Nvidia Vera Rubin NVL72 spec page now lists 19.2 TB/s HBM4 per GPU, down from 22 TB/s
Nvidia's own Vera Rubin page now shows 12.7% less memory bandwidth and 17% less NVLink bandwidth per GPU than earlier figures, with capacity and compute unchanged and no explanation.
· 8 min read
Written by AI from the sources linked below. How we use AI
On this page
Nvidia's own product page for the Vera Rubin NVL72 now lists lower memory and interconnect bandwidth than the figures the company showed earlier this year. Each Rubin GPU is listed at 19.2 TB/s of HBM4 bandwidth, down from 22 TB/s, and 3 TB/s of NVLink 6 bandwidth, down from 3.6 TB/s. Memory capacity and the headline compute numbers did not change. Nvidia has not said why. The change matters because bandwidth, not raw compute, is what limits many large-model inference workloads, and because the rival AMD part is now listed with more of it.
Key points
- Nvidia's live Vera Rubin NVL72 page lists 19.2 TB/s of HBM4 bandwidth per Rubin GPU; an earlier version of the page, reported by FourWeekMBA from an Internet Archive snapshot dated 12 August 2026, listed 22 TB/s, a drop of about 12.7%.
- NVLink 6 bandwidth per GPU is listed at 3 TB/s, down from 3.6 TB/s (about 17%); rack-level HBM4 bandwidth is 1,400 TB/s, down from 1,580 TB/s, and rack NVLink is 216 TB/s, down from 260 TB/s.
- HBM4 capacity (288 GB per GPU, 20.7 TB per rack) and the NVFP4 figures (50 PFLOPS inference, 35 PFLOPS training per GPU) are unchanged on the NVL72 page.
- The 8-GPU HGX Rubin NVL8 took a larger cut: ServeTheHome reports HBM4 bandwidth falling from 176 TB/s to 130 TB/s (about 26%) and capacity from 2.3 TB to 2.0 TB.
- Nvidia has given no reason; reported explanations (memory suppliers not reaching the original pin speed, or a push to ship more volume) are speculation from outlets.
What happened
Nvidia's Vera Rubin NVL72 page carries a specification table with three columns: the full rack (72 GPUs and 36 Vera CPUs), the Vera Rubin Superchip (two GPUs and one CPU) and a single Rubin GPU. We read the live table on 1 November 2026. It lists, per Rubin GPU, 288 GB of HBM4 with 19.2 TB/s of bandwidth, 3 TB/s of NVLink 6, 0.45 TB/s of networking and, for the NVFP4 number format, 50 PFLOPS of inference (a sparse figure) and 35 PFLOPS of training (a dense figure). The superchip is listed at 576 GB, 38.5 TB/s and 6 TB/s. The rack is listed at 20.7 TB of HBM4, 1,400 TB/s of memory bandwidth, 216 TB/s of NVLink and 32.4 TB/s of networking.
The earlier figures come from outside reporting, because Nvidia did not announce a change. The tech-news outlet FourWeekMBA compared the live page with Internet Archive snapshots. It reports that a snapshot from 12 August 2026 showed 22 TB/s per GPU, 44 TB/s per superchip, 1,580 TB/s per rack, 3.6 TB/s of NVLink per GPU and 260 TB/s per rack, and that snapshots from 2 October onward show the new values. We could not open the archive snapshots ourselves from our systems, so those "before" numbers rest on that report and on ServeTheHome, which independently lists the same preliminary figures.
The change was noticed in stages. The Chinese outlet IT Home published a comparison on 26 October, according to ServeTheHome and FourWeekMBA, using Nvidia's current documentation and a June version saved by a third party. ServeTheHome followed on 31 October. ServeTheHome says Nvidia had described the original numbers as preliminary, with "all values are up to and subject to change".
| Spec (per Rubin GPU) | Earlier figure | Current page | Change |
|---|---|---|---|
| HBM4 capacity | 288 GB | 288 GB | none |
| HBM4 bandwidth | 22 TB/s | 19.2 TB/s | about -12.7% |
| NVLink 6 bandwidth | 3.6 TB/s | 3 TB/s | about -16.7% |
| Networking | 0.4 TB/s | 0.45 TB/s | see below |
| NVFP4 inference | 50 PFLOPS | 50 PFLOPS | none |
| NVFP4 training | 35 PFLOPS | 35 PFLOPS | none |
The percentages are our arithmetic from the listed figures. The networking number went up, but FourWeekMBA notes the current page footnotes it as bidirectional bandwidth, so part of the difference may be how it is counted rather than a faster link. ServeTheHome reports it as a 12.5% increase.
Background
Rubin is Nvidia's successor to the Blackwell generation, and the NVL72 is the rack-scale product that links 72 GPUs into one NVLink domain so a model can be split across them. Two numbers decide how fast such a rack can generate text. Memory bandwidth sets how quickly the weights and the attention cache can be streamed to the compute units for each generated token, and NVLink bandwidth sets how quickly the GPUs can exchange the pieces when a model is split across them. Both matter most for large mixture-of-experts models and long contexts, where each token touches a lot of memory.
The 22 TB/s figure implied very fast HBM4 stacks. HBM4 is the newest high-bandwidth memory standard, and each Rubin GPU uses eight stacks of 36 GB, which is where the 288 GB comes from. ServeTheHome converts the old and new bandwidth into per-pin data rates of 10.75 Gbps and 9.375 Gbps. That is consistent with the percentage drop above and points at the memory stacks running slower than first planned. It is an inference, not an Nvidia statement.
