Huawei Ascend 950 vs NVIDIA B300 and B200 for LLM Inference: TPS, Memory Bandwidth, and Cost Comparison (2026)

The Atlas 950 SuperNode was cited as part of Huawei's compute infrastructure behind some DeepSeek-family models. That connection brought the Ascend 950 (also called the 950PR) onto the radar of LLM inference teams who had never looked at non-NVIDIA hardware before. This post gives you the spec breakdown, benchmarks, and cost math to decide whether Ascend 950 is worth pursuing or whether NVIDIA Blackwell and Hopper are the practical choice. One thing to state clearly upfront: most teams outside China cannot legally procure Ascend hardware. US export control rules generally restrict acquisition of Huawei's advanced AI chips, and Ascend 950 is not offered through any globally accessible GPU cloud provider.

This is the third entry in a series covering non-NVIDIA AI accelerators for LLM inference. See Intel Gaudi 3 vs H200 and B200 for the SynapseAI stack analysis and AMD MI400 vs B300 for the CDNA 5 vs Blackwell Ultra comparison.

Quick Answer: Ascend 950 vs B200 vs B300 at a Glance

What Is the Huawei Ascend 950 (950PR)?

The Ascend 950 is Huawei's current flagship AI accelerator, built on the Da Vinci 3.0 architecture. Its key specs, per Huawei announcement, are: 1.56 PFLOPS FP4 compute, 112 GB HiBL (Huawei's proprietary HBM-class memory) at 1.4 TB/s, and a 600W TDP. The chip uses HCCS for intra-node communication and LingQu for cluster-scale interconnect. There is no Western-accessible datasheet for the Ascend 950. All specs above come from Huawei's Chinese-language announcements and third-party analysis, and should be read as Huawei-claimed figures rather than independently verified numbers.

The Atlas 950 SuperNode

The Atlas 950 SuperNode scales 8,192 Ascend 950 cards through Huawei's LingQu proprietary fabric, delivering 16 EFLOPS FP4 in aggregate (per Huawei announcement). Note: the SuperNode is built using the Ascend 950DT (decode/training) variant, not the 950PR (prefill) card profiled throughout this post; the specs and throughput estimates in this article apply to the 950PR and do not reflect the 950DT cards in the SuperNode. It was cited as part of the training compute infrastructure behind some DeepSeek-family models. To be precise: the SuperNode connection is to DeepSeek's training infrastructure, not inference serving. The Atlas 950 SuperNode is not available as a commercial cloud rental product outside Huawei Cloud in China.

Full Spec Comparison: Ascend 950 vs B200 vs B300

Note on FP4 figures: Huawei's FP4 implementation on Ascend and NVIDIA's NVFP4 are different formats with different numeric ranges and rounding behavior. Peak TFLOPS figures are not directly comparable across vendors. The Ascend 950's FP8 figure (~1 PFLOPS) is per Huawei's published specification and is the basis for the widely cited "2.8x H20" performance claim. NVIDIA's B200 FP4 figure (20 PFLOPS sparse) is NVIDIA's published sparse peak; dense throughput is approximately half that.

Memory Bandwidth: Why 1.4 TB/s vs 8 TB/s Decides Inference Throughput

Decode (token generation) is memory-bandwidth-bound at small batch sizes. Each forward pass loads model weights from GPU memory, and token throughput scales roughly linearly with how fast that load completes. At batch size 1, nearly all time is spent loading weights and KV cache rather than doing matrix math.

Llama 3.3 70B in FP8 weighs approximately 70 billion × 1 byte = 70 GB. At 1.4 TB/s (Ascend 950), the weights move through the memory system roughly 20 times per second at batch 1. At 8 TB/s (B200), that same load completes approximately 114 times per second. That 5.7x bandwidth ratio (8 TB/s divided by 1.4 TB/s) maps almost directly to a 5.7x throughput gap at small batches, which is the figure used throughout this post.

The gap does not close at long context. KV cache grows with sequence length, and loading the accumulated KV cache per decode step also scales with bandwidth. For applications like long-document Q&A or multi-turn conversations with large context windows, the bandwidth gap remains decisive. For deeper reading on why LLM weights and KV cache dominate GPU memory requirements, see the GPU memory requirements for LLMs guide.

LLM Inference Benchmarks

Independent Ascend 950 inference benchmarks are scarce outside Chinese-language technical sources as of June 2026. Huawei has not submitted Ascend 950 to MLCommons or MLPerf data center inference benchmarks. The Ascend 950 figures below are bandwidth-scaling projections derived from H200 baselines using the 1.4 TB/s vs 4.8 TB/s ratio, and are labeled as estimates throughout. Note: at batch 32 and above, inference begins to shift from purely bandwidth-bound toward partially compute-bound behavior, meaning actual throughput on the Ascend 950 could exceed the pure bandwidth-ratio projection, but without measured data, the estimates here follow the bandwidth ratio conservatively.

Llama 3.3 70B FP8: Throughput by Batch Size

B200 and B300 both carry 8.0 TB/s per chip, so their batch-1 and batch-8 decode throughput is similar. At batch 32 and above, B300's larger compute die provides a modest edge for compute-bound work. At small batches, bandwidth is the binding constraint and the two chips perform equivalently.

DeepSeek V4-Pro: Multi-GPU Configuration

DeepSeek V4-Pro at FP8 requires roughly 500 GB VRAM to hold model weights, which rules out single-card inference on any current hardware. The minimum configurations are: 5x Ascend 950 (560 GB combined HiBL), 3x B200 SXM6 (576 GB combined HBM3e), or 4x H200 SXM5 (564 GB combined HBM3e).

