s.
  • m9g.xlarge (EBS gp3) and m9gd.xlarge (local NVMe) share the Neoverse-V3: Rust even within run-to-run spread (1m 47s–1m 52s on EBS gp3, 1m 45s–1m 56s on local NVMe); TypeScript 50s on local NVMe vs 58s; Docker 6m 19s on local NVMe vs 9m 04s.
  • m9g.2xlarge (EBS gp3) and m9gd.2xlarge (local NVMe) share the Neoverse-V3: Rust even within run-to-run spread (1m 13s–1m 23s on EBS gp3, 1m 11s–1m 17s on local NVMe); TypeScript 30s on local NVMe vs 46s; Docker 4m 24s on local NVMe vs 6m 03s.
  • m9g.large (EBS gp3) and m9gd.large (local NVMe) share the Neoverse-V3: Rust even within run-to-run spread (2m 50s–3m 07s on EBS gp3, 2m 50s–2m 56s on local NVMe); TypeScript even within run-to-run spread (1m 39s–2m 01s on EBS gp3, 1m 40s–1m 53s on local NVMe).
  • m9gd.2xlarge (local NVMe) and m9g.2xlarge (tmpfs) share the Neoverse-V3: Rust even within run-to-run spread (1m 11s–1m 17s on local NVMe, 1m 13s–1m 20s on tmpfs); TypeScript even within run-to-run spread (30s–34s on local NVMe, 28s–38s on tmpfs); Docker even within run-to-run spread (4m 17s–5m 12s on local NVMe, 3m 57s–7m 21s on tmpfs).
  • m9gd.xlarge (local NVMe) and m9g.xlarge (tmpfs) share the Neoverse-V3: Rust even within run-to-run spread (1m 45s–1m 56s on local NVMe, 1m 46s–2m 14s on tmpfs); TypeScript even within run-to-run spread (49s–51s on local NVMe, 46s–53s on tmpfs).
  • c8a.large (EBS gp3) and r8a.large (tmpfs) share the EPYC 9R45 (RAM differs: 4 vs 16 GiB): Rust even within run-to-run spread (3m 06s–3m 22s on EBS gp3, 2m 53s–3m 20s on tmpfs).
  • c8a.xlarge (EBS gp3) and r8a.xlarge (tmpfs) share the EPYC 9R45 (RAM differs: 8 vs 32 GiB): Rust even within run-to-run spread (1m 47s–1m 55s on EBS gp3, 1m 47s–1m 58s on tmpfs); TypeScript 49s on tmpfs vs 1m 12s.
  • c8a.2xlarge (EBS gp3) and m8a.2xlarge (gp3 1000 MiB/s) share the EPYC 9R45 (RAM differs: 16 vs 32 GiB): Rust even within run-to-run spread (1m 12s–1m 23s on EBS gp3, 1m 12s–1m 21s on gp3 1000 MiB/s); TypeScript 35s on gp3 1000 MiB/s vs 48s; Docker 4m 39s on gp3 1000 MiB/s vs 7m 49s.
  • c8a.2xlarge (EBS gp3) and m8a.2xlarge (tmpfs) share the EPYC 9R45 (RAM differs: 16 vs 32 GiB): Rust even within run-to-run spread (1m 12s–1m 23s on EBS gp3, 1m 13s–1m 18s on tmpfs); TypeScript 32s on tmpfs vs 48s; Docker 3m 32s on tmpfs vs 7m 49s.
  • c8a.2xlarge (EBS gp3) and r8a.2xlarge (tmpfs) share the EPYC 9R45 (RAM differs: 16 vs 64 GiB): Rust even within run-to-run spread (1m 12s–1m 23s on EBS gp3, 1m 13s–1m 22s on tmpfs); TypeScript 32s on tmpfs vs 48s; Docker even within run-to-run spread (7m 16s–9m 27s on EBS gp3, 3m 36s–11m 51s on tmpfs).
  • m8a.2xlarge (EBS gp3) and r8a.2xlarge (tmpfs) share the EPYC 9R45 (RAM differs: 32 vs 64 GiB): Rust even within run-to-run spread (1m 13s–1m 21s on EBS gp3, 1m 13s–1m 22s on tmpfs); TypeScript 32s on tmpfs vs 45s; Docker even within run-to-run spread (5m 33s–6m 28s on EBS gp3, 3m 36s–11m 51s on tmpfs).
  • m8a.large (EBS gp3) and r8a.large (tmpfs) share the EPYC 9R45 (RAM differs: 8 vs 16 GiB): Rust even within run-to-run spread (2m 51s–3m 07s on EBS gp3, 2m 53s–3m 20s on tmpfs); TypeScript 1m 26s on tmpfs vs 1m 49s.
  • m8a.xlarge (EBS gp3) and r8a.xlarge (tmpfs) share the EPYC 9R45 (RAM differs: 16 vs 32 GiB): Rust even within run-to-run spread (1m 46s–1m 58s on EBS gp3, 1m 47s–1m 58s on tmpfs); TypeScript 49s on tmpfs vs 1m 03s; Docker 4m 55s on tmpfs vs 9m 08s.
  • m9g.xlarge (EBS gp3) and r9g.xlarge (tmpfs) share the Neoverse-V3 (RAM differs: 16 vs 32 GiB): Rust even within run-to-run spread (1m 47s–1m 52s on EBS gp3, 1m 49s–1m 56s on tmpfs); TypeScript 49s on tmpfs vs 58s; Docker even within run-to-run spread (8m 36s–9m 31s on EBS gp3, 5m 08s–8m 42s on tmpfs).
  • m9g.large (EBS gp3) and r9g.large (tmpfs) share the Neoverse-V3 (RAM differs: 8 vs 16 GiB): Rust even within run-to-run spread (2m 50s–3m 07s on EBS gp3, 2m 55s–5m 01s on tmpfs); TypeScript even within run-to-run spread (1m 39s–2m 01s on EBS gp3, 1m 23s–1m 44s on tmpfs).
  • m8a.2xlarge (gp3 1000 MiB/s) and r8a.2xlarge (tmpfs) share the EPYC 9R45 (RAM differs: 32 vs 64 GiB): Rust even within run-to-run spread (1m 12s–1m 21s on gp3 1000 MiB/s, 1m 13s–1m 22s on tmpfs); TypeScript even within run-to-run spread (34s–35s on gp3 1000 MiB/s, 30s–34s on tmpfs); Docker even within run-to-run spread (4m 13s–7m 11s on gp3 1000 MiB/s, 3m 36s–11m 51s on tmpfs).
  • m9gd.xlarge (local NVMe) and r9g.xlarge (tmpfs) share the Neoverse-V3 (RAM differs: 16 vs 32 GiB): Rust even within run-to-run spread (1m 45s–1m 56s on local NVMe, 1m 49s–1m 56s on tmpfs); TypeScript even within run-to-run spread (49s–51s on local NVMe, 48s–55s on tmpfs); Docker even within run-to-run spread (6m 10s–6m 54s on local NVMe, 5m 08s–8m 42s on tmpfs).
  • m9gd.large (local NVMe) and r9g.large (tmpfs) share the Neoverse-V3 (RAM differs: 8 vs 16 GiB): Rust even within run-to-run spread (2m 50s–2m 56s on local NVMe, 2m 55s–5m 01s on tmpfs); TypeScript even within run-to-run spread (1m 40s–1m 53s on local NVMe, 1m 23s–1m 44s on tmpfs).

How it's measured

Methodology

The same fio tests as the runner table, inside the job workspace, on every size of each family. The jobs ran on RunsOn's default capacity: 13 on spot and 2 on on-demand, its fallback. None was interrupted.