Competition is part of the context. In its deep dive on AMD's Instinct MI455X, ServeTheHome gives 432 GB of HBM4 and about 23.3 TB/s of bandwidth per GPU, described there as slightly ahead of the roughly 22 TB/s it cites for Rubin. Against Rubin's current 19.2 TB/s the gap is wider: by our arithmetic, about 21% more bandwidth and 50% more capacity on AMD's side. We could not confirm AMD's figures against AMD's own documentation, so treat that comparison as ServeTheHome's. Both vendors' racks are now reaching customers; our earlier report on Meta's equity deals with Nvidia and AMD and the first Helios racks covers the AMD side of that race.
The numbers
The chart shows per-GPU HBM4 bandwidth, using Nvidia's earlier and current figures and the MI455X figure from ServeTheHome.
The cut is bigger at the eight-GPU server level. ServeTheHome reports that HGX Rubin NVL8 went from 2.3 TB of HBM4 at 176 TB/s to 2.0 TB at 130 TB/s, plus smaller reductions in training throughput (NVFP4 training from 280 to 270 PFLOPS and FP8 training from 140 to 130 PFLOPS). Eight GPUs at the new per-GPU rate would be about 154 TB/s (8 x 19.2), so 130 TB/s is lower than the per-GPU figure alone would give. ServeTheHome's reading is that about seven of each GPU's eight memory stacks are active in that configuration, which would be roughly 250 GB and about 16.8 TB/s per GPU. That is the outlet's estimate; Nvidia lists no per-GPU figure for the NVL8 part.
| TB/s | |
|---|---|
| HBM4 earlier | 1580 |
| HBM4 now | 1400 |
| NVLink earlier | 260 |
| NVLink now | 216 |
Source: Nvidia Vera Rubin NVL72 page; FourWeekMBA (earlier figures)
What it means
For buyers and cloud customers. A rack with 11% less aggregate memory bandwidth is not 11% slower at every job. Training that is limited by compute is unaffected, since the compute figures did not move. Inference that is limited by memory bandwidth, which describes much of large-model token generation, is where the lower figure is most likely to show. How much depends on the model and the batch size, and independent benchmarks, not spec sheets, will settle it. Customers who signed on the basis of the earlier numbers may want to check which table their contracts and capacity plans reference.
For Nvidia. Spec sheets are published before hardware ships, and revising them is normal; Nvidia flagged the first numbers as subject to change. What drew attention is the lack of an announcement and that the page appears to have changed quietly, as the first production systems were reaching customers. ServeTheHome reports that cloud providers are starting to make NVL72 systems generally available, citing an Oracle announcement of 27 October that we could not open.
For the HBM supply chain. If the reported cause is right, the change tells you something about memory suppliers. A lower required pin speed lets more HBM4 parts qualify, which helps availability. IT Home is cited by FourWeekMBA as expecting exactly that, and one Japanese outlet argues Nvidia is trading bandwidth for shipment volume. Neither is confirmed by Nvidia or the memory makers.
For AMD and others. On paper the MI455X now leads on per-GPU memory bandwidth and capacity. Rack-level results, software maturity and availability will matter at least as much, and we have no benchmark data to compare them.
What we don't know yet
- Why Nvidia changed the figures. The company has not commented in any source we found, and FourWeekMBA says it did not contact Nvidia either.
- Exactly when the page changed. Sources disagree. FourWeekMBA places it between 12 August and 2 October; ServeTheHome says Nvidia posted the revised specs in September; IT Home compared against a June document. Nvidia's documentation is also reported to be inconsistent, with some pages still showing the original figures.
- What shipping hardware actually delivers. A published peak bandwidth is a ceiling. Real sustained bandwidth is lower and depends on the memory stack and the software.
- The pin speed. One X post cited by FourWeekMBA claims a drop from 11.7 to 9.3 Gbps; ServeTheHome's own derivation gives 10.75 to 9.375 Gbps. The two do not match, and Nvidia lists bandwidth, not pin speed.
- Whether the NVL8 cut is a different configuration or a different chip. ServeTheHome's seven-stack reading is an estimate.
What to watch next
- Independent inference benchmarks on production Vera Rubin systems, particularly for large mixture-of-experts models, compared with Blackwell and the AMD MI455X.
- Any Nvidia clarification, a datasheet revision or a note on the NVL72 page.
- Memory-maker commentary on HBM4 volume and speeds, which would confirm or contradict the supply explanation.
- Cloud providers' published instance specifications, which will show which numbers they quote to customers.
Sources
- NVIDIA Vera Rubin NVL72 product page (primary source; live specification table)
- ServeTheHome: NVIDIA Updates Rubin GPU Specs as First Production Hardware Arrives
- FourWeekMBA: Nvidia Cuts Vera Rubin Memory Bandwidth to 19.2 TB/s per GPU
- IT Home: comparison of the Rubin specification documents, 26 October 2026
- ServeTheHome: AMD Instinct MI455X deep dive
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next
Intel Q3 2026: revenue $16.5B, but Q4 guide of $15-16B sends shares down 28%
Data center sales jumped 78% while PC revenue fell 5%; Intel says memory prices and substrate supply will keep the PC slowdown going into 2027.
audiotale.net team6 min read
Google Project Suncatcher satellite returns first AI response from a TPU in orbit: a Gemma 4 haiku
Google's prototype space data center is online and answered a prompt from orbit; the tests that matter, heat and radiation, are only starting.
audiotale.net team7 min read
Meta $97.5B equity plan: Nvidia invests $3B, AMD $1B, Helios racks installed
Meta priced $25.0B of stock and $12.5B of convertible preferred, plans up to $60B more, and its chip suppliers Nvidia and AMD are putting in $3B and $1B.
audiotale.net team7 min read