NVIDIA · accelerator
- Accelerators
- 1 × H100 SXM
- Architecture
- Hopper
- Memory
- 80 GB
- Bandwidth
- 3.4 TB/s
- Accelerator interconnect
- 0.9 TB/s
- Power
- 1.275 kW: 0.7 kW for the accelerator and a 0.575 kW allowance for the host
- Acquisition
- not published
- Deployment multiplier
- ×1
Unpriced accelerator; use an explicit scenario acquisition price.
Official specifications.
Power: 0.7 kW is the accelerator alone: NVIDIA rates the H100 SXM at "Up to 700W (configurable)". An SXM module runs inside an eight-GPU server, so the 0.575 kW host allowance is one eighth of the host share of NVIDIA's own eight-GPU Hopper system. NVIDIA's DGX H200 datasheet gives 10.2 kW maximum system power for its standard configuration, and its eight H200s are rated "Up to 700W" each on NVIDIA's H200 page: 5.6 kW of accelerators, which leaves 4.6 kW for the CPUs, memory, NVSwitch chips, networking, fans and power-supply losses.
NVIDIA · node
- Accelerators
- 8 × H200
- Architecture
- Hopper
- Memory
- 1,128 GB
- Bandwidth
- 38.4 TB/s
- Power
- 10.2 kW: the whole system at full load
- Acquisition
- $370,000
- Reported range
- $320,000 – $420,000
- Deployment multiplier
- ×1.2
Independent typical 8-GPU estimate.
Memory/bandwidth use official H200 values and the 8-GPU price range is independent research. REGRADED 2026-09-25 from assumption: the assumption was the power figure, which now has its own record citing NVIDIA's DGX H200 datasheet, so what this record still rests on is a price estimate.
Power: NVIDIA's DGX H200 datasheet: "System power usage 10.2kW max" for the standard configuration, host included; its Custom Thermal Solution variant is specified up to 14.3 kW. This entry prices a generic eight-GPU H200 server, so the figure is NVIDIA's reference system rather than any one OEM build. It carried the same 10.2 kW before 2026-09-25, recorded as a planning assumption with no source.
NVIDIA · node
- Accelerators
- 8 × B200
- Architecture
- Blackwell sm_100/sm_103
- Memory
- 1,440 GB
- Bandwidth
- 64 TB/s
- Accelerator interconnect
- 14.4 TB/s
- Power
- 14.3 kW: the whole system at full load
- Acquisition
- not published
- Deployment multiplier
- ×1.2
Full-node acquisition price not published; planning assumption required.
Official node specifications; price intentionally missing.
Power: The DGX B200 User Guide's power specification gives "14.3 kW max." for the system, drawn through six 3.3 kW supplies in 5+1 redundancy, with each GPU capped at 1,000 W. The supplies' combined rating is capacity, not draw.
NVIDIA · node
- Accelerators
- 8 × B300
- Architecture
- Blackwell sm_100/sm_103
- Memory
- 2,304 GB
- Bandwidth
- 64 TB/s
- Accelerator interconnect
- 14.4 TB/s
- Power
- 14.5 kW: the whole system at full load
- Acquisition
- $350,000
- Reported range
- $300,000 – $500,000
- Deployment multiplier
- ×1.2
No NVIDIA MSRP exists. Independent 2026 reports cluster at $300k-350k baseline with resellers quoting up to $500k; $350k is the recorded planning point.
[CORRECTED at the 2026-08-01 refresh. This entry previously carried $550,000, a figure that re-verification could not trace to any source; current independent reporting puts an 8-GPU B300 node at $300k-350k baseline and up to $400-500k from resellers, so the point value is now $350k with the spread recorded.] Memory (2,304 GB), 64 TB/s aggregate HBM bandwidth and 14.5 kW are all exact matches to NVIDIA's own DGX B300 User Guide. Note that the 14.4 TB/s figure circulating in secondary write-ups is the NVLink switch fabric, not memory bandwidth. Deployment multiplier is an editable planning assumption.
