Discovery by Oak Ridge National Laboratory: status and timeline

Discovery

Give U.S. scientists several times Frontier's capability for simulation and AI, from fusion and reactors to materials and medicine, with codes ready to run on day one.

ProtoStage 2 of 5

Ordered Oct 2025; nine early science teams picked Jun 2026. Delivery 2028, operations 2029.

Updated 17 Jun 2026Checked 11 Oct0 updates this week

Milestones

Next · Code readiness on early AMD/HPE hardware

Frontier passes one exaflopMay 2022Complete.
DOE, HPE and AMD announce Discovery and Lux27 Oct 2025Complete.
Call for early science applications opens12 Jan 2026Complete.
Nine day-one science projects selected17 Jun 2026Complete.
Code readiness on early AMD/HPE hardwareTarget 2026Current milestone.
Discovery delivered and installedTarget 2028Not yet reached.
Full science operations beginTarget 2029Not yet reached.

Most important updates

Upcoming

2026Code readiness on early AMD/HPE hardware (next)
2028Discovery delivered and installed
2029Full science operations begin

Current obstacles

Codes must be rewritten for new chipsScience programs must be tuned to new GPUs and memory to hit 3-5x Frontier, so teams start years early with AMD and HPE.
Hardware not yet shippingMI430X and Venice are new chips; any slip in them, or in HPE's new platform, pushes the 2028 delivery.

Physics limits

Moving data costs more than computing itChips can do arithmetic far faster than memory and networks can feed them numbers, so many science codes wait on data, not math.
Power and heat cap the sizeA top machine already draws about 20-30 MW. Every operation turns into heat that must be carried away, so energy per operation sets how big it can get.
Not every problem splits into piecesParts of a calculation that must run in order cannot be spread across more chips, so adding nodes gives smaller and smaller gains.

How it works

3 parts

Chips

GPUs built for precise math

The MI430X supports the full 64-bit arithmetic that physics simulations need alongside the low-precision math AI uses, with 432 GB of fast HBM4 memory per chip.

Blades

Dense liquid-cooled racks

HPE's GX5000 packs CPU-GPU blades tightly and cools them fully with liquid, since each blade can draw more than 25 kW.

Network

Thousands of nodes act as one

The next generation of HPE's Slingshot network links the nodes so one simulation can be split across the whole machine.

Spec sheet

vs Frontier

Spec	Discovery	Frontier	Change
GPU	AMD Instinct MI430X^R (reported)	AMD Instinct MI250X (Frontier)	—no comparable change
Memory per GPU	432 GB HBM4^R (reported)	—	—no comparable change
CPU	AMD EPYC Venice^R (reported)	—	—no comparable change
Power per compute blade	Over 25 kW^R (reported)	—	—no comparable change
Cost	About $500M^R (reported)	—	—no comparable change
Day-one code speed-up target vs Frontier	3-5x^R (reported)	—	—no comparable change

^R reported by the company

Update log

3 updates

About Oak Ridge National Laboratory

The team behind Discovery

Oak Ridge National Laboratory

Frontier and Discovery supercomputers

The largest U.S. Department of Energy science and energy lab by size, in Tennessee, run by UT-Battelle. It hosts Frontier, the first exascale supercomputer, and has ordered its successor, Discovery, from HPE and AMD for 2028.

Frontier (2022) was the first machine to pass one exaflop on the standard benchmark, at 1.102 exaflops.
Discovery (HPE Cray GX5000, AMD Venice CPUs + MI430X GPUs) is due in 2028 and in service in 2029; about $500M.
Lux, an AI cluster of about 4,000 AMD MI355X GPUs, is being installed; half its time is reserved for paying partners.

ornl.gov (opens ornl.gov)Company profile →

Founded: 194383 yrs
Headquarters: United States
Status: Governmentof U.S. Department of Energy
Valuation: government
Staff: ~7,000Jan 2026
Works in: ComputingAI factories & supercomputers
Coverage: 1 program · 3 updateslatest 17 Jun 2026checked 11 Oct
Partners: Oracle Cloud Infrastructure (opens ornl.gov)
Suppliers: Hewlett Packard Enterprise (opens hpe.com)AMD (opens ornl.gov)
People: Stephen Streiffer (opens ornl.gov)Laboratory Director

Spec

Discovery

Frontier

Change

GPU

AMD Instinct MI430X^R (reported)

AMD Instinct MI250X (Frontier)

—no comparable change

Memory per GPU

432 GB HBM4^R (reported)

—

—no comparable change

CPU

AMD EPYC Venice^R (reported)

—

—no comparable change

Power per compute blade

Over 25 kW^R (reported)

—

—no comparable change

Cost

About $500M^R (reported)

—

—no comparable change

Day-one code speed-up target vs Frontier

3-5x^R (reported)

—

—no comparable change

Discovery

Milestones

Most important updates

Upcoming

Current obstacles

Physics limits

How it works

GPUs built for precise math

Dense liquid-cooled racks

Thousands of nodes act as one

Spec sheet

Update log

Wed 17 Jun

Mon 12 Jan

Mon 27 Oct 2025

About Oak Ridge National Laboratory

Current obstacles

Physics limits

How it works

GPUs built for precise math

Dense liquid-cooled racks

Thousands of nodes act as one

Spec sheet

Update log

Wed 17 Jun

Mon 12 Jan

Mon 27 Oct 2025

About Oak Ridge National Laboratory