SC26에서 NVIDIA는 Rubin HPC 칩을 발표했습니다. 과학 및 공학 연산에 특화된 서버 GPU로, 50,000개 이상의 CUDA 코어를 갖추며 2028년에 출시될 예정입니다.
Adding to a busy SC26 conference, NVIDIA has announced that it is developing a new GPU targeted specifically at high-performance computing (HPC) workloads. Named the Rubin HPC, the future chip will be an HPC-optimized version of the company's Rubin GPU architecture. It would mark NVIDIA's first significant HPC-class part since the Hopper generation, almost four years ago.
This week's announcement comes amid broader soul-searching over the past couple of years about the future of classic HPC workloads and the hardware that will run them. As commercial GPU server demand has shifted toward AI-related tasks and low-precision math over the last half-decade, server GPU designs have favored large arrays of tensor cores and other features that better suit AI needs. This has left traditional science and engineering users with fewer options than they previously enjoyed, as for most of the prior decade server GPU designs treated HPC as a first-class citizen, if not a top customer.
NVIDIA's announcement of a dedicated HPC chip is therefore significant. After the company all but sat out the HPC market with its Blackwell generation of parts, it was unclear whether the leader of the AI revolution would bother with the smaller HPC market. The answer is an unambiguous "yes," though not until 2028, when the Rubin HPC chip is released. Nonetheless, it is a clear signal that NVIDIA does not intend to permanently cede the HPC (or supercomputer) market to rivals like AMD and Intel, who have been carving out that market in recent years.
**Rubin HPC: 50K+ CUDA Cores, Coming in 2028**
With the Rubin HPC chip not slated to arrive until 2028, this week's announcement from the company, led by a very visible VP of Hyperscale and HPC, Ian Buck, is in effect a teaser with more details to come. NVIDIA shared only a few key details about the chip.
The headline news is that NVIDIA's future HPC chip will feature more than 50,000 CUDA cores, the company's ALU arrays for classic vector/SIMD processing. This is around twice as many CUDA cores as NVIDIA's current flagship Rubin SoC, and it matters greatly to the HPC market, as few HPC workloads can make heavy use of matrix multiplication and NVIDIA's tensor cores. An ideal HPC processor needs to be a strong vector processing machine, not just a tensor processing machine, so the announcement of a large number of CUDA cores underscores what kind of GPU this will be.
Beyond that, NVIDIA is not sharing any further details about the expected performance or memory subsystem of the HPC chip. Instead, the company is touting a "balanced design" that grows compute and data movement capabilities together, rather than ramping up compute throughput on its own.
This is presumably a reference to the chip's other marquee feature, NVLink 7 support. NVIDIA's current Rubin server GPU uses NVLink 6, and in fact is the first chip to offer it. That means Rubin HPC's scale-up networking technology will be a generation beyond that when it launches in 2028.
Reading between the lines, the high-level specifications for Rubin HPC appear to be more closely tied to NVIDIA's mid-generation refresh for the Rubin lineup, Rubin Ultra. With that chip not set to be available for another year, there is little information available on it as well, but notably, we already know it will be the first chip to ship with NVLink 7 support. This lends credence to the idea that the 2028 Rubin HPC chip will be closer to a Rubin Ultra derivative than to the original Rubin server chip.
The 2028 timeframe also lines up with giving Rubin HPC access to newer HBM4e memory, rather than the HBM4 used on the current Rubin server chip.
NVIDIA also included a render of the chip in its SC26 presentation and newsroom article. It looks very similar to the company's current Rubin GPU renders. NVIDIA chip renders are often more aspirational or illustrative than technical this far out, but if the render is at least physically accurate, we are looking at eight stacks of HBM spread across multiple dies, the same as the current Rubin server chip.
Finally, NVIDIA briefly touted a "sparse data acceleration" capability to keep the GPU cores fed, though it offered no other details on how this works. The company's tensor cores have long supported sparse matrices, but this would seem to be something beyond that, especially given the emphasis on vector math and CUDA cores.
**Specification Comparison**
| Accelerator | Rubin HPC | B200 | H200 |
|---|---|---|---|
| CUDA Cores | >50,000 | 18,944 | 16,896 |
| Tensor Cores | Fewer? | 592 | 528 |
| Memory Clock | HBM4E? | 8 Gbps HBM3E | 6.4 Gbps HBM3 |
| Memory Bus Width | 2x 8192-bit? | 2x 4096-bit | 6144-bit |
| Memory Bandwidth | TBD | 8 TB/sec | 4.8 TB/sec |
| VRAM | TBD | 192 GB | 141 GB |
| FP32 Vector | TBD | 75 TFLOPS | 67 TFLOPS |
| FP64 Vector | A whole lot | 37 TFLOPS | 34 TFLOPS |
| FP64 Tensor | TBD | 37 TFLOPS | 67 TFLOPS |
| INT8/FP8 Tensor | TBD | 9 P(FL)OPS | 3.96 P(FL)OPS |
| Interconnect | NVLink 7 | NVLink 5, 18 links (1,800 GB/sec) | NVLink 4, 18 links (900 GB/sec) |
| GPU | Rubin HPC | B200 | GH100 |
|---|---|---|---|
| Transistor Count | ? | 208B (2x 104B) | 80B |
| Manufacturing Process | TSMC 3nm? | TSMC 4NP | TSMC 4N |
| Architecture | Rubin | Blackwell | Hopper |
In lieu of specific performance figures, NVIDIA emphasizes that Rubin HPC will support a wide range of precisions, from NVFP4 up to FP64. While FP64 (and to a lesser extent FP32) are the primary focus for classic HPC workloads, more recent software efforts have examined lower-precision math as well, if only to account for how GPU hardware has changed over the years. It was always extremely unlikely that NVIDIA would remove low-precision format support entirely, if only for binary compatibility reasons, and the emphasis on supporting the full range of precisions suggests NVIDIA will not heavily compromise them either.
Still, the top-line figure to watch will be FP64 performance, for both vector and tensor operations. NVIDIA has essentially plateaued at a bit under 40 TFLOPS of FP64 since Hopper (with Blackwell Ultra being especially poor here at less than 2 TFLOPS). A fresh chip with a much larger number of CUDA cores would give NVIDIA the means to significantly raise its FP64 performance, depending largely on how many FP64 CUDA cores it bakes into the architecture, and how much die space it takes from tensor cores to make room for that FP64 hardware.
**Final Words**
The Rubin HPC announcement means NVIDIA is firmly putting its hat back in the ring for HPC after a few years of relative absence. With the company so heavily focused on AI performance during the Blackwell generation, the Rubin generation offers a chance to reset its position in the market and take advantage of the technologies it has developed. NVIDIA will not be without competition, however. AMD plans to launch its HPC-focused Instinct MI430X accelerator in 2027, and Intel is preparing its PCIe-based GPU D-Series. NVIDIA will have a fight on its hands if it wants to reclaim a large portion of the HPC market, or its close cousin, the classic supercomputer market.