Unmarked gray slab on a plain studio floor representing the measured cache tier cliff on a Ryzen 9 5900X

56.99 GB/s in Cache, 12.93 GB/s in DRAM: What a Measured Cache Cliff Says About the Tier You Paid For

Two price signals, opposite directions

TrendForce published its fourth-quarter memory outlook on September 30, 2026. Conventional DRAM contract prices are projected to rise 10 to 15 percent quarter over quarter, NAND flash at 15 to 20 percent. Enterprise SSDs are the only category where it expects the increase to accelerate, on demand from cloud providers.

The spot market is not confirming that picture. The reported average spot price for a 16 Gb DDR5 chip on November 5 was $60.667, down 1.09 percent from the previous session. Contract prices for the same chip sit in the $48 to $50 range, with a reported Korean fixed price of $50 in October, up from $48.

Those two readings are not the same measurement and should not be averaged. A contract price is negotiated between a fabricant and a large buyer on a quarterly cycle. A spot price is what someone paid for a tray of chips on an afternoon. Contract prices lag spot at both cycle peaks and troughs, so the gap between them is the signal. Contract momentum is decelerating from 90 to 95 percent quarter over quarter in Q1 2026 to 13 to 18 percent in Q3, and the Q4 print is a 10 to 15 percent forecast against a spot market that ticked down in one session.

Module makers are not reading the small number as relief. TeamGroup CEO Chen Ching-wen said on an earnings call on November 2 that the most severe shortage runs from Q4 2026 through Q2 2027, that supply and demand may not balance until the second half of 2028, and that no downward price trend is expected in 2027. One soft session does not change a capacity allocation already committed elsewhere.

That is the market side. What I can measure is what a memory tier does once installed, and the answer has a cliff in it.

What I measured on the serving box

The machine that serves my models is a Ryzen 9 5900X with 64 GB of DDR4 at 2667 MT/s in two modules. I ran a single-threaded copy pass on that box over SSH, not on the machine I am writing from. The box I sent the request from is a Ryzen 7 2700X with 16 MB of L3, and a run there would have described a different chip. The numbers below are the 5900X’s.

A copy pass touches memory twice: it reads the source and writes the destination. Every figure below is reported as copy bandwidth and as memory traffic, because quoting one as the other is how these numbers get inflated. Each buffer size is warmed with 2 passes, then measured over 11, and the median of those 11 is recorded. I ran the whole sweep five times on five separate connections. The range column is the spread of the five run-medians, not the spread of individual passes.

The small-buffer medians landed on the same value in all five runs, which does not mean the passes were identical: at 4 KB individual passes ranged from 5.12 to 13.65 GB/s.

Buffer Copy GB/s (median of 5 runs) Range across runs Memory traffic GB/s Tier the working set sits in
4 KB 10.24 identical 20.48 per-core L1/L2 boundary
16 KB 27.31 identical 54.62 L1
32 KB 40.96 identical 81.92 L1
64 KB 50.41 50.41 to 54.61 100.82 L1
128 KB 59.58 56.99 to 62.41 119.16 per-core L2 (512 KB)
256 KB 56.99 55.77 to 56.99 113.98 per-core L2, at capacity
512 KB 20.40 19.64 to 20.48 40.80 spills to shared L3
1 MB 20.93 20.64 to 20.97 41.86 L3
2 MB 21.12 19.49 to 21.12 42.24 L3
4 MB 21.17 20.99 to 21.20 42.34 L3
6 MB 21.23 21.18 to 21.23 42.46 L3
8 MB 21.11 20.95 to 21.14 42.22 L3, inside one CCD
16 MB 21.14 20.85 to 21.18 42.28 L3, inside one CCD
32 MB 19.15 16.67 to 19.99 38.30 working set spans both CCDs
64 MB 17.95 16.20 to 18.14 35.90 DRAM
128 MB 17.27 16.93 to 17.52 34.54 DRAM
256 MB 16.92 13.66 to 17.09 33.84 DRAM
512 MB 16.61 14.09 to 16.81 33.22 DRAM
1024 MB 15.72 12.93 to 16.79 31.44 DRAM

The peak is at 128 KB: 59.58 GB/s of copy across the five runs, with individual runs landing between 54.61 and 62.41. The working set there is 256 KB, which fits inside one core’s 512 KB of L2. At 512 KB the working set is 1 MB, which no single core can hold, and the rate falls to 20.40 GB/s. That is a 65.8 percent drop.

