inference/research
verifiedmedium · 0.65updated 2026-10-03

Moore Threads MTLink (MTT S4000 card-to-card fabric)

120.0 GB/s

Fields

id
interconnect/moore-threads-mtlink
Kind
other
Version
MTLink, MTT S4000 generation (MTLink Bridge for 2/4/8-card topologies)
Bandwidth (GB/s, per link, unidirectional)
120
Bandwidth basis
PER CARD, PER DIRECTION, AND THE 2x IS A GENUINELY CONTESTED READING - flagged here because the vendor does not state a direction for this figure. The Moore Threads S4000 page (cn-mthreads-s4000) publishes '240 GB/s I/O bandwidth' and, separately, names 'MTLink interconnect' as the card-to-card mechanism, with the accelerator record (accelerators/moore-threads-s4000) noting that 'the 240 GB/s card-to-card figure equals the stated I/O bandwidth'. THE AMBIGUITY: Moore Threads' own consumer page for the MTT S80 uses explicitly bidirectional language for its PCIe figure - 'the highest BIDIRECTIONAL 128GB/s transfer bandwidth' on a PCIe Gen 5 x16 interface, which is exactly 2 x the 64 GB/s per-direction rate that PCI-SIG's convention gives - so this vendor DOES use 'bidirectional' deliberately and knows the distinction. On that precedent, and because 240 GB/s equals the S4000's total stated I/O bandwidth for a PCIe 5.0 x16 host link plus the MTLink card-to-card path, the recorded value is the HALF, 240 / 2 = 120 GB/s per direction per card, on the reading that 240 is a bidirectional aggregate. THE COMPETING READING, retained rather than discarded: if 240 is already per-direction, the per-card fabric rate is 240 GB/s and B* in docs/11-interconnect-decision.md falls to ~3,860 instead of ~7,720 on the S4000's 768 GB/s of memory - still far above any serving batch either way, so the decision conclusion is unchanged by which reading is right. That is the honest way to record it: the direction convention does not change the recommendation, so the record is verified on the bandwidth being firmly in the low hundreds, with the 2x specifically flagged as convention-dependent. STATUS IS VERIFIED RATHER THAN CONTESTED BECAUSE THE TWO READINGS DISAGREE BY A CONSTANT FACTOR OF 2 ON A FIGURE WHOSE ORDER OF MAGNITUDE IS NOT IN DOUBT AND WHOSE DECISION-RELEVANCE IS INSENSITIVE TO IT - the vendor figure itself is first-party and uncontradicted.
Topology
Physical card-to-card links via an MTLink Bridge accessory. The S4000 page states the MTLink Bridge supports 2-, 4- and 8-card topologies, which is a small direct-attached domain comparable to the 8-way meshes AMD's Infinity Fabric provides and smaller than what NVSwitch reaches. No switch part is named in the retrieved sources.
Scale-up
yes
Scale-out
no
Switching
None published. The domain is the 2/4/8-card MTLink Bridge configuration, built from direct links.
notes
Added 2026-10-03 to resolve a DANGLING REFERENCE: both moore-threads-s4000 and moore-threads-s5000 list 'moore-threads-mtlink' in their interconnect arrays and no such record existed. The prior audit pass left a documented negative that this fabric had no sourceable record; the vendor page already held in this repo (cn-mthreads-s4000) does in fact name MTLink and give the 240 GB/s figure, so the negative is now closed. WHAT THE VENDOR PUBLISHES (S4000 page, primary): 48 GB memory, 768 GB/s memory bandwidth, 240 GB/s I/O bandwidth, PCIe 5.0 x16, MTLink interconnect, 450 W max power, 266 x 112 x 39 mm, and supported precisions FP64/FP32/TF32/FP16/BF16/INT8 - but NO throughput figure for any precision. MTLink Bridge supports 2/4/8-card topologies. THE SCALE-UP-TO-MEMORY RATIO IS THE ACTUALLY INTERESTING NUMBER, and it is stated PER DIRECTION to match the convention of every other fabric in this file: the recorded 120 GB/s per direction against 768 GB/s of local memory is a ratio of 0.156, versus H100's 450/3350 = 0.134 and A100's 300/2039 = 0.147 and MI355X's 537.6/8000 = 0.067. CORRECTION, because the version of this note written earlier in the audit used the 240 GB/s published figure directly and got 0.31 - which is exactly the aggregate-as-per-direction error this record is otherwise careful about, applied to its own ratio claim. At the corrected 0.156 this part is roughly level with A100 and H100 rather than '2.3x better than H100', so the honest summary is that MTLink's ratio is ordinary for its generation, not exceptional. Its TP crossover is B* ~3,860 on 7B geometry, still far above any serving batch, so the decision conclusion is unchanged - but the part is not the standout the earlier note implied. TWO THINGS A READER SHOULD NOT INFER. First, the S4000's own page declines to publish a single TFLOPS number for any precision, which is unusual and means any compute claim about this part comes from elsewhere; the sibling S5000 page does publish tok/s under an explicit methodology, so this is a per-part choice rather than a company that cannot be specific. Second, MTLink's scale-up domain tops out at 8 cards, so this is a single-node scale-up story with no path to rack scale comparable to NVSwitch or the Atlas 950 SuperPoD - the honest framing is 'healthy ratio, small domain'. KIND: filed as 'other' because MTLink is a proprietary vendor interconnect with no entry in the schema's enum and no analogue in the named families.
way
each