Power: The DGX B300 User Guide lists "Power consumption: 14.5 kW" for the system and specifies its AC and DC inputs for "15 kW system max."; the twelve 3.2 kW supplies are capacity, not draw. Read in the guide's HTML edition on 2026-09-25.
NVIDIA · rack
- Accelerators
- 72 × GB200
- Architecture
- Blackwell sm_100/sm_103
- Memory
- 13,400 GB
- Bandwidth
- 576 TB/s
- Accelerator interconnect
- 130 TB/s
- Power
- 120 kW: the whole system at full load
- Acquisition
- not published
- Deployment multiplier
- ×1.25
Rack price not published; planning assumption required.
[CORRECTED at the 2026-08-01 refresh: power is now NVIDIA's own approximately 120 kW from the DGX GB200 NVL72 User Guide, replacing the 132 kW previously recorded.] CONFLICT preserved: infrastructure partners co-designing reference racks with NVIDIA (Vertiv, Schneider Electric, HPE) provision for up to 132 kW nominal with far higher transient peaks, so 132 kW is the right number for facility sizing and 120 kW for the rack itself. Memory and fabric are official; the 13.4 TB rack aggregate is NVIDIA's own published figure and deliberately is not 72 x a per-GPU value.
Power: The DGX GB200 User Guide: "The rack power consumption is approximately 120kW." That is the whole rack: compute trays, NVLink switch trays and the eight power shelves feeding them. Partners provision up to 132 kW, which is facility capacity rather than draw.
NVIDIA · rack
- Accelerators
- 72 × GB300
- Architecture
- Blackwell sm_100/sm_103
- Memory
- 20,736 GB
- Bandwidth
- 576 TB/s
- Accelerator interconnect
- 130 TB/s
- Power
- 135 kW: the whole system at full load
- Acquisition
- $3,500,000
- Reported range
- $3,000,000 – $4,000,000
- Deployment multiplier
- ×1.25
No MSRP. Press and analyst reporting spans $3.0-4.0M per rack; $3.5M is the recorded midpoint.
[CORRECTED at the 2026-08-01 refresh. The previous $3.0M simply reused the GB200 estimate, while reporting consistently puts GB300 at a premium over its predecessor ($3.0-4.0M, with one analyst note derived from a large customer order suggesting $3.7-4.0M); the midpoint is now recorded with the spread. Power moved from 132 kW to 135 kW: NVIDIA publishes no rack figure for GB300 at all, OEM reference documentation from HPE and Lenovo converges on 132-135 kW nominal, and other trackers report 140-142 kW — treat 132-142 kW as the honest range.] Memory is derived as 72 x 288 GB; NVIDIA's own page rounds it to 20 TB.
Power: NVIDIA publishes no rack power for GB300 NVL72. OEM reference documentation from HPE and Lenovo converges on 132-135 kW nominal for the whole rack and other trackers report 140-142 kW; 135 kW is a point in that range, not a vendor figure.
Vera Rubin NVL72 (VR200)
preview NVIDIA · rack
- Accelerators
- 72 × Rubin (VR200)
- Architecture
- Rubin
- Memory
- 20,736 GB
- Bandwidth
- 1,580 TB/s
- Accelerator interconnect
- 260 TB/s
- Power
- 190 kW: the whole system at full load
- Acquisition
- not published
- Deployment multiplier
- ×1.25
No vendor price published. An analyst estimate of $3.5-4.0M per rack circulates but is not NVIDIA's figure; supply your own quote.
