What actually happened
At the OCP Global Summit in San Jose, October 12 to 15, 2026, Samsung and SK hynix each showed CXL memory pools serving KV cache across more than one inference server. SK hynix unveiled a second-generation CMM-DDR5, 256 GB, on CXL 3.2, which it says more than doubles bandwidth to 80 GB/s. Samsung claimed an L3.5 composable memory tier delivered 5 to 7 times KV cache utilization. Those are vendor figures from vendor booths.
TechInsights, covering the same event, stated plainly that CXL memory pooling has moved from forecast to production in 2026, running on next-generation Intel and AMD server platforms. The CXL tier is not faster, and at current street prices not cheaper either. It is a capacity tier with a fixed per-access penalty. I measured that penalty, and one published deployment number settles the rest.
The published tier table, and where it contradicts itself
Every buying guide for 2026 lays out the same four tiers. QSCompute’s August 2026 guide is representative:
| Memory tier | Typical capacity | Latency | Cost per GB (Q3 2026) | What caps it |
|---|---|---|---|---|
| GPU HBM3e | 80 to 192 GB per GPU | about 1 microsecond | $15 to $25 | Soldered and fixed |
| Native DDR5 RDIMM | Up to about 3 TB per socket | 80 to 110 ns | $10 to $14 | DIMM slot count |
| CXL-attached DRAM | 2 to 4 TB per socket | 170 to 250 ns | $9 to $12 | Needs a CXL-capable CPU |
| NVMe SSD | 15 to 122 TB per drive | 10 to 100 microseconds | $0.05 to $0.10 | Too slow for live KV cache |
Look at the two price columns. The CXL tier is quoted at $9 to $12 per gigabyte and the native tier at $10 to $14. The bands overlap. The CXL tier is not cheaper per gigabyte. It is cheaper per slot, because it puts capacity where a DIMM slot cannot go.
Street prices confirm it. A Samsung CMM-D 256 GB module lists at roughly $2,600 to $3,200, or $10.16 to $12.50 per gigabyte. Micron’s CZ120 is about $11.72 at 128 GB and $10.94 at 256 GB. SK hynix’s CMM-DDR5 256 GB is $10.94 to $12.89. Every one lands inside the native band the same guide quotes. There is no arbitrage. There is only the slot.
What I measured on my own hardware
I cannot buy a CXL module, so I measured the tier I have: local DRAM. The machine is my own host, an AMD Ryzen 7 2700X with 32 GB of DDR4 and 16 MiB of L3 across two instances, confirmed through lscpu. It has no CXL root complex, which is part of the finding.
I ran two loops, which answer different questions, so I label them separately. The latency loop is a random pointer chase over a working set of N bytes of 8-byte slots, 200,000 accesses per pass, 11 passes, median. Accesses do not overlap, so this is what one dependent read costs. The throughput loop is a memcpy of N bytes, 200 repetitions per pass, 9 passes, median. Repetitions overlap, so that measures sustained rate.
| Working set | Median ns per dependent access (latency loop) | Median copy GB/s (throughput loop) |
|---|---|---|
| 8 KB | 2.91 | not measured |
| 64 KB | 3.93 | not measured |
| 256 KB | 4.67 | not measured |
| 1 MB | 9.20 | 36.68 |
| 4 MB | 11.61 | not measured |
| 8 MB | 22.35 | 16.89 |
| 16 MB (L3 size) | 55.70 | not measured |
| 32 MB | 76.59 | 13.02 |
| 128 MB | 93.73 | 11.75 |
| 512 MB | 104.80 | 12.18 |
The DRAM tier on this box costs 104.80 ns per dependent access. That is inside the published native range of 80 to 110 ns, the check I wanted before trusting either number. The jump from 8 MB to 16 MB is a factor of 2.49, and from 256 KB to 1 MB a factor of 1.97. Those are cache boundaries, not memory speed. Counted as memory traffic, a copy touches each byte twice, so the DRAM-side figures are 33.78, 26.04, 23.50 and 24.36 GB/s. The 1 MB row never reaches DRAM, so it is a rate inside cache, not a share of the channel.
Why the penalty amortizes, and why it does not
Take a 1 MB KV cache block: 16,384 cache-line reads. Read one at a time, dependent, at 104.80 ns each: 1.717 milliseconds. Read them overlapped at the measured copy rate of 12.18 GB/s: 86.1 microseconds. The ratio is 19.9, meaning about 20 accesses in flight get the throughput number instead of the latency number. This is arithmetic from the two loops above, not a measurement.
The same for the CXL tier. A PCIe 5.0 x8 link is 256 GT/s, which with 128b/130b encoding is 251.9 Gb/s, or 31.51 GB/s per direction. That is my own arithmetic, not a published figure. A 1 MB block serialized at 200 ns per access is 3.277 milliseconds. Overlapped at the link rate it is 33.3 microseconds. The ratio is 98.5, so it needs about 99 accesses in flight to reach its own throughput.
The CXL penalty is a per-access penalty, and a KV cache read is a bulk transfer. If the engine keeps enough requests in flight, the bulk rate is set by the link, and the link is not slower than what a consumer DRAM channel delivers. My measured copy rate is 12.18 GB/s against a theoretical 31.51 GB/s link. On a consumer platform the CXL tier is not meaningfully faster for bulk moves either.
