Step 5 Preview: Self-Hosting Estimate
Step 5 Preview (StepFun) on different GPUs: configuration, estimated performance and use cases.
Data updated: 2026-09-16B300 288GB
288 GB · 8,000 GB/s · Top-tier Blackwell card (no official channels to China)- Suggested precision
- FP16/BF16 (full precision)
- GPUs needed
- 8priced as 1 × 8-GPU HGX node
- Est. decode speed
- 1,185 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 2,015 tok/sin 43,333 tok/s · many concurrent requests
- Hardware list price
- $497,836Cloud rental (live) ×8: $79.52–$120.01/hr (2026-10-24)
- Hardware market price
- $1,048,061
- Typical use
- Enterprise-grade ultra-large-scale training and high-concurrency inference clusters
GB200 192GB (single card equivalent)
192 GB · 8,000 GB/s · Top-tier Blackwell card (no official channels to China)- Suggested precision
- FP16/BF16 (full precision)
- GPUs needed
- 8
- Est. decode speed
- 1,185 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 2,015 tok/sin 48,148 tok/s · many concurrent requests
- Hardware list price
- $239,748–$359,623
- Hardware market price
- $718,670–$1,437,341
- Typical use
- Ultra-large-scale trillion-parameter model training and inference clusters (networked with Grace CPU)
B200 180GB
180 GB · 8,000 GB/s · Top-tier Blackwell card (no official channels to China)- Suggested precision
- INT8
- GPUs needed
- 8priced as 1 × 8-GPU HGX node
- Est. decode speed
- 2,370 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 4,030 tok/sin 43,333 tok/s · many concurrent requests
- Hardware list price
- $240,769–$563,955Cloud rental (live) ×8: $50.02–$60.02/hr (2026-10-24)
- Hardware market price
- $449,169–$524,031
- Typical use
- Enterprise-grade large-scale online inference and training
H200 141GB
141 GB · 4,800 GB/s · Top-tier export-controlled card (channels to China are volatile)- Suggested precision
- INT8
- GPUs needed
- 8priced as 1 × 8-GPU HGX node
- Est. decode speed
- 1,422 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 2,418 tok/sin 19,047 tok/s · many concurrent requests
- Hardware list price
- $248,598–$451,084Cloud rental (live) ×8: $27.41–$40.43/hr (2026-10-24)
- Hardware market price
- $224,585–$233,568
- Typical use
- Enterprise-grade production environment high-concurrency inference
H100 SXM5 80GB
80 GB · 3,350 GB/s · Top-tier export-controlled card (no official channels to China)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8priced as 1 × 8-GPU HGX node
- Est. decode speed
- 1,985 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 3,375 tok/sin 19,047 tok/s · many concurrent requests
- Hardware list price
- $216,521–$834,005Cloud rental (live) ×8: $17.09–$29.98/hr (2026-10-24)
- Hardware market price
- $404,252–$419,224
- Typical use
- Enterprise-grade production environment high-concurrency inference/training
H800 SXM 80GB
80 GB · 3,350 GB/s · Top-tier export-controlled card (discontinued)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8priced as 1 × 8-GPU HGX node
- Est. decode speed
- 1,985 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 3,375 tok/sin 19,047 tok/s · many concurrent requests
- Hardware list price
- $293,457–$311,424
- Hardware market price
- $404,252–$419,224
- Typical use
- Enterprise-grade production deployment (original China-compliant model, now discontinued)
H20 96GB
96 GB · 4,000 GB/s · Export-compliant special supply card (significantly reduced compute power)- Suggested precision
- INT8
- GPUs needed
- 8
- Est. decode speed
- 1,185 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 2,015 tok/sin 2,850 tok/s · many concurrent requests
- Hardware list price
- n/a
- Hardware market price
- $95,823–$131,756
- Typical use
- Suboptimal choice in export-compliant environments, medium-to-large scale inference deployment (FP16 compute power is only about 15% of H100, but higher memory bandwidth offers an advantage in decode-intensive scenarios)
A800 PCIe 80GB
80 GB · 1,935 GB/s · Compliant special supply card (export compliance status adjusted multiple times)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8
- Est. decode speed
- 1,147 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 1,170 tok/sin 3,605 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $119,778–$124,570
- Hardware market price
- $173,679
- Typical use
- Medium-scale inference/training deployment in export-compliant environments
Ascend 910C 128GB
128 GB · 3,200 GB/s · Huawei dual-chip data center card (domestically produced)- Suggested precision
- INT8
- GPUs needed
- 8
- Est. decode speed
- 948 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 1,612 tok/sin 14,483 tok/s · many concurrent requests
- Hardware list price
- n/a
- Hardware market price
- n/a
- Typical use
- Domestically produced large model training and inference clusters, CloudMatrix and other super-node solutions
Ascend 910B 64GB
64 GB · 400 GB/s · Domestic alternative (not subject to US export controls)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8
- Est. decode speed
- 237 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 403 tok/sin 6,163 tok/s · many concurrent requests
- Hardware list price
- n/a
- Hardware market price
- $131,756–$143,734