NAMING: NVIDIA announced this rack as "NVL144" at GTC 2025 counting compute dies, then as "NVL72" at CES 2026 counting GPU packages. They are the same rack, not two products. Memory is 72 × 288 GB of HBM4, which NVIDIA's datasheet and product page both give as 20.7 TB. Bandwidth is the 1,580 TB/s the datasheet prints for the rack; 72 × its 22 TB/s per GPU would be 1,584 TB/s, and the printed rack figure is the one stored. The 260 TB/s NVLink 6 fabric is the datasheet's and NVIDIA's CES release's. CONFLICT preserved: NVIDIA's product page, read 2026-09-26, gives 19.2 TB/s per GPU and 1,400 TB/s for the rack, and 3 TB/s of NVLink per GPU and 216 TB/s for the rack. The same page's "100 MW AI Factory Specifications" give "12 PB HBM4 | 800 PB/s" for 40K GPUs, 20 TB/s per GPU, which fits the page's own 19.2 TB/s, 768 PB/s across 40,000 GPUs, and not the datasheet's 22 TB/s, 880 PB/s; it is the same page as the 19.2, so it adds no independent document to that side. The datasheet that page links calls its specifications "Preliminary information" whose values are "up to and subject to change", and neither document says which describes the racks now shipping. The stored figures are the datasheet's, which every earlier report also gave, and confidence is low because NVIDIA's own two documents disagree. The 190 kW power draw is NOT vendor-published — it is the low end of a supply-chain analyst range of roughly 190 kW (Max-Q) to 230 kW (Max-P); the Power draw field under Operations replaces it with a quote. SHIPPING: at CES on 2026-01-05 NVIDIA said "Rubin-based products will be available from partners the second half of 2026". On 2026-07-21 it said racks were running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius, and an update on the same post, headed "Tuesday, July 24, 1:30 p.m. PT", says Nebius "got early access". That header's weekday and date disagree: 2026-07-24 is a Friday, and the Tuesdays either side are 2026-07-21, the post's own date, and 2026-07-28, the last change its metadata records. Its investor presentation of 2026-08-26 says NVIDIA "Commenced production shipments of Vera Rubin in early August", which is 2026-08-01 to 2026-08-10 read as the first third of the month, and on 2026-09-15 it said "Vera Rubin is in full production and scaling across the ecosystem". It entered the rack in MLPerf Inference v6.1, published 2026-09-16, as a preview system rather than an available one; MLCommons calls a system available only once it is priced, has shipped to a third party and can be bought by further third parties without early-access approval. So the rack ships, NVIDIA has not said it is generally available, and it has published no price: this catalog's preview status, released but not generally available. [CORRECTED 2026-09-26: memory bandwidth was stored as 1,584 TB/s, 72 × the datasheet's 22 TB/s per GPU, a rack figure NVIDIA does not print. The datasheet prints 1,580 TB/s for the rack, and where a vendor prints a rack figure this catalog stores it, as it does AMD's 1,670 TB/s for Helios.] [CORRECTED 2026-09-26: the entry was announced, which this catalog defines as not yet shipping, and stored 2026-07-01 as the day it became available. That is the first day of the half NVIDIA named at CES, not a day NVIDIA gave. The note said "Declared in full production at CES 2026 with volume availability in the second half of 2026", citing only ServeTheHome and SemiAnalysis.] [UPGRADED 2026-09-26: estimated to official-claim, at low confidence still. Memory, bandwidth and fabric are now read from NVIDIA's datasheet, product page and CES release rather than from two outlets' reports of them, so the reason for the downgrade below no longer holds.] [DOWNGRADED 2026-08-29: this record was graded official-claim at medium confidence while citing no primary source at all — both sourceIds are independent-research, ServeTheHome and SemiAnalysis, so nothing here was read from NVIDIA. The gb300-nvl72 record carries the identical caveat about an unpublished rack power figure, cites one primary source more than this one, and is graded estimated at low confidence; the entry with the weaker evidence held the stronger grade. Nothing about the figures changed — 288 GB HBM4 x 72 and the NVLink 6 fabric are still what NVIDIA announced at CES 2026 and what those two outlets reported — only the claim this catalog makes about how it knows them.]