The two numbers in this article’s title are the medians from the first run and sit inside the ranges above. I am reporting ranges because the absolute rates move between runs and the ratio does not. The five runs produced drops of 64.2, 65.5, 65.8, 65.8 and 67.2 percent. A single run gives a number. Five runs give a shape, and the shape is the part that survives being run again.

The second cliff is smaller and less stable. Up to 16 MB the working set fits inside one CCD’s 32 MB of L3 and holds around 21 GB/s in every run. At 32 MB the working set is 64 MB, the whole L3 pool, and the copy crosses between the two CCDs through the Infinity Fabric. The median there is 19.15 GB/s, 9.4 percent below the plateau. Across the five runs that drop was 4.1, 8.2, 9.5, 11.8 and 21.1 percent. One run would have called that a cliff. Five runs call it a shoulder. Cache that exists is not cache a single core can reach, and crossing between cores costs far less than crossing out of L2 does.

On the DRAM rows the median copy rate at 1 GB is 15.72 GB/s, with runs between 12.93 and 16.79. Counted as memory traffic that is 73.6 percent of the theoretical channel at 1 GB and 84.1 percent at 64 MB. Counted as data moved it is 36.8 percent and 42.0 percent. The channel arithmetic is mine: 2667 MT/s, 8 bytes per transfer, two channels, 42.7 GB/s, read from the machine rather than a vendor page.

Cache-tier rows are rates achieved inside cache. Only the rows from 64 MB upward are actual DRAM traffic, and only those can be expressed as a share of the channel. A buffer that never reaches DRAM cannot be a percentage of the DRAM channel, though published tables present it as one.

Why the cliff matters for a model that does not fit

A model’s weights do not fit in 512 KB or in 64 MB. A 125 billion parameter mixture of experts quantized to 4 bits is roughly 62 GB of weights, and every token that touches an expert reads that expert from somewhere. On a build that splits the model across GPU, system RAM and storage, system RAM is the tier that holds the experts. The measured rate for that tier on my hardware is 12.93 GB/s of copy.

The cache tiers are faster by a factor of 3.79 at the median and between 3.44 and 4.41 across runs, and they are useless for weights. They matter for the parts of a request that do fit: the attention state for a short context, the tokenization buffers, the routing decision that picks which expert to load. That is why cache policy is the lever on a small model. When the model fits inside a tier, the price you paid is irrelevant. When it does not, the tier you get is the one that costs money per gigabyte, and the price is the whole story.

That connects the two halves. The market is pricing a tier between 3.4 and 4.4 times slower than the tier nobody can use, because the same wafer capacity is being pulled toward HBM. That is how a 16 Gb DDR5 chip went from about $8.70 in October 2025 to a reported $48 contract price in September 2026, roughly 450 percent in eleven months.

What a gigabyte of that tier costs right now

Three listings I verified, all desktop UDIMM kits, the correct form factor for a desktop board. The Kingston FURY Beast RGB 64 GB kit (2×32 GB, DDR5-6000 CL36, AMD EXPO) is listed at about $1,199.99 as of the tracker reading, which is $18.75 per gigabyte. Its lowest recorded price on that ASIN was $231.86 on July 12, 2025. A lowest-ever against a current listing compares two measurement types, so I am naming which is which.

The Crucial Pro 64 GB kit (2×32 GB, 5600 MT/s, UDIMM 288-pin, desktop) is at about $1,177.00 on the November 7 scan, which is $18.39 per gigabyte. Its lowest recorded price was $129.00 in September 2023. The Crucial Pro 32 GB kit (2×16 GB, 6000 MHz, UDIMM, desktop) is at about $609.99 on the October 22 scan, which is $19.06 per gigabyte.