- Typical use
- Domestic alternative, not subject to US export controls, medium-to-large scale inference/training deployment, memory bandwidth significantly lower than NVIDIA cards of similar VRAM capacity
RTX PRO 6000 Blackwell 96GB
96 GB · 1,792 GB/s · Professional workstation/server card (China-specific version is 6000D 84GB)- Suggested precision
- INT8
- GPUs needed
- 8
- Est. decode speed
- 531 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 542 tok/sin 5,778 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $128,309Cloud rental (live) ×8: $9.62–$19.76/hr (2026-10-24)
- Hardware market price
- n/a
- Typical use
- Single-card large VRAM inference, local deployment of medium MoE models, private all-in-one machine
RTX PRO 5500 Blackwell 84GB
84 GB · 1,398 GB/s · Professional workstation card (specs are NVIDIA preliminary data)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8
- Est. decode speed
- 828 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 845 tok/sin 5,223 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- n/a
- Hardware market price
- $101,812–$113,789
- Typical use
- Workstation large VRAM inference, departmental private deployment
RTX PRO 5000 Blackwell 48GB
48 GB · 1,344 GB/s · Professional workstation card- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8
- Est. decode speed
- 796 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 812 tok/sin 2,981 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- n/aCloud rental (live) ×8: $4.85–$9.63/hr (2026-10-24)
- Hardware market price
- $64,680
- Typical use
- Small to medium model inference, development testing, single-machine multi-card private deployment
L40S 48GB
48 GB · 864 GB/s · Compliant inference card- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8
- Est. decode speed
- 512 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 522 tok/sin 4,183 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $53,900–$77,856Cloud rental (live) ×8: $3.33–$6.42/hr (2026-10-24)
- Hardware market price
- $53,900–$77,856
- Typical use
- SME teams building inference services, medium-low concurrency
L20 48GB
48 GB · 864 GB/s · Compliant inference card- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8
- Est. decode speed
- 512 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 522 tok/sin 1,381 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $27,965–$28,125
- Hardware market price
- $29,945–$41,922
- Typical use
- SME teams building inference services, medium-low concurrency
RTX 5090 32GB
32 GB · 1,792 GB/s · Consumer flagship (not sold in China, 5090D for domestic market)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 16 (exceeds one node: multi-node needed)
- Est. decode speed
- 2,124 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 2,166 tok/sin 9,684 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- n/aCloud rental (live) ×16: $7.50–$10.74/hr (2026-10-24)
- Hardware market price
- $83,845–$129,361
- Typical use
- Local inference and development for individuals/small teams, multi-card VRAM pooling for quantized MoE
RTX 5090D 24GB
24 GB · 1,344 GB/s · Consumer-grade/small-scale self-built- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 16 (exceeds one node: multi-node needed)
- Est. decode speed
- 1,593 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 1,625 tok/sin 6,864 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $39,524Cloud rental (live) ×16: $7.95–$11.55/hr (2026-10-24)
- Hardware market price
- $67,076–$95,823
- Typical use
- Personal development/small-scale verification/POC, not recommended for production-level high concurrency
RTX 4090D 24GB
24 GB · 1,008 GB/s · Consumer-grade/small-scale self-built- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 16 (exceeds one node: multi-node needed)
- Est. decode speed
- 1,195 tok/ssingle-stream theoretical ceiling · vendor API 229 tok/s
- Est. batched throughput
- out 1,219 tok/sin 6,795 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $31,140–$45,513Cloud rental (live) ×16: $6.43–$7.78/hr (2026-10-24)
- Hardware market price
- $62,285–$71,867
- Typical use
- Personal development/small-scale verification/POC, not recommended for production-level high concurrency
FAQ
FAQ
How is the estimated decode speed calculated?
Estimated speed = (memory bandwidth × GPU count) ÷ (active parameters × bytes per parameter): the memory-bound upper limit for one request (roofline). It scales linearly with GPU count and ignores interconnect overhead, batching and framework losses, so real results are usually well below it.
Can I trust the speed numbers for consumer GPUs?
The linear assumption is reasonable for NVLink data-center cards (H/B series) but too optimistic for consumer PCIe cards without NVLink (RTX 5090D/4090D); multi-GPU consumer setups lose far more to communication, so equal-looking numbers do not mean equal performance.
How is “Suggested precision” chosen?
It is the highest precision that fits within 8 GPUs; beyond that it is flagged as multi-node.
Where do the prices come from?
USD converted at ¥6.68 per USD. Cloud rental is the international vast.ai market rate, not a China purchase price.