Power: NVIDIA has published no rack power for Vera Rubin NVL72: its datasheet, its product page and its MLPerf Inference v6.1 system description, read on 2026-09-26, state none. 190 kW is the low end of a supply-chain analyst range of roughly 190 kW (Max-Q) to 230 kW (Max-P) for the whole rack. The system description does rate this rack's GPUs, "2300W TGP Per GPU": 165.6 kW for the 72 GPUs alone, before the 36 Vera CPUs, the NVLink switch trays and the networking, so 190 kW is not below the GPUs' own rating. The product page gives a cross-check: its "100 MW AI Factory Specifications" put "40K GPUs" in 100 MW, footnoted "Specifications based on an at-scale AI factory using NVIDIA DSX with MaxLPS", which is 2.5 kW per GPU for everything those 100 MW cover, about 180 kW per 72 GPUs and 5% under the stored figure. The page does not say whether the 100 MW is the IT load or the whole facility's power, cooling included, and does not define MaxLPS, so the figure cannot replace the stored one: the calculator applies PUE separately, and a figure that may already carry a facility's overhead is not a rack's own draw. The Power draw field under Operations replaces it with a quote.
Instinct MI455X Helios rack
announced AMD · rack
- Accelerators
- 72 × MI455X
- Architecture
- CDNA 5
- Memory
- 31,104 GB
- Bandwidth
- 1,670 TB/s
- Accelerator interconnect
- 260 TB/s
- Power
- 190 kW: the whole system at full load
- Acquisition
- not published
- Deployment multiplier
- ×1.25
- Expected
- 2026-08-04 to 2026-09-26
No vendor price published. An analyst estimate of $5.0-5.5M per rack circulates but is not AMD's figure; supply your own quote.
NAMING: AMD's own press releases call this generation both "MI450 Series" (the OpenAI and Anthropic partnership announcements) and "MI400 Series" with the MI455X as flagship (the July 2026 launch). Same hardware. 432 GB HBM4 per accelerator across 72 accelerators is a vendor figure, as is the 260 TB/s UALink-over-Ethernet fabric; AMD's MI455X and Helios pages state them. BANDWIDTH: the stored 1,670 TB/s is AMD's own rack figure. Its MI400 Series page says Helios offers "1.67 PB/s of memory bandwidth", and a tile on that page and AMD's Helios launch release of 2026-07-23 round it to 1.7 PB/s. The 23.3 TB/s per GPU that the MI455X page, the MI400 Series page and the Helios page's specification table give agrees with it: 72 × 23.3 TB/s is 1,677.6 TB/s, within 0.5%. CONFLICT preserved on bandwidth: the compute-tray section of the same Helios page gives the MI455X "19.6 TB/s memory bandwidth per GPU", which would make the rack 1,411.2 TB/s. That is the figure AMD gave when it previewed the MI400 Series on 2025-06-12, and it cannot produce the launch release's claim of "6% more HBM bandwidth" than the leading competitive solution, which its footnote names as NVIDIA's Vera Rubin NVL72: AMD's pages set 23.3 TB/s against 22.0 TB/s for Vera Rubin, 5.9% more, where 19.6 TB/s would be 11% less. So the entry carries 1.67 PB/s, graded as AMD's claim at medium confidence: apart from that one section and the 2025 preview, every AMD statement read here agrees with it. AMD's product pages show no date; their metadata gives 2026-09-01 as their last update, and they were read on 2026-09-26. CONFLICT preserved on power: coverage of the same launch event cites both roughly 140 kW and 225-245 kW of bus-bar capacity under load; AMD has published no single authoritative rack figure, so 190 kW is recorded as a midpoint planning assumption, and the Power draw field under Operations replaces it with whatever a vendor quote states. AMD does not sell the rack itself: its Form 10-Q for the quarter to 2026-06-27 says it supplies products used in a Helios system and does not manufacture or sell the completed rack, which customers' manufacturing partners build. When AMD said it would ship is its release window. [CORRECTED 2026-09-26: this said the rack was "declared in full production 2026-07-23 with shipments from Q3 2026", a paraphrase that cited no AMD source, and the entry stored 2026-09-01 as the day it became available, a day no AMD page names.]
Power: AMD has published no rack figure. Coverage of the same launch cites roughly 140 kW and 225-245 kW of bus-bar capacity under load; 190 kW is a midpoint this repository chose between them, for the whole rack. The Power draw field under Operations replaces it with whatever a vendor quote states.