This is why Samsung’s own white paper reads the way it does. It reports that a CXL switch-based pool maintained a high cache hit rate without recomputation overhead and delivered performance comparable to DRAM, while also stating that as KV cache size grows, performance is limited by PCIe bandwidth. Those are not in conflict: one is the overlapped case, the other the point where the link becomes the constraint. The same split shows up on the storage side in what block size says about the KV cache tier: a throughput loop and a latency loop give different per-request costs on the same hardware.
The deployed number is worse than the marketing band
The CXL consortium’s own FMS 2026 panel presentation carries a real configuration rather than a spec band. A Xeon 6 host with 2 TB of native DDR5 against a Liqid memory chassis of 10 TB: total read bandwidth 250 GB/s, memory latency 434 ns.
434 ns against my measured 104.80 ns is a factor of 4.14. That is not the 170 to 250 ns band the guides print. It is what a switch-based pool costs in a deployed system, and the difference is switch hops, not DRAM.
The same panel reports the workload: a 3 TB dataset held in the pool runs in 4 ms, the same dataset on NVMe runs in 46 ms. That is a factor of 11.5, and labels the configuration as 8 times performance per dollar. Both numbers are the panel’s own.
Those two figures are not in tension. The latency number is one dependent access. The workload number is a bulk scan. The scan wins because it keeps enough requests in flight, which is the 98.5 ratio above. A tier costing 4.14 times the latency can still win by 11.5 when the pattern is a scan rather than a chase.
A University of Wisconsin paper on first-generation CXL deployment argues that CPU and DRAM internals, not the CXL protocol, dominate latency and variability. My numbers have the same shape: the per-access cost is set by which cache level the working set fits in, from 2.91 ns at 8 KB to 104.80 ns at DRAM. The protocol adds on top of the memory number. It does not replace it.
What it means for a local build
For most readers the CXL tier is not a purchase. CXL 2.0 requires a CXL-capable root complex: AMD EPYC 9005, Intel Xeon 6 and AmpereOne, with production Xeon 6 platforms typically exposing 64 CXL lanes per socket, configured as four times 16. A Ryzen 7 2700X has no CXL support, so the tier is unreachable here at any price. A module does not become system memory because it fits in a slot.
So the real decision for a home build is between two tiers you can buy. The DRAM tier costs $18.75 to $21.87 per gigabyte at current retail. The cold tier costs about $0.16. That is a factor of 116 to 136, and it is why inference engines tier KV cache to disk instead of paying for more DIMMs. The CXL band sits between them and is unreachable on the platform most readers own.
The part that transfers is the policy, not the hardware. The memory wall on a small machine is a cache policy problem before a capacity problem, and the engines that work well on limited RAM are the ones that decide which pages stay hot. That is the same lesson as the memory wall being a cache policy problem. Capacity without a placement policy is a larger tier the hot path cannot use.
And the number that decides a build is not price per gigabyte. It is bandwidth per dollar, which is where a $7,000 card and a 256 GB module are not comparable. A CXL module buys bytes you cannot reach any other way. It does not buy the rate at which you can touch them.
What to buy, if you are on a consumer platform
At current prices a 64 GB desktop kit is the whole decision. The cheapest confirmed listing is the G.Skill Trident Z5 Neo RGB 64 GB (2x32GB) DDR5-6000 CL30 UDIMM kit at $1,199.99, which is $18.75 per gigabyte. The Kingston FURY Beast RGB 64 GB (2x32GB) DDR5-6000 CL30 desktop kit is at $1,399.99, or $21.87 per gigabyte. Both are UDIMM kits, the correct form factor for a desktop board; a SODIMM kit is laptop memory.
Both figures are third-party new prices from a price-history tracker, scanned November 4 and October 21, 2026, quoted as of November 5, 2026. The Kingston kit shows out of stock at the Amazon price, so treat these as tracker figures, not confirmed listing prices.
The cold tier is where the money is not spent. A WD_Black SN7100 2 TB NVMe drive is at $329.99 on the same tracker as of October 18, 2026, which is $0.16 per gigabyte. That is 136 times cheaper than the Kingston kit, and the tier serving frameworks push stale KV entries into. The penalty is 10 to 100 microseconds per access instead of 104.80 nanoseconds: a factor of 95 to 954 in latency against 136 in price. On a bulk overlapped read that latency factor collapses into a bandwidth question and the cheap tier wins. On the hot path it does not.
The OCP announcements are a real production milestone. They are not a speed improvement, and at current street prices not a price improvement either. They are the only way past the slot count. If your platform supports CXL and you are capacity-bound, that is worth buying. If you are latency-bound, no module changes that. The DRAM price squeeze is what put consumer kits at $18 to $22 per gigabyte, which is why the CXL band looks competitive against desktop retail rather than against the server RDIMM pricing it was quoted against. The two bands describe different buyers.
Products for this workload
G.Skill Trident Z5 Neo RGB 64 GB (2x32GB) DDR5-6000 CL30 UDIMM โ the hot tier, at $1,199.99, $18.75 per gigabyte.
Kingston FURY Beast RGB 64 GB (2x32GB) DDR5-6000 CL30 desktop kit โ same capacity, $1,399.99, $21.87 per gigabyte.
WD_Black SN7100 2 TB NVMe M.2 2280 โ the cold tier, at $329.99, $0.16 per gigabyte. This is the tier the placement policy actually writes to.


Pingback: 7,200 GB/s of Link, 16 TB of Pool: What a Photonic Memory Tier Says About Bandwidth per Byte - GenX AI Tools