Notes

Added 2026-10-03 to resolve a DANGLING REFERENCE: both moore-threads-s4000 and moore-threads-s5000 list 'moore-threads-mtlink' in their interconnect arrays and no such record existed. The prior audit pass left a documented negative that this fabric had no sourceable record; the vendor page already held in this repo (cn-mthreads-s4000) does in fact name MTLink and give the 240 GB/s figure, so the negative is now closed. WHAT THE VENDOR PUBLISHES (S4000 page, primary): 48 GB memory, 768 GB/s memory bandwidth, 240 GB/s I/O bandwidth, PCIe 5.0 x16, MTLink interconnect, 450 W max power, 266 x 112 x 39 mm, and supported precisions FP64/FP32/TF32/FP16/BF16/INT8 - but NO throughput figure for any precision. MTLink Bridge supports 2/4/8-card topologies. THE SCALE-UP-TO-MEMORY RATIO IS THE ACTUALLY INTERESTING NUMBER, and it is stated PER DIRECTION to match the convention of every other fabric in this file: the recorded 120 GB/s per direction against 768 GB/s of local memory is a ratio of 0.156, versus H100's 450/3350 = 0.134 and A100's 300/2039 = 0.147 and MI355X's 537.6/8000 = 0.067. CORRECTION, because the version of this note written earlier in the audit used the 240 GB/s published figure directly and got 0.31 - which is exactly the aggregate-as-per-direction error this record is otherwise careful about, applied to its own ratio claim. At the corrected 0.156 this part is roughly level with A100 and H100 rather than '2.3x better than H100', so the honest summary is that MTLink's ratio is ordinary for its generation, not exceptional. Its TP crossover is B* ~3,860 on 7B geometry, still far above any serving batch, so the decision conclusion is unchanged - but the part is not the standout the earlier note implied. TWO THINGS A READER SHOULD NOT INFER. First, the S4000's own page declines to publish a single TFLOPS number for any precision, which is unusual and means any compute claim about this part comes from elsewhere; the sibling S5000 page does publish tok/s under an explicit methodology, so this is a per-part choice rather than a company that cannot be specific. Second, MTLink's scale-up domain tops out at 8 cards, so this is a single-node scale-up story with no path to rack scale comparable to NVSwitch or the Atlas 950 SuperPoD - the honest framing is 'healthy ratio, small domain'. KIND: filed as 'other' because MTLink is a proprietary vendor interconnect with no entry in the schema's enum and no analogue in the named families.

Cross-references

This record references no other records — it is a leaf. It is cited by 3.