Release: on 2026-08-04
AMD said “Helios is now in production, with initial shipments on track to begin later this quarter and ramp through the fourth quarter and into 2027 to meet very strong customer demand.”
Lisa Su said it on AMD's second-quarter earnings call and closed the same remarks with “initial Helios shipments set to begin this quarter”. That quarter is AMD's fiscal third. AMD's fiscal year ends on the last Saturday in December, 2026-12-26 this year, 52 weeks, and its Form 10-Q gives the second quarter as the 13 weeks to 2026-06-27, so the third quarter's 13 weeks end on 2026-09-26, four days before the calendar quarter. The window runs from the day she said it to that date. Two wider statements agree with it: the 10-Q expects Helios “to start shipping in the second half of fiscal year 2026”, and AMD's release with Microsoft of 2026-07-20 says “AMD will begin shipping Helios to customers, including Microsoft, in the second half of 2026.” AMD has announced no Helios shipment since: neither its newsroom, read to its item of 2026-09-24, nor its investor-relations releases, the latest of 2026-08-31, reports one.
Instinct MI300X 8-GPU platform
AMD · node
- Accelerators
- 8 × MI300X
- Architecture
- CDNA 3
- Memory
- 1,536 GB
- Bandwidth
- 42.4 TB/s
- Power
- 10.6 kW: 6 kW for the 8 accelerators and a 4.6 kW allowance for the host
- Acquisition
- not published
- Deployment multiplier
- ×1.2
Acquisition price not published.
Memory and bandwidth are official. REGRADED 2026-09-25 from assumption: that grade was the 8.5 kW full-node planning value, and power now carries its own record, the data sheet's 6 kW for eight GPUs plus a cited host allowance.
Power: 6 kW is the accelerators alone: eight MI300X at the 750 W maximum board power of AMD's platform data sheet. The 4.6 kW host allowance comes from the eight-GPU system whose accelerators are closest to 750 W. NVIDIA's DGX H200 datasheet gives 10.2 kW maximum system power for its standard configuration, and its eight H200s are rated "Up to 700W" each on NVIDIA's H200 page: 5.6 kW of accelerators, which leaves 4.6 kW for the CPUs, memory, NVSwitch chips, networking, fans and power-supply losses. Four NVSwitch chips are part of that share and this platform has none, so it errs slightly high here. It replaces an 8.5 kW full-node planning value whose 2.5 kW for the host had no source.
Instinct MI325X 8-GPU platform
AMD · node
- Accelerators
- 8 × MI325X
- Architecture
- CDNA 3
- Memory
- 2,048 GB
- Bandwidth
- 48 TB/s
- Power
- 14.3 kW: 8 kW for the 8 accelerators and a 6.3 kW allowance for the host
- Acquisition
- not published
- Deployment multiplier
- ×1.2
Acquisition price not published.
Memory and bandwidth are official. REGRADED 2026-09-25 from assumption: that grade was the 10.5 kW full-node planning value, and power now carries its own record, the data sheet's 8 kW for eight GPUs plus a cited host allowance.
Power: 8 kW is the accelerators alone: eight MI325X at the 1,000 W maximum of AMD's platform data sheet. The 6.3 kW host allowance comes from the eight-GPU system whose accelerators draw the same 1,000 W. NVIDIA's DGX B200 User Guide gives 14.3 kW maximum system power and caps each of its eight GPUs at 1,000 W: 8.0 kW of accelerators, which leaves 6.3 kW for everything else in the node. It replaces a 10.5 kW full-node planning value whose 2.5 kW for the host had no source.
Instinct MI355X 8-GPU platform
AMD · node
- Accelerators
- 8 × MI355X
- Architecture
- CDNA 4
- Memory
- 2,304 GB
- Bandwidth
- 64 TB/s
- Power
- 17.5 kW: 11.2 kW for the 8 accelerators and a 6.3 kW allowance for the host
- Acquisition
- $325,000
- Deployment multiplier
- ×1.25
Node acquisition planning assumption; not MSRP.
