1,800-Petaflop El Capitan Quadruples Fugaku: U.S. Leads Exascale Race—But GPU Hack Hijacks Crypto in 12k Steps

TL;DR

• El Capitan Supercomputer Surpasses 1,800 Petaflops, Setting New Exascale Benchmark in U.S.
• University of Toronto researchers demonstrate GPUBreach, a Rowhammer attack exploiting GPU GDDR6 memory to escalate privileges

🚀 El Capitan Crushes 1.8-Exaflop Barrier, Outruns Fugaku 4-to-1: U.S. Reclaims Supercomputing Crown

1,800 petaflops = 4× Japan’s Fugaku 🚀—El Capitan just rewrote the supercomputer record book. That’s every person on Earth doing 230M calc/sec. Power bill: 10 MW. U.S. labs & your tax dollar—ready to model fusion, climate, nukes. Ready for the next exascale sprint?

Lawrence Livermore’s new machine cranks 1,800 double-precision petaflops, pushing every GPU node to 2.57 petaflops and widening the U.S. lead over every rival on the planet.

How the numbers stack up

• 2.57 petaflops/node × 700 nodes = 1.8 exaflops LINpack, 98 % of theoretical peak
• HDR InfiniBand shuttles 200 Gb/s per port; NVMe tier feeds data at >200 TB/s
• 10 MW liquid-cooled racks hold power draw inside DOE’s exascale envelope (~0.5 GFLOPS/W)

Impacts so far

Science: full-scale fusion plasma runs that once took a week now finish overnight
Security: stockpile models that chewed 30 % of annual cycles drop to single-digit share
Market: 80 % leap over Frontier re-centers global procurement on U.S. GPU designs

Gaps & fixes

Software: legacy codes hit 40 % of node peak; ASC co-design teams port 12 flagship apps this quarter
Power: 10 MW exceeds campus baseload; 30 % renewable contract slated for 2027
Supply: sole-source NVIDIA risk—LLNL pilots AMD Instinct nodes for 2028 refresh

Timelines

• Q3 2026: FP8 tensor paths go live, doubling AI training throughput
• 2027: HDR200+ cuts latency 15 %, pushing aggregate toward 2.3 exaflops
• 2029: Hopper-X upgrade targets 3 PFLOPS/node, lifting system past 3 exaflops
• 2030: immersion cooling pilot aims for PUE <1.1, trimming $1 M/year in power bills

Double-precision bragging rights today translate into faster bombs-to-breakthroughs tomorrow; if the codes keep pace, El Capitan’s 1.8 exaflops will soon feel as routine as a 4-minute mile.

😱 One Bit Flip Hijacks RTX GPUs: Rowhammer Bypasses IOMMU, Threatens Cloud

RTX GPUs flip bits with just 12k activations—enough to hijack host memory & bypass IOMMU 😱 One rogue CUDA job can read your crypto keys or root the box. Cloud gamers & AI tenants, patch NOW—how fast will your provider shield you?

On 13 April in Oakland, University of Toronto researchers showed that a single flipped bit inside an NVIDIA RTX A6000’s GDDR6 can turn the card into a skeleton key for the whole host. By repeatedly activating two neighboring DRAM rows 12 000 times in a few milliseconds, they forced a page-table bit to invert, promoted a read-only entry to read-write, and walked straight past the IOMMU. Microsoft paid the team $600; attackers would pay nothing.

How one bit topples a system

• Trigger: CUDA kernels allocate 2 MB slabs until the target page-table row is physically adjacent; double-sided “hammer” accesses follow.
• Flip: ≥8 bits twitch per run; one suffices to corrupt a PTE.
• Win: The GPU now reads or writes any host page via PCIe BAR1; kernel code injection finishes the job.

Impacts at a glance

Confidentiality: cleartext keys, credentials, and VM images exposed → cloud tenant separation collapses.
Integrity: arbitrary kernel writes → root shell, persistent firmware implants.
Availability: corrupted page tables → crash or ransom-ready DoS.
Economics: 25 tested gamer-class cards and $6 000-data-center GPUs alike remain vulnerable until NVIDIA ships a patch.

What defenders are doing—and not doing

NVIDIA has a July 2025 microcode update in test; AWS and Google Cloud temporarily pin one GPU per customer and shrink PCIe apertures. AMD, Apple, and Qualcomm already ship ECC by default, reducing flip rates to near zero, but NVIDIA’s GDDR6 lines omit the feature. No mainstream hypervisor yet alerts on 12 k-row activation bursts.

Outlook

• Q3 2025: Patch drops; expect <30 % uptake among workstations, >80 % among major clouds.
• 2026: New GDDR6 wafers add Target-Row-Refresh, raising hammer threshold >20 k; older cards stay exposed.
• 2027: JEDEC row-hammer tests likely become mandatory; GPU-ECC and on-die integrity checks spread to budget lines.

Bottom line

A GPU bought for gaming or AI can now pry open the server room. Until silicon or firmware hardens every GDDR6 bit, treat high-performance graphics cards as untrusted peripherals—because one electron nudge is all it takes to swap “denied” for “root.”

In Other News

• Texas launches Hydrogen City, the world’s first large-scale green hydrogen production facility targeting 280,000 tons annually by 2029