Nvidia's new China chips: Inside the B20 and B40 specifications

Nvidia CEO Jensen Huang told CNN's Anna Stewart in Paris on Thursday that "going forward, our forecasts will not include the China market." Yet the importance of China to Nvidia goes without saying—otherwise, Huang wouldn't be going to such lengths to amplify Huawei as a competitive threat. His assertion that Huawei's chips are comparable to Nvidia's most advanced processors doesn't align with the reality on the ground in China.

Our industry checks reveal that Huawei's CloudMatrix384—which Huang positioned as comparable to Nvidia's Blackwell architecture—has proven to be a difficult sell in the Chinese market. While some companies have expressed interest, none have placed substantial orders yet. In fact, Huawei's 910B and 910C chips have struggled to gain traction, with major potential buyers including internet giants remaining cautious about procurement from the company for various reasons.

What has now become clear is that Nvidia's new China-specific chips are based on the GB202 GPU—the same foundation used for the RTX 6000 and RTX 5090.

According to our own checks, Nvidia has been developing this next-generation China chip—the successor to the H20—since at least mid-2024. The months of advance preparation, combined with the fact that these chips are built on the existing GB202 base GPU already shipping in products like the RTX 5090, position Nvidia to rapidly deploy these new China-market processors. Industry consensus suggests the new chips will become available as early as July.

Built on the GB202 die, Nvidia plans to release two variants. The first is what industry insiders are calling the B20, while the second is referred to as either the B40 or B30—the naming convention remains in flux. Two industry participants told us that Nvidia may ultimately brand these chips simply as RTX 6000 variants, a move apparently designed to obscure from regulators and markets that these are part of the Blackwell series.

The latest U.S. export controls implemented in April introduced restrictions on two additional metrics beyond total compute capacity: memory bandwidth and interconnect speed. Regarding memory specifications, we've learned that both the B20 and B40 will use GDDR memory rather than the HBM memory found in the previously China-tailored H20 chip, which was banned following the April regulations.

However, one surprising development is that the B40 will retain NVLink connectivity, giving it the same interconnect capabilities as the H20, according to a transcript from a June 13 expert interview conducted by Chinese research group AceCamp. This means the B40 could remain highly effective in cluster configurations, potentially supporting four-card single-machine setups and high-density solutions like NVL72 or NVL144 arrangements.

The following technical details for the B20 and B40 are based on the transcript from AceCamp's June 13 expert interview:

B20 configuration

The B20 utilizes Nvidia's ConnectX-8 for interconnect functionality, achieving connectivity through Ethernet with a maximum interconnect bandwidth of 800Gbps. Each B20 chip connects to a ConnectX-8 chip via NVLink bus, forming a discrete module equivalent to a PCIe card that enables interconnection in PCIe card format.

This design is optimized for small-scale clusters of 8 to 16 cards per machine, primarily targeting inference tasks and post-training for smaller models. The B20 is fundamentally unsuitable for model training because its inter-chip interconnect bandwidth is significantly less efficient than NVLink, resulting in substantial losses in overall parallelization efficiency compared to NVLink-based solutions.

B30/B40 configuration

The B30/B40 support NVLink interconnect, though at reduced speeds compared to standard specifications. The NVLink bandwidth is expected to match the H20's 900Gbps rather than achieving the B-series' full 1.8Tbps capability, though final specifications will depend on official release information.

Both variants maintain the OAM (OCP Accelerator Module) form factor with design aesthetics consistent with standard NVLink architecture, likely supporting four-card single-machine configurations that can scale to high-density solutions like NVL72 or NVL144 clusters. However, this frequency-reduced NVLink, combined with the B30's inherent computational and memory bandwidth limitations, results in cluster performance that falls short of the H20. These chips likely cannot support the large-scale cluster training that the H20 enables, with maximum capacity probably limited to post-training and fine-tuning on clusters of several hundred chips at most.

Memory specifications

The GDDR capacity configurations are expected to include 24GB, 36GB, and 48GB variants. The 48GB configuration appears most likely, with each B20 chip anticipated to feature six 8GB GDDR7 modules. Should firmware modifications become available in the future, there's potential for high-density memory solutions similar to the dual-sided memory configuration seen in consumer cards like the RTX 4090.

Expected demand

Reuters reported last month that the new China chips will be priced between $6,500 and $8,000, which aligns closely with our industry checks. If the B30 is priced at that level, it would represent exceptional value, coming in well below the $10,000-$12,000 that the H20 commanded. While the B40 delivers approximately 85% of the H20's performance, its significantly lower price point should guarantee continued client demand.

Our industry sources indicate that complete server configurations with these new chips will sell for around $80,000-$100,000, with Ethernet-connected systems priced near $80,000 and OAM NVLink configurations approaching $100,000. However, these quotes are based on preliminary estimates and already include certain discounts.

Early customer interest

According to AceCamp's research, major Chinese tech companies are showing differentiated interest in the chip variants. Tencent appears to favor the B20, as the company's inference-heavy workloads and high requirements for network scheduling and management make the B20's price-performance ratio particularly attractive. ByteDance is focusing more on the B30 and B40, hoping these higher-end models can partially address the market demand left by the discontinued H20.

Meanwhile, Alibaba has not yet expressed clear preferences for specific models, indicating strong overall chip demand but providing no detailed allocation plans. This market segmentation directly reflects each company's distinct business deployment requirements.

Where We Stand

The true test lies ahead, as none of the major Chinese clients have yet received cards for testing. The evaluation process typically takes about a month, and only after completing these assessments will companies begin placing substantial orders.

For Nvidia, the China market represents too significant a revenue source to abandon entirely. While Jensen Huang may be overstating Huawei's capabilities in an attempt to persuade U.S. officials to reconsider sanctions, Chinese competitors like Huawei are hardly standing still in their own development efforts.