Official GPU specifications; the node price is a planning assumption. Power carries its own record since 2026-09-25, AMD's 1,400 W per GPU plus a cited host allowance, in place of a 14 kW planning value.
Power: 11.2 kW is the accelerators alone: eight MI355X at the 1,400 W typical board power on AMD's product page. The 6.3 kW host allowance comes from the highest-power eight-GPU system for which a vendor publishes both figures. NVIDIA's DGX B200 User Guide gives 14.3 kW maximum system power and caps each of its eight GPUs at 1,000 W: 8.0 kW of accelerators, which leaves 6.3 kW for everything else in the node. The MI355X is the liquid-cooled part, and an air-cooled host spends part of that share on fans, so the allowance likely errs high here. It replaces a 14 kW planning value whose 2.8 kW for the host had no source.
Intel · node
- Accelerators
- 8 × Gaudi 3
- Architecture
- Gaudi 3
- Memory
- 1,024 GB
- Bandwidth
- 29.6 TB/s
- Accelerator interconnect
- 4.2 TB/s
- Power
- 13.5 kW: 7.2 kW for the 8 accelerators and a 6.3 kW allowance for the host
- Acquisition
- not published
- Deployment multiplier
- ×1.2
- Available from
- 2025-05-19
US$125,000 list for the 8-OAM baseboard incl. networking (Computex 2024); host system excluded, full-node price not published.
128 GB HBM2e and 900 W per accelerator are official. REGRADED 2026-09-25 from assumption: that grade was a 9.5 kW full-node planning value, and power now carries its own record, the 7.2 kW the eight accelerators draw plus a cited host allowance.
Power: 7.2 kW is the accelerators alone: eight Gaudi 3 at 900 W each. The 6.3 kW host allowance comes from the eight-GPU system whose accelerators are closest to 900 W. NVIDIA's DGX B200 User Guide gives 14.3 kW maximum system power and caps each of its eight GPUs at 1,000 W: 8.0 kW of accelerators, which leaves 6.3 kW for everything else in the node. It replaces a 9.5 kW full-node planning value whose 2.3 kW for the host had no source.
NVIDIA · accelerator
- Accelerators
- 1 × RTX PRO 6000 Blackwell Workstation Edition (GB202)
- Architecture
- Blackwell sm_120/sm_121
- Memory
- 96 GB
- Bandwidth
- 1.8 TB/s
- Power
- 0.68 kW: 0.6 kW for the accelerator and a 0.08 kW allowance for the host
- Acquisition
- $16,000
- Deployment multiplier
- ×1
NVIDIA's own price for this card on its US store, $16,000.00, read on 2026-09-25. Pre-orders opened around $8,000–8,565 in March 2025; press reporting attributes the doubling to a GDDR7 shortage.
96 GB GDDR7 ECC, 1,792 GB/s and 600 W are NVIDIA's own specifications, and so is the price, $16,000.00 on NVIDIA's own store. Until 2026-09-25 this entry said NVIDIA published no price and carried a range, $8,000 to $16,000, from the March 2025 pre-orders to press reporting of the September 2026 street price. Beside the vendor's one price, the low end describes what the card cost eighteen months earlier rather than a price on offer, so no range is carried. Carries no NVLink: the datasheet lists PCIe 5.0 x16 as the whole of its system interface, and the Server Edition's product brief states 'NVIDIA NVLink: Not supported' outright for the same silicon. NVLink left this product line two generations ago — the RTX A6000 had a bridge for two cards at 112 GB/s, the RTX 6000 Ada that replaced it did not — so a second card in the same box talks over the host bus at 128 GB/s, not over a fabric.
Power: 0.6 kW is the card alone, its 600 W maximum on NVIDIA's datasheet. The 0.08 kW host allowance is a measured desktop overhead: a study that metered an RTX 5080 workstation at the wall beside the GPU's own reported power found the difference at 68-80 W across its ordinary ML and HPC workloads (arXiv 2606.19262), and the allowance takes the top of that range.
