A dark server rack in an empty plain room

The Cache on a Separate Die: What Fujitsu Monaka Estimates About Per-Core Bandwidth

What Fujitsu actually announced

Fujitsu opened sales on the FUJITSU-MONAKA server line in November 2026. The chip itself was detailed at Hot Chips on August 24, and the architecture is unusual for a reason that has nothing to do with core count: the entire last-level cache sits on a separate 5nm die, stacked face-to-face under the 2nm compute dies, along with the low-dropout regulators that feed each core. Four core dies carry 36 cores each. The 2nm silicon is under 30 percent of total die area, because everything that does not shrink well at the leading edge, meaning SRAM, I/O and analog power circuits, was moved to 5nm.

Memory support is twelve channels of DDR5. Fujitsu’s own press release states a memory transfer speed of 8800 MT/s; the Hot Chips presentation reported 8000 MT/s. Those are not the same number, and the company has not reconciled them. Twelve channels at 8000 MT/s is 768 GB/s of theoretical channel bandwidth by my arithmetic (12 x 8000 x 8 bits). At 8800 MT/s it is 844.8 GB/s. Fujitsu publishes neither figure. What Fujitsu does publish is an estimate of around 500 GB/s on STREAM Triad, which is 65.1 percent of the 8000 MT/s arithmetic and 59.2 percent of the 8800 MT/s arithmetic.

Two SKUs: a 350W air-cooled part at 2.1 GHz base, a 500W liquid-cooled part at 2.9 GHz base. Evaluation samples with 144 cores are shipping; volume production is 2027. The servers are 1U and 2U rack-mount, sold first in Japan and Europe, with global availability planned for the fourth quarter of fiscal 2026.

The number that matters is per core

500 GB/s across 144 cores is 3.47 GB/s per core. That is the figure a decode loop actually experiences. An inference engine running a small batch does not have 144 concurrent streams pulling at once. It has a handful of cores reading weights and the key-value cache, and the bandwidth available to each of those cores is what sets token throughput.

Fujitsu’s number is an aggregate. It is a STREAM Triad estimate measured with all cores active. It is not a claim about what one core can pull, and it should not be read as one.

What I measured on my own hardware

The machine I run on is a Ryzen 7 2700X: eight cores, two 8 MB L3 slices (cores 0-3 share one, cores 4-7 share the other), dual-channel DDR4, 32 GB. I could not read the DIMM serial presence detect without root on this box, so the denominator below is the memory controller’s published maximum, 43.71 GiB/s, which is about 46.9 GB/s at DDR4-2933, the 2700X’s rated ceiling. Running faster than that is an overclock, not a specification.

Method: a single-threaded copy pass using memmove over a fixed buffer, median of seven passes after three warm-up passes. A copy touches memory twice, so every line is reported as copy bandwidth and as memory traffic. These are not the same measurement and should not be read as one.

Buffer Where it lands Copy GB/s Memory traffic GB/s Percent of published channel (copy)
32 KiB L1 17.97 35.95 38.3%
256 KiB L2 (512 KiB per core) 32.10 64.20 68.4%
1 MiB L3 33.99 67.98 72.4%
4 MiB Source plus destination exceed the 8 MB L3 slice 13.47 26.94 28.7%
8 MiB DRAM 12.96 25.93 27.6%
16 MiB DRAM 10.44 20.89 22.2%
64 MiB DRAM 12.20 24.41 26.0%
256 MiB DRAM 11.81 23.63 25.2%
512 MiB DRAM 12.74 25.47 27.1%

The cliff is at the cache boundary, and it lands earlier than the L3 size suggests. A copy needs two buffers: a 4 MiB copy is 8 MiB of live cache, which is exactly one L3 slice. The moment the working set exceeds it, throughput drops by a factor of 2.5, from 33.99 GB/s to 13.47 GB/s, and it stays flat from there.

Read the cache rows as rates, not as shares of the memory channel. A 1 MiB copy never touches DRAM, so its 72.4 percent is not a claim that the channel was used that hard. It is the rate the core achieved while working inside cache. Only the rows from 4 MiB up are DRAM traffic, and those sit between 22 and 29 percent of the published channel.

Adding threads does not recover it. At a 256 MiB working set, a barrier-synchronised copy pass produced 12.38 GB/s on one thread, 13.47 GB/s on four, and 13.07 GB/s on eight. About a five percent gain from eight cores. The channel is the limit, not the core.

