Every November, the supercomputing world gathers for one conference — SC, the International Conference for High Performance Computing — and the vendors save their biggest silicon for it. This year’s edition, SC26 in Chicago, delivered all three at once: NVIDIA, AMD, and Intel each announced flagship hardware within about 24 hours of each other. The numbers are absurd, and that’s the point. But read past the core counts and a pattern emerges: the race is no longer about raw arithmetic. It’s about feeding the cores.
NVIDIA: Rubin HPC and the 50,000-core GPU
The headline act is Rubin HPC, a new GPU in NVIDIA’s Vera Rubin platform announced on November 16. It packs more than 50,000 CUDA cores on a single GPU — nearly double the 28,672 in the current Rubin chip — plus a new sparse-data acceleration feature and seventh-generation NVLink to tie GPUs together. NVIDIA expects it to be available in 2028.
What’s interesting is how NVIDIA is framing it. The pitch isn’t “faster AI training” — it’s agents doing science. The idea: AI agents propose candidate materials, molecules, or designs, and then GPU simulations filter out the ones worth actually building in a lab. Rubin HPC is built for that loop.
Two honest caveats. First, NVIDIA published no memory capacity, power draw, price, or FP64 performance figures for the chip — so there’s nothing yet that tells you how fast it will run your code, only how many cores it has. Second, core count is the wrong number to stare at for the workloads NVIDIA names. On the HPCG benchmark — which measures the sparse, memory-bound math real science actually runs, as opposed to dense peak FLOPS — not one of the top 10 systems in the world cracks 3% of its theoretical peak. That’s exactly why the sparse-data acceleration matters more than the core count: the cores have been waiting on data for years.
AMD: Venice ships, and Venice-X brings a gigabyte of cache
AMD used SC26 to show off its 6th Gen EPYC “Venice” server processors, which started shipping in OEM and ODM systems on November 10. The specs are a generational jump: up to 256 cores per socket on the dense Zen 6c design, boost clocks up to 5.3 GHz, up to 1.6 TB/s of memory bandwidth, and PCIe Gen 6 — all on TSMC’s 2nm process. AMD says the new Zen 6 core delivers about 27% more instructions per clock than Zen 5.
The part to watch is Venice-X, due in the second half of 2027: up to 96 cores with 1,152 MB of L3 cache per socket, thanks to stacked 3D V-Cache. That’s more than a gigabyte of cache sitting right next to the cores, aimed squarely at the workloads that choke on memory latency — computational fluid dynamics, chip design, and other simulation work where the data set almost fits but not quite. For perspective, the top standard Venice chip already carries 1,024 MB of L3; Venice-X is the cache-maximalist version of the same silicon.
There’s a neat manufacturing detail in here too. Each 32-core Venice chiplet physically carries a 33rd spare core (plus a spare slice of cache) that AMD never enables — it’s a yield safeguard, letting a chiplet with one defective core still ship as a full 32-core part. When your dies are this big and this expensive, you plan for defects.
Intel: the D270 bets on cheap, plentiful memory
Intel’s SC26 story is the GPU D270, the data-center accelerator formerly codenamed Crescent Island. It’s a 350W air-cooled PCIe card built on the Xe3P architecture, and it makes an unusual bet: instead of expensive stacked HBM memory like every other AI accelerator, it uses 160 GB of LPDDR5X — the same kind of memory in your phone — delivering 1.5 TB/s of bandwidth, with partner boards scaling up to 480 GB. It also keeps real double-precision (FP64) compute alongside the low-precision AI formats, which most inference cards have abandoned.
Intel published its first benchmarks at SC26, claiming up to 5.9x the FP64 throughput of NVIDIA’s RTX PRO 6000 Blackwell Server Edition on financial workloads like options pricing. Treat that number the way you’d treat any vendor benchmark: it’s Intel’s own testing on pre-production silicon, and trade press has already flagged quirks in the comparison setup. The card is sampling with customers now, with volume ramping into 2027 and no price or general-availability date yet. Still, the architecture is the story — memory capacity over memory exoticism — and Intel says it already has over $2 billion in orders, per the CEO’s remarks at last week’s SGLang Summit.
Intel also outlined Diamond Rapids, its next Xeon: up to 256 performance cores, 1.6 TB/s of memory bandwidth, and 1.28 GB of last-level cache. No launch date attached.
What this means for the rest of us
Nothing announced this week is something you can buy — these are 2027 and 2028 parts, priced in the five figures, destined for national labs and AI factories. But supercomputing has always been where consumer hardware is prototyped in public, and the trickle-down is easy to trace:
- Zen 6 is coming to the desktop. Trade reporting points to Ryzen versions in the first half of 2027, and 3D V-Cache has already proven itself a gaming story. If you’re planning a CPU upgrade, the next generation is now a known quantity, not a rumor.
- Cache is the new core count. All three vendors are, in different ways, spending transistors on keeping data close to the compute — AMD with stacked cache, NVIDIA with sparse-data acceleration, Intel with a mountain of LPDDR5X.
- The memory story continues. Last week’s RAM shortage post was about AI demand eating the memory market. SC26 is the other side of that coin: when memory is the bottleneck and the scarce resource, every vendor redesigns around it.
The supercomputer of 2028 won’t be interesting because it has 50,000 cores on a chip. It’ll be interesting because, for the first time in a while, the vendors are all admitting the cores were never the hard part — getting data to them was. This week, they all showed their answers.