RTX PRO 6000 Blackwell Max-Q
NVIDIA · accelerator
- Accelerators
- 1 × RTX PRO 6000 Blackwell Max-Q Workstation Edition (GB202)
- Architecture
- Blackwell sm_120/sm_121
- Memory
- 96 GB
- Bandwidth
- 1.8 TB/s
- Power
- 0.38 kW: 0.3 kW for the accelerator and a 0.08 kW allowance for the host
- Acquisition
- $16,000
- Deployment multiplier
- ×1
NVIDIA's own price for this card on its US store, $16,000.00, read on 2026-09-25, the same as the full-power Workstation Edition's. Until then this entry carried $7,673 to $8,900, retailer and marketplace quotes gathered by search rather than read off any store's own page; NVIDIA's price is about twice that, so the quotes are not kept as a range.
Same 96 GB and the same 1,792 GB/s as the full-power Workstation Edition at half the board power: a power-capped bin for slim chassis, not a cut-down memory system. It sells for the same price, too: $16,000.00 on NVIDIA's own store. Compute throughput at 300 W is lower, and this catalog holds no measurement of by how much. Carries no NVLink: the datasheet lists PCIe 5.0 x16 as the whole of its system interface, and the Server Edition's product brief states 'NVIDIA NVLink: Not supported' outright for the same silicon. NVLink left this product line two generations ago — the RTX A6000 had a bridge for two cards at 112 GB/s, the RTX 6000 Ada that replaced it did not — so a second card in the same box talks over the host bus at 128 GB/s, not over a fabric.
Power: 0.3 kW is the card alone, the Max-Q's 300 W board power. The 0.08 kW host allowance is a measured desktop overhead: a study that metered an RTX 5080 workstation at the wall beside the GPU's own reported power found the difference at 68-80 W across its ordinary ML and HPC workloads (arXiv 2606.19262), and the allowance takes the top of that range.
4× RTX PRO 6000 Blackwell Max-Q
NVIDIA · node
- Accelerators
- 4 × RTX PRO 6000 Blackwell Max-Q Workstation Edition (GB202)
- Architecture
- Blackwell sm_120/sm_121
- Memory
- 384 GB
- Bandwidth
- 7.2 TB/s
- Accelerator interconnect
- 0.128 TB/s
- Power
- 1.28 kW: 1.2 kW for the 4 accelerators and a 0.08 kW allowance for the host
- Acquisition
- $64,000
- Deployment multiplier
- ×1.25
Four times NVIDIA's own $16,000.00 for the Max-Q card on its US store, read on 2026-09-25. Nobody sells this build as a product, so there is no build price to quote. The 1.25 multiplier stands for the host workstation the four cards need and is an assumption, not a configured quote.
Four is the count NVIDIA itself names: the Max-Q datasheet says "Scale up to four RTX PRO 6000 Max-Q GPUs", and the 600 W Workstation Edition datasheet makes no such claim, which is why this build is the 300 W part. Memory and bandwidth are four times the card’s own published figures, and since 2026-09-25 so is the price: four cards at NVIDIA's own $16,000.00, where it was four times a range of retailer quotes, $30,692 to $35,600. The grade is `assumption` for the 1.25 host multiplier, the one figure in this entry this repository chose. Power is the four cards' board power plus a host allowance, which its own record cites. The H100 page is cited for one thing only, and a review was right that nothing said so: NVIDIA prints "NVLink: 900GB/s | PCIe Gen5: 128GB/s" in one spec table, which fixes both the figure and the convention it is on, since the 900 is the bidirectional total this catalog already stores for H100 SXM. The A6000 page dates the loss of NVLink from the line.
Power: 1.2 kW is the four cards alone, four times the Max-Q's 300 W board power. The 0.08 kW host allowance is a measured desktop overhead: a study that metered an RTX 5080 workstation at the wall beside the GPU's own reported power found the difference at 68-80 W across its ordinary ML and HPC workloads (arXiv 2606.19262), and the allowance takes the top of that range. That machine held one card; a host for four carries a larger CPU and more fans, so 0.08 kW is a floor for this build, and no measurement of a four-card host was found.