There is an additional complication on Ascend: the vLLM Ascend backend has incomplete support for DeepSeek V4's fine-grained MoE routing as of June 2026. Teams running DeepSeek V4-Pro on Ascend would likely need custom CANN kernels or patches to the community backend. For the full vLLM expert parallelism setup on NVIDIA hardware, see the Deploy DeepSeek V4 on GPU cloud guide.

Interconnect: LingQu vs NVLink 5

Huawei has cited ~2 TB/s chip-to-chip bandwidth for the LingQu/UnifiedBus interconnect; this figure has not been independently confirmed. NVSwitch in the B200/B300 NVL72 delivers 1.8 TB/s per-GPU all-to-all within a 72-GPU rack. For fine-grained MoE expert dispatch like DeepSeek V4-Pro (256+ experts), this tight coupling keeps expert-routing latency low.

For teams outside China operating at sub-SuperNode scale (8 to 16 GPUs), NVLink 5 on B200 or H200 is the proven interconnect option. LingQu at that scale is not accessible as a standalone product through any globally available channel.

Software Reality: CANN vs CUDA

torch_npu patches PyTorch's device dispatch to route compute to CANN. Standard ops work. Custom CUDA kernels require full rewrites using Ascend C, Huawei's kernel language equivalent to CUDA C++.

The vLLM Ascend backend is a community project, not Huawei-official. It lags upstream vLLM by several release cycles. Speculative decoding, MRV2, and SGLang are not available on Ascend. TensorRT-LLM is NVIDIA-only and will never run on Ascend hardware. Teams using TensorRT-LLM in production cannot migrate to Ascend without a complete stack swap.

Estimated porting time: 2 to 4 weeks for a standard PyTorch vLLM deployment on Ascend, assuming no prior CANN experience. Custom attention implementations or non-standard model architectures add 4 to 8 weeks. For comparison, deploying vLLM on NVIDIA hardware typically takes a few hours. See the vLLM production deployment guide for the NVIDIA baseline.

Availability and Export Controls

Ascend 950 is not sold outside China and is not available on any globally accessible GPU cloud provider. Channels where it is available: Huawei Cloud (China region), Alibaba Cloud (China region), and ByteDance Volcengine.

US export control updates from October 2022 and October 2023 from the Bureau of Industry and Security (BIS) restrict Huawei's access to advanced chip manufacturing equipment and restrict export of advanced AI accelerators involving Huawei's supply chain. The specific legal exposure depends on organization type, jurisdiction, and the nature of the transaction. US export control rules generally restrict export of Ascend hardware, but teams should verify their specific situation with legal counsel.

For teams inside China: Ascend 950 on Huawei Cloud is a real and competitive option, particularly for Chinese-language model inference where the community has CANN-optimized serving paths for Qwen, Baichuan, and DeepSeek variants.

Cost-Per-Token Analysis

Spot CPM formula: CPM = (spot $/hr) / (tokens_per_sec × 3600 / 1,000,000)

B200 and B300 produce similar cost-per-token at batch 32 despite B300's higher hourly rate, because B300's larger compute die delivers more tokens per second at that batch size. H200 is the most established option with the widest vLLM support. For measured baseline numbers from independent inference benchmarks, see the GPU cost per token benchmark 2026 post.

Pricing fluctuates based on GPU availability. The prices above are based on 10 Jun 2026 and may have changed. Check current GPU pricing → for live rates.

When Ascend 950 Makes Sense

• Teams inside China with existing CANN toolchain experience and direct Huawei Cloud access
• Large-scale MoE inference within Huawei's Atlas 950 SuperNode infrastructure
• Chinese-language model serving (Qwen, Baichuan, DeepSeek) where the community has CANN-optimized inference paths
• Organizations where access to NVIDIA hardware is the binding constraint (rare outside specific jurisdictions)

When NVIDIA Wins

• Any team outside China: Ascend 950 is not procurable through legal channels for most organizations
• Fine-grained MoE inference (DeepSeek V4-class models) at multi-GPU scale where NVLink 5's 1.8 TB/s reduces expert dispatch latency
• Deployments using TensorRT-LLM, SGLang, upstream vLLM features, or speculative decoding
• Teams without CANN experience who cannot absorb a 2 to 8 week porting investment
• Long-context inference at small batch sizes: B200's 8 TB/s bandwidth produces roughly 5.7x higher decode throughput than Ascend 950's 1.4 TB/s
• Any workload requiring English-language documentation, active community support, and mature ecosystem tooling

Decision Matrix

For teams outside China deciding now, the practical path comes down to memory and budget. B300 GPU rental delivers 288 GB HBM3e at 8.0 TB/s, which handles 200B+ parameter models at FP16 and large MoE models at FP8 without scaling past a single card. If your budget is tighter or you need lower per-GPU cost at equivalent bandwidth, B200 SXM6 on Spheron offers the same 8 TB/s with 192 GB at $2.68/hr. For teams running Llama 3 or DeepSeek V3-class models that fit in under 141 GB, on-demand H200 instances remain the most battle-tested option in the NVIDIA cloud stack with the widest vLLM coverage and the most community benchmarks to compare against.

The Ascend 950 is real hardware with real performance inside Huawei's infrastructure. But export controls make it unavailable to most teams, and CANN's porting cost eats the budget advantage for anyone starting from a CUDA stack. For DeepSeek V4-Pro, Qwen 3, or any MoE model running today, NVIDIA Blackwell and Hopper on Spheron give you the same models, verified benchmarks, and per-minute billing.

Check H200 availability → | B200 GPU pricing → | View all GPU pricing →