Instances›AWS EC2›g4dn›g4dn.12xlarge
Compute
vCPUs48
Memory192 GiB
Mem / vCPU4.00 GiB
ProcessorXeon Family
Clock speed2.5 GHz
Architecturex86_64
Storage
Type1 × 900 GB NVMe SSD
Bandwidth9,500 Mbps
Throughput1187.5 MB/s
IOPS40,000
Networking
Performance50 Gigabit
Max ENIs8
IPv6Yes
GPUNVIDIA T4 Tensor Core
Family sizes
g4dn.xlarge4 vCPU · 16 GiB
g4dn.2xlarge8 vCPU · 32 GiB
g4dn.4xlarge16 vCPU · 64 GiB
g4dn.8xlarge32 vCPU · 128 GiB
g4dn.12xlarge48 vCPU · 192 GiB
g4dn.16xlarge64 vCPU · 256 GiB
g4dn.metal96 vCPU · 384 GiB
g4dn.12xlarge

g4dn.12xlarge is AWS's GPU EC2 size at 48 vCPU and 192 GiB. In the name, "g" is graphics, "4" is the generation, "dn" marks local NVMe storage and enhanced networking and "12xlarge" is the size within the g4dn family. It runs on Intel Xeon Family at 2.5 GHz, and AWS introduced it in September 2019. AWS has since shipped g5, g6 and g7; g5.12xlarge is the direct replacement at this size, scoring 895,147 CoreMark/s against this instance's 715,777 — roughly 25% more throughput for the same vCPU count. Rackspace Spot runs one GPU class, gpu.vs2.xlargeplusplus — a single NVIDIA A30 with 24 vCPU and 64 GiB, in San Jose only. Against this instance's 4 × NVIDIA T4 Tensor Core that is a downgrade per card, honest for inference and not for the training this instance was bought for — it prices $4.69/hr on AWS in US West (N. California), from $0.191/hr on Rackspace Spot.

Match qualityWeaker GPU

g4dn.12xlarge puts 4 Turing, Turing, Turing, Turing GPUs (64 GiB VRAM) in one machine. Spot has a single GPU class and no multi-GPU node, so there is no shape here that places 4 cards behind one host. Per card the NVIDIA A30 (Ampere, 24 GB VRAM) is a step down from what g4dn.12xlarge carries, and the aggregate is not reachable at all.

That makes this honest for data-parallel inference and embarrassingly parallel batch work, where each card is independent, and dishonest for anything that shards a model across cards or synchronises every step. The catch is location rather than hardware: US West, San Jose is the only Spot region that offers it, so latency to your users and egress to your data are the things to check.

No local diskg4dn.12xlarge comes with 1 × 900 GB NVMe SSD attached to the machine. Spot nodes use network storage only, so that disk doesn't come with it. If you were using it as fast scratch space, you'd move to a persistent volume or object storage, which are slower to read.
Pricing by region
1 hr
RegionAWS on-demandAWS Spot1-yr Reserved3-yr ReservedRackspace SpotYou save
US West, San Jose$4.694$1.827$3.303$2.281$0.191
$4.503 savings
When to use it
Good for
  • ●Machine learning experimentation and fine-tuning
  • ●Remote graphics workstations and rendering
  • ●CUDA-accelerated batch jobs
  • ●Workloads that fit inside 192 GiB of RAM
  • ●Jobs that want 1 × 900 GB of local NVMe scratch space
  • ●Network-heavy services — AWS rates this size at 50 Gigabit
Avoid if
  • ●Starting a new deployment. g5.12xlarge is the newer generation at this size.
  • ●The working set needs more than 192 GiB of RAM
  • ●More than 4.00 GiB of RAM per vCPU is needed — that ratio is fixed for the whole family
  • ●The work is single-threaded. Most of these 48 vCPU would sit idle.
  • ●Anything the CPU can do fast enough. GPU hours are the most expensive on the price list.
Stop overpaying AWS. This compute costs $3,427/mo on AWS vs $139/mo on Rackspace Spot.
  • Spot sells spare capacity at market prices. It is the cheapest compute on the internet.
  • Prices are set by supply and demand. You pay what the market decides.
  • Compute is delivered as fully managed Kubernetes. All infrastructure is taken care of.
  • Egress costs $0.01/GB after the free allowance. AWS charges $0.09/GB.
Build your Cloudspace
Frequently asked questions