01TPUs and AI accelerators4
02Custom silicon and ASICs3
03GPUs3
04Interconnect and fabric0
05Memory and storage0
06Networking0
RANKED BY WHAT WE HOLD · THE 3 WITH NOTHING SORT LAST AND KEEP THEIR ROW
TPUs and AI acceleratorsTHIN4
Vendor accelerators offered as a platform rather than a card you buy: TPUs, Gaudi and their kind, usually rented inside the vendor cloud.
Ordered by attributes answered — the column we hold for most of them. Blank where we hold none.
SHOWING 4 OF 4 · ALL OF THEM
Custom silicon and ASICsTHIN3
Accelerators that are not GPUs, from companies building their own silicon for training or inference. Google TPU and Intel Gaudi sit in TPUs and accelerators; this is the rest.
Ordered by attributes answered — the column we hold for most of them. Blank where we hold none.
SHOWING 3 OF 3 · ALL OF THEM
GPUsTHIN3
The GPUs themselves as parts — generations, memory, form factor, the makers of them. Renting one by the hour is a GPU cloud, and every rate we hold is on the GPU price pages.
Ordered by attributes answered — the column we hold for most of them. Blank where we hold none.
SHOWING 3 OF 3 · ALL OF THEM
Interconnect and fabricEMPTY0
The links between accelerators and inside a node — NVLink, InfiniBand fabrics, chip-to-chip. Datacentre-level networking between machines is its own category.
Listed as a subcategory because the layer declares one. Nothing on record yet.
Nothing listed in interconnect and fabric
3 of the 6 subcategories here are empty. That is a fact about our coverage rather than a rendering fault, so the rail keeps them and this panel says so.
Suggest a product →0 OF 0 · THE RAIL STILL LISTS IT, BECAUSE AN ABSENT OPTION READS AS A MISSING FEATURE RATHER THAN AN HONEST GAP
Memory and storageEMPTY0
The memory and storage a training run is bottlenecked on: HBM, high-throughput local storage, parallel filesystems. Object storage sold by the terabyte-month is a Compute and Cloud category.
Listed as a subcategory because the layer declares one. Nothing on record yet.
Nothing listed in memory and storage
3 of the 6 subcategories here are empty. That is a fact about our coverage rather than a rendering fault, so the rail keeps them and this panel says so.
Suggest a product →0 OF 0 · THE RAIL STILL LISTS IT, BECAUSE AN ABSENT OPTION READS AS A MISSING FEATURE RATHER THAN AN HONEST GAP
NetworkingEMPTY0
Moving data between machines at training scale: switches, NICs, RDMA and the software around them. Links inside a node or between accelerators are interconnect and fabric.
Listed as a subcategory because the layer declares one. Nothing on record yet.
Nothing listed in networking
3 of the 6 subcategories here are empty. That is a fact about our coverage rather than a rendering fault, so the rail keeps them and this panel says so.
Suggest a product →0 OF 0 · THE RAIL STILL LISTS IT, BECAUSE AN ABSENT OPTION READS AS A MISSING FEATURE RATHER THAN AN HONEST GAP
Intel Gaudi4 ATTRIBUTES ANSWERED
AWS Trainium3 ATTRIBUTES ANSWERED
Google TPU1 ATTRIBUTE ANSWERED
AMD InstinctNO ANSWER ON RECORD