Those three figures are flat per gigabyte: $18.39, $18.75, $19.06. Price indexes report that 64 GB kits carry a higher per-gigabyte price than 32 GB kits on the reasoning that large kits have thinner demand and fatter margins. On the listings I could resolve, the spread runs the other way and it is under 4 percent.

Here is the arithmetic that makes the cliff and the price the same argument. The $1,199.99 kit delivers a median 15.72 GB/s of copy bandwidth, which is $76.34 per gigabyte per second, or $38.17 counted as memory traffic. Across the five runs that figure ranged from $71.47 to $92.81. The 6 MB of L2 on the same CPU delivers 56.99 GB/s at no marginal cost, because it came with the chip. That comparison is not fair, since L2 cannot hold a model, but it is why a build that keeps its hot state inside a cache tier behaves differently from one that does not.

What to actually do with this

If you are buying memory for a local build in November 2026, the decision is not whether to spend the money. It is whether the tier you are buying is the tier your workload reaches. A kit sized so the model plus its attention state fits inside one tier behaves like a different machine than one sized so the model spills. The measured difference between those two situations on my hardware is between 3.4 and 4.4, and it held in every run. It is not a property of the memory module. It is a property of where the data lands.

On the price side, a 10 to 15 percent quarterly contract increase against a spot market that fell 1.09 percent in one session is not a forecast of your street price. Street pricing depends on who has stock and when they bought it. On the same Crucial Pro 64 GB ASIN, the used price was $1,049.99 on the November 6 scan against $1,177.00 new on November 7. Both are true at the same time.

The capacity allocation behind that outlook is the mechanism in HBM is eating your RAM. The cache-policy side, where a model that fits inside a tier stops being a bandwidth problem at all, is covered in the memory wall as a cache policy problem. The tier below system RAM behaves differently than its rated sequential spec suggests, which is measured in what block size says about the KV cache tier.

The Kingston FURY Beast RGB 64 GB kit (2×32 GB, DDR5-6000 CL36, desktop DIMM) is the one that matches a build sized to hold a 4-bit 125B model in system RAM, at about $1,199.99 as of the November price-tracker reading. The Crucial Pro 64 GB kit (2×32 GB, 5600 MT/s, UDIMM 288-pin, desktop) is the cheaper per gigabyte of the three at about $1,177.00 on the November 7 scan, and it is a UDIMM rather than a laptop kit, which is the distinction that matters on a desktop board.

If the model fits in 32 GB, the Crucial Pro 32 GB kit (2×16 GB, 6000 MHz, UDIMM, desktop) at about $609.99 from the October 22 scan is the same tier at half the capacity. The measured bandwidth does not change with the number of gigabytes you bought. It depends on whether the working set fits in a cache tier, and on a 32 GB build with a model that fits, part of it does.

The tier below RAM serves a quantized model that does not fit in 64 GB. The TEAMGROUP MP44 2 TB M.2 2280 PCIe 4.0 drive is at about $325.99 on the October 21 scan, which is $163 per terabyte. Against the 15.72 GB/s median measured on the DRAM tier, a drive listed at up to 7000 MB/s sequential is a different tier entirely, and block size decides which of its two published rates you get. Prices here are price-tracker readings with their scan dates named, not live listings: Amazon served an interstitial to the browser the day I checked.

The measured cliff is the part of this that is not a market opinion. 59.58 GB/s inside a cache a single core can reach, 20.40 GB/s the moment the working set crosses 512 KB, 15.72 GB/s once it reaches the memory you paid $18.75 per gigabyte for. Those three numbers came from the same machine across five separate runs, and the ratio between the first and the second held between 64.2 and 67.2 percent in every one of them. The rates are not reproducible to the digit. The cliff is, and that is why two builds with the same parts list are not the same build.

4 Comments

Leave a Reply

Your email address will not be published. Required fields are marked *