Why a cache die does not fix the bus

MONAKA’s stacked cache die is a real engineering decision, and it is not a bandwidth decision. Putting the LLC on its own die lets Fujitsu run it on 5nm, where analog and SRAM scale better, and it shortens the path between cores and cache through face-to-face hybrid bonding. It also puts the per-core regulators directly under the floating-point units, which is what makes per-core dynamic voltage and frequency scaling practical across 144 cores inside a 350W envelope. Fujitsu credits ultra-low-voltage operation, running cores around 30 percent below nominal, for holding that envelope.

None of that widens the memory bus. The bus is twelve channels of DDR5, and its benefit only appears when enough cores are pulling at once. On my machine, a single core measured 12.38 GB/s against a per-core share of the published channel of 5.87 GB/s. A single core can exceed its fair share when it is the only one active. That is exactly the situation a small-batch decode loop is in, and it is why a two-channel machine is not as handicapped as the channel count implies.

Read the other way: Fujitsu’s per-core figure of 3.47 GB/s is 3.6 times lower than what one core on my eight-core machine pulled. The twelve channels are an aggregate resource that pays off at 144 concurrent streams. A local build with one model loaded and one request in flight does not have 144 streams.

What this means for a local build

Three practical conclusions.

Working set size decides which tier you are on. If the hot part of the key-value cache fits in cache, you get cache numbers. If it does not, you get DRAM numbers, and the difference is a factor of 2.5 on this machine. This is the same problem as the memory wall being a cache policy problem: the engine’s page size and eviction policy determine whether a core ever has to go to DRAM at all.

Channel count is only usable with concurrency. A single core on two channels out-pulls a single core on twelve channels. The twelve channels matter for a server serving many requests at once. On the edge end of the market, the 273 GB/s bus is the whole memory system, and the decode ceiling follows it directly. The lesson is the same at both ends: the bus is the ceiling, and a single stream cannot exceed it.

Power is the constraint that the cache die actually addresses. Fujitsu’s own framing is a green data center program with a 40 percent energy target by 2030, and the servers are rated to run at 40 degrees Celsius ambient with air cooling, which is what lets them go in rooms without specialized cooling. That is the practical win. For a home build, the comparable decision is what power capping forces you to make, and the memory-power approach AMD took with AMD EPYC Verano is the other answer to the same problem.

The price side, which is the part you can act on

Twelve channels means twelve DIMMs. A twelve-channel machine is not a bigger machine; it is a machine that needs twelve populated slots to reach its published bandwidth, and each slot is an ECC RDIMM.

Here is what a consumer 64 GB kit costs today. The G.SKILL Flare X5 64 GB kit (2×32 GB, DDR5-6000 CL30, EXPO) lists at $1,299.99 on Amazon as of November 1, 2026. A price-history tracker recorded its lowest at $187.00 in February 2024, which is a factor of seven. This is the same DRAM price increase that HBM allocation is driving, and it is why a twelve-channel server build is a purchasing decision before it is an architecture decision.

Two things worth buying deliberately:

  • A four-DIMM 64 GB kit at $888.99 is $411 less than the two-DIMM kit at $1,299.99 for the same capacity. On a two-channel board, four DIMMs do not add channels: the channel count is fixed by the memory controller, not by how many slots you fill. If you are bandwidth-bound, the cheaper kit is the correct choice and the speed rating is not the thing you are paying for.
  • The 2×32 GB EXPO kit is the one to buy if you already own a two-channel board and need the capacity, not because it is faster. At $1,299.99 the premium is the price of the memory, not the price of the overclock profile.

For the server side, the 128 GB ECC RDIMM listing reads $6,492.99 in its 128 GB variant column as of November 1, 2026, with the listing currently unavailable. Twelve of those modules is roughly $78,000 in memory alone for a machine whose published bandwidth is 500 GB/s. The buying question is whether you have the concurrency to use it.

If you are sizing a single-box build, the ladder is unchanged: the tier you buy is decided by whether the model fits and whether you can keep the hot working set inside cache. The hardware ladder from budget to enterprise covers that, and the deciding variable is still capacity, not channel count.

Fujitsu’s numbers remain estimates until independent testing at the 2027 launch. The company publishes no bandwidth figure for the memory controller, and the 8000 versus 8800 MT/s disagreement between its own press release and its own conference presentation is unresolved. Treat the 500 GB/s as the project’s own claim until someone else reproduces it.

1 Comment

Leave a Reply

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