GEM-4C 558 -4%
GPT-61S 555 -3.8%
CL-SN55 553 -3.8%
CL-OP55 551 -4%
GPT-61TERR 547 -3.5%
GEM-4CM 546 -4%
CL-OP55EH 542 -3.9%
GPT-61SOLXH 541 -3.6%
GPT-61SOLH 527 -3.7%
CL-SN55XH 523 -3.5%
GEM-4CL 520 -3.9%
GPT-61TERRXH 507 -3.4%
CL-OP55H 506 -4%
GPT-61SOLM 500 -3.5%
KIMI-K31P 499 -3.5%
GEM-4AR 498 -3.9%
GPT-61TERRH 490 -3.2%
CL-OP55M 483 -3.6%
MUSE-WONDFUL 479 -3.8%
CL-SN55H 467 -3.3%
CL-HK55 463 -2.9%
GPT-61TERRM 462 -3.1%
GPT-6A 451 -4.7%
MUSE-WOND 451 -3.6%
CL-HK55XH 450 -3%
GPT-6ASTRXH 440 -4.8%
GROK-47 440 -3.1%
GROK-47H 438 -2.9%
GPT-6ASTRH 437 -4.8%
DSK-V-340615 436 -2.9%
QWEN-38X 428 -2.9%
GPT-61SOLL 423 -3.2%
GLM-53 420 -3%
MIMO-V26P 420 -2.6%
CL-HK55H 419 -2.8%
LING-31F1019 417 -2.6%
GEM-38F 416 -2.8%
GPT-6ASTRM 415 -4.6%
MMAX-M31F 414 -2.8%
GEM-38FM 412 -2.8%
GPT-61TERRNO 411 -2.8%
GROK-47L 410 -2.8%
DSK-V-3E48BD 410 -2.6%
GPT-6L 407 -2.6%
GLM-53F 407 -2.6%
STEP-5 405 -2.6%
CL-FABL51XH 403 -4.7%
GPT-61TERRL 403 -2.9%
MIMO-V26F 401 -2.4%
CL-FABL51 399 -4.5%
CL-SN55M 395 -2.9%
CL-OP55L 393 -3.4%
MIST-LARG4 390 -2.7%
CL-FABL51H 386 -4.5%
CL-HK55M 386 -2.3%
GPT-6LUNAXH 377 -2.6%
MUSE-SPAROPE 375 -2.6%
MIST-LARG4M 371 -2.4%
DSK-V4P 365 -2.4%
GPT-6LUNAH 365 -2.1%
GLM-53L 364 -2.4%
CL-SN55L 362 -2.7%
GEM-38FL 361 -2.4%
GPT-6ASTRL 359 -4.5%
CL-FABL51M 359 -4.5%
QWEN-827B 349 -2.2%
CL-HK55L 347 -2%
MMAX-M3 345 -2%
GPT-6LUNAM 344 -2.3%
DSK-V-02811D 344 -2%
K2-HORI375BA 342 -2%
REFL-BEAM 338 -2%
QWEN-827BM 332 -1.8%
CL-FABL51L 329 -4.4%
QWEN-827BL 320 -1.8%
NEMO-35SUPEV 313 -1.9%
INKL 295 -1.7%
NVID-NEMO3UL 295 -1.3%
GEM-35FLITE 295 -1.3%
GPT-6LUNAL 287 -1.4%
GPT-6LUNANON 286 -1%
NANB-333B 285 -2.7%
QWEN-827BNON 282 -1.4%
COMM-APLUS 266 -1.1%
MUSE-GLIM 263 -1.1%
NEMO-35LIGH 261 -0.8%
GRAN-428B 258 -0.8%
MIST-SMAL4 252 -0.4%
INKL-SM 235 -2.5%
GEMM-431B 225 —
GEMM-4E4B 225 —
GEMM-4E4BNON 225 —
LING-30TINY 225 —
MINI-33B 225 —
NORT-MCODE 225 —
NOVA-MICR 225 —
NVID--ADE3B4 225 —
QWEN-54B 225 —
QWEN-54BNONR 225 —
TINY-AYAGLOB 225 —
GEM-4C 558 -4%
GPT-61S 555 -3.8%
CL-SN55 553 -3.8%
CL-OP55 551 -4%
GPT-61TERR 547 -3.5%
GEM-4CM 546 -4%
CL-OP55EH 542 -3.9%
GPT-61SOLXH 541 -3.6%
GPT-61SOLH 527 -3.7%
CL-SN55XH 523 -3.5%
GEM-4CL 520 -3.9%
GPT-61TERRXH 507 -3.4%
CL-OP55H 506 -4%
GPT-61SOLM 500 -3.5%
KIMI-K31P 499 -3.5%
GEM-4AR 498 -3.9%
GPT-61TERRH 490 -3.2%
CL-OP55M 483 -3.6%
MUSE-WONDFUL 479 -3.8%
CL-SN55H 467 -3.3%
CL-HK55 463 -2.9%
GPT-61TERRM 462 -3.1%
GPT-6A 451 -4.7%
MUSE-WOND 451 -3.6%
CL-HK55XH 450 -3%
GPT-6ASTRXH 440 -4.8%
GROK-47 440 -3.1%
GROK-47H 438 -2.9%
GPT-6ASTRH 437 -4.8%
DSK-V-340615 436 -2.9%
QWEN-38X 428 -2.9%
GPT-61SOLL 423 -3.2%
GLM-53 420 -3%
MIMO-V26P 420 -2.6%
CL-HK55H 419 -2.8%
LING-31F1019 417 -2.6%
GEM-38F 416 -2.8%
GPT-6ASTRM 415 -4.6%
MMAX-M31F 414 -2.8%
GEM-38FM 412 -2.8%
GPT-61TERRNO 411 -2.8%
GROK-47L 410 -2.8%
DSK-V-3E48BD 410 -2.6%
GPT-6L 407 -2.6%
GLM-53F 407 -2.6%
STEP-5 405 -2.6%
CL-FABL51XH 403 -4.7%
GPT-61TERRL 403 -2.9%
MIMO-V26F 401 -2.4%
CL-FABL51 399 -4.5%
CL-SN55M 395 -2.9%
CL-OP55L 393 -3.4%
MIST-LARG4 390 -2.7%
CL-FABL51H 386 -4.5%
CL-HK55M 386 -2.3%
GPT-6LUNAXH 377 -2.6%
MUSE-SPAROPE 375 -2.6%
MIST-LARG4M 371 -2.4%
DSK-V4P 365 -2.4%
GPT-6LUNAH 365 -2.1%
GLM-53L 364 -2.4%
CL-SN55L 362 -2.7%
GEM-38FL 361 -2.4%
GPT-6ASTRL 359 -4.5%
CL-FABL51M 359 -4.5%
QWEN-827B 349 -2.2%
CL-HK55L 347 -2%
MMAX-M3 345 -2%
GPT-6LUNAM 344 -2.3%
DSK-V-02811D 344 -2%
K2-HORI375BA 342 -2%
REFL-BEAM 338 -2%
QWEN-827BM 332 -1.8%
CL-FABL51L 329 -4.4%
QWEN-827BL 320 -1.8%
NEMO-35SUPEV 313 -1.9%
INKL 295 -1.7%
NVID-NEMO3UL 295 -1.3%
GEM-35FLITE 295 -1.3%
GPT-6LUNAL 287 -1.4%
GPT-6LUNANON 286 -1%
NANB-333B 285 -2.7%
QWEN-827BNON 282 -1.4%
COMM-APLUS 266 -1.1%
MUSE-GLIM 263 -1.1%
NEMO-35LIGH 261 -0.8%
GRAN-428B 258 -0.8%
MIST-SMAL4 252 -0.4%
INKL-SM 235 -2.5%
GEMM-431B 225 —
GEMM-4E4B 225 —
GEMM-4E4BNON 225 —
LING-30TINY 225 —
MINI-33B 225 —
NORT-MCODE 225 —
NOVA-MICR 225 —
NVID--ADE3B4 225 —
QWEN-54B 225 —
QWEN-54BNONR 225 —
TINY-AYAGLOB 225 —

