SK Hynix CMM-Ax with Marvell: 5.5x Throughput Claim Sparks 12% MRVL Rally
CXL PNM moves compute to memory to solve the long-context KV-cache bottleneck. Flow flipped to accumulation today, options chase upside, but adoption hinges on server and software support.
SK Hynix announced CMM-Ax, an ASIC-based CXL-PNM solution built with Marvell, claiming up to 5.5x higher throughput than conventional GPU-only systems for long-context LLM inference. The market reacted fast on Aug 4: MRVL +12.6% to $218.09 and SKHY +5.6% to $150.68 on heavy volume.
This is not just another memory module. It is an attempt to fix where inference actually bottlenecks today.
What CMM-Ax is
CMM-Ax is described by SK Hynix as the world's first PNM (Processing Near Memory) solution with compute added to CXL-based memory. Key specs from the company: CXL Type 3 / PCIe Gen4 x8, up to 512GB capacity, controller with ML acceleration HW and Cortex-A72 dual cores. See the original spec sheet in SK Hynix's CXL post.
Why it matters: long-context decode is KV-cache bound. Unlike model weights, you cannot batch your way out of KV-cache bandwidth. PNM puts compute next to memory and minimizes data movement. That is the same bottleneck researchers have flagged as the open problem in LLM inference.
Marvell's role is the custom ASIC controller. That fits Marvell's AI infrastructure push where Data Center was 76% of Q1 FY27 revenue at $1.83B.
Flow today: accumulation, not distribution
Both names were in distribution in late July. Today flipped.
| Ticker | Price | Change | VWAP | Dist Score | Regime | Volume |
|---|---|---|---|---|---|---|
| MRVL | $218.09 | +12.55% | $215.44 | 8.5 | calm | 17.1M |
| SKHY | $150.68 | +5.58% | $149.47 | 14.0 | calm | 19.4M |
MRVL traded above VWAP all day with signed flow OFI -0.04, 91 block prints ($25.1M buys vs $31.2M sells). SKHY also above VWAP with no systematic distribution flag. This is buyers in control after a washout.
For context on memory positioning, see Memory Is Not a Commodity: Why MU and SK Hynix Are Priced for a Recession That Isn't Coming and SK Hynix Printed 76% Margins. The Market Sold It Anyway.
Options: upside chasing, not hedging
MRVL:
- Put/call volume 0.22 today, 82,483 calls vs 18,452 puts, total 100,935 contracts
- Positioning: net GEX +$49.4M long-gamma, flip $180.44, call wall $240, put wall $200, ATM IV 105.3%
- Largest OI adds: Nov $310 call +3,864, Aug $240 call +3,604, Aug $250 call +2,827, Jan 2028 $300 LEAPS call +2,253
- LEAPS total 268k OI, call OI 157k vs put 112k, put/call 0.71, long-dated bullish tilt
SKHY:
- Put/call volume 0.75, 27,626 calls vs 20,718 puts, total 48,344 contracts
- Positioning: net GEX -$5.4M short-gamma, flip $154.31, call wall $200, put wall $125, ATM IV 116.4%
- Largest OI adds: Dec $210 call +12,813 (now 37k OI), Jan 2027 $250 call +5,378, Jan 2027 $180 call +5,045
- LEAPS total 32k OI, call 20k vs put 11k, put/call 0.56
No unusual put buying that would signal a fade. Flow is momentum chasing calls into strength. MRVL's long-gamma regime suggests pinning near $200-$240 into Aug expiry, SKHY's short-gamma suggests amplified moves.
Is this a game-changer?
Directionally yes, commercially still early.
Why it could be:
- Inference economics are shifting from training FLOPs to memory bandwidth and KV-cache capacity. CXL memory expansion is one of few architectural fixes that scales beyond HBM.
- Validates two theses: SK Hynix beyond HBM into CXL, and Marvell custom silicon as AI infrastructure. The memory leverage unwind note flagged positioning was washed out before this catalyst.
Why it is not instant:
- CXL ecosystem needs server OEMs, CPU support, and software (vLLM, SGLang, PyTorch) to expose PNM. Adoption is a 2026-27 ramp, not Q3 revenue.
- Competition: Samsung CMM-D, Micron, Astera Labs CXL controllers, plus HBM4 and LPDDR5X approaches.
- The 5.5x claim is workload specific vs GPU-only baseline. Real gains depend on context length, batch, and model.
Valuation anchor:
- SKHY: ~$1.01T market cap, forward P/E 4.27, trailing 19.5
- MRVL: $174B market cap, forward P/E 31.0, trailing 64.6, priced for AI growth
What to watch
- Server qualification: any hyperscaler or OEM design win for CMM-Ax
- Software: vLLM or SGLang adding CXL-PNM offload paths
- Marvell custom ASIC pipeline commentary next earnings
- Options levels: MRVL gamma flip $180.44 as support, $240 call wall resistance, SKHY $125 put wall and $200 call wall
More research at bargo.ai/research.