Delayed ticker data. Last calculated 5 Nov 2026, 11:04 pm Sydney time.

← Back to feed

GLM-5.2 on AMD MI355X: 2,626 Tokens/Second at 2x Lower Cost Than Blackwell as AI Closes the NVIDIA Software Gap

Wafer AI published benchmark results for GLM-5.2 running on AMD MI355X hardware via TensorWave. The headline: 2,626 tokens per second per node at a 2.4 RPS saturation point, with latency staying under 5 seconds TTFT across all load levels. Cost per GPU: roughly 2.75x cheaper than NVIDIA B300 (MI355X vs B300 comparison).

That translates to over 2x the inference value per dollar compared to Blackwell, at 80% of B200 throughput.

The Numbers

RPSAggregate tok/s/nodeTTFT p50 / p95Success
0.54490.59s / 0.60s100%
1.09740.60s / 0.81s100%
1.51,9130.62s / 1.03s100%
2.01,9440.62s / 1.05s100%
2.252,0890.63s / 1.23s100%
2.4 (saturation)2,6260.81s / 2.22s100%

Single-stream throughput following Artificial Analysis methodology: 213 tok/s on 10K input / 1.5K output. Not a leaderboard topper, but competitive for an AMD-based deployment.

How the Optimization Was Done

The ROCm software stack for frontier MoE models does not come ready out of the box. Wafer quantized the base bf16 GLM-5.2 weights to MXFP4 using AMD Quark. Their lossless test:

EvalFP8 baselineMXFP4Delta
GSM8K (5-shot)0.9650.955-0.010
GPQA-Diamond0.92170.9026-0.019
tau2 macro0.8190.834+0.015

Effectively lossless on capability while enabling hardware-native MXFP4 paths. Inference framework: sglang. vLLM had no working MXFP4 + GlmMoeDsa path on ROCm; ATOM degraded at long context.

Why This Matters

NVIDIA’s hardware advantage is real. But the advantage is not primarily in silicon: the MI355X has comparable specs to the B300. The gap has been in software — ROCm tooling, day-0 model support, inference framework maturity. Providers historically saw weeks of engineering work before achieving usable throughput on AMD hardware, by which point a newer model had shipped.

What Wafer is demonstrating is that AI-assisted kernel optimization is closing that gap in real time. The optimization loop that once required specialist compiler engineers can increasingly be handled by coding agents working against a clear target metric.

For the inference economics argument, this is significant. GLM-5.2 is a 744B open-weight model that leads the open-weight Intelligence Index. Running it at over 2x cost efficiency versus Blackwell, at 80% of the throughput, changes the unit economics of deploying frontier-tier open-weight inference.

The Broader Dynamic

DeepSeek V4 on Huawei Ascend was Jensen Huang’s stated nightmare scenario. GLM-5.2 on AMD MI355X is the same argument applied to a less politically charged hardware pair. If frontier open-weight models can be served at Blackwell-comparable quality for half the cost on AMD hardware, the NVIDIA inference moat narrows from both directions simultaneously: high-end proprietary models (Blackwell required, vendor-locked), and open-weight frontier models (AMD-viable, software-unlocked).

The 2x cost gap is today’s number. The engineering work to close the software gap is ongoing.