RTX PRO 6000 Blackwell Server Edition
NVIDIA · accelerator
- Accelerators
- 1 × RTX PRO 6000 Blackwell Server Edition (GB202, passively cooled)
- Architecture
- Blackwell sm_120/sm_121
- Memory
- 96 GB
- Bandwidth
- 1.6 TB/s
- Power
- 1.175 kW: 0.6 kW for the accelerator and a 0.575 kW allowance for the host
- Acquisition
- not published
- Deployment multiplier
- ×1.2
Sold through server OEMs only; no public price.
The rack card carries the same 96 GB GDDR7 as the two workstation editions but publishes 1,597 GB/s against their 1,792 GB/s — a lower memory clock, stated on both NVIDIA's own page and an OEM product guide, so it is not a typo on one page. Decode is bound by that figure, which makes the server part the slowest of the three despite the datacenter label. Carries no NVLink: the datasheet lists PCIe 5.0 x16 as the whole of its system interface, and the Server Edition's product brief states 'NVIDIA NVLink: Not supported' outright for the same silicon. NVLink left this product line two generations ago — the RTX A6000 had a bridge for two cards at 112 GB/s, the RTX 6000 Ada that replaced it did not — so a second card in the same box talks over the host bus at 128 GB/s, not over a fabric.
Power: 0.6 kW is the card alone, its 600 W maximum. The card runs in eight-slot servers, so the 0.575 kW host allowance is one eighth of the host share of the eight-GPU system whose accelerators are closest to 600 W. NVIDIA's DGX H200 datasheet gives 10.2 kW maximum system power for its standard configuration, and its eight H200s are rated "Up to 700W" each on NVIDIA's H200 page: 5.6 kW of accelerators, which leaves 4.6 kW for the CPUs, memory, NVSwitch chips, networking, fans and power-supply losses.
NVIDIA · accelerator
- Accelerators
- 1 × RTX 6000 Ada Generation (AD102)
- Architecture
- Ada Lovelace
- Memory
- 48 GB
- Bandwidth
- 0.96 TB/s
- Power
- 0.38 kW: 0.3 kW for the accelerator and a 0.08 kW allowance for the host
- Acquisition
- $6,800
- Deployment multiplier
- ×1
Reseller listing. NVIDIA's own store, which prices both RTX PRO 6000 editions, does not list this card.
Previous generation and still sold new. It is here because "the 96 GB one or the 48 GB one" describes two generations rather than two configurations of one card: 96 GB is Blackwell, 48 GB is Ada, and the Ada part also has barely half the memory bandwidth.
Power: 0.3 kW is the card alone, its 300 W maximum. The 0.08 kW host allowance is a measured desktop overhead: a study that metered an RTX 5080 workstation at the wall beside the GPU's own reported power found the difference at 68-80 W across its ordinary ML and HPC workloads (arXiv 2606.19262), and the allowance takes the top of that range.
NVIDIA · accelerator
- Accelerators
- 1 × GeForce RTX 5090 (GB202, consumer)
- Architecture
- Blackwell sm_120/sm_121
- Memory
- 32 GB
- Bandwidth
- 1.8 TB/s
- Power
- 0.655 kW: 0.575 kW for the accelerator and a 0.08 kW allowance for the host
- Acquisition
- not published
- Reported range
- $1,999 – $4,000
- Deployment multiplier
- ×1
$1,999 is the MSRP; the upper bound is the September 2026 street price, which has stayed well above MSRP through the same memory shortage that moved the RTX PRO 6000.
Same GB202 die and the same 1,792 GB/s as the RTX PRO 6000, with a third of the memory. That pairing is why it is listed: on a consumer card it is capacity that runs out first, not bandwidth. NVIDIA's own page rendered neither price nor bandwidth in this session's fetch, so both are corroborated rather than read off the vendor.
Power: 0.575 kW is the card alone, its 575 W total graphics power. The 0.08 kW host allowance is a measured desktop overhead: a study that metered an RTX 5080 workstation at the wall beside the GPU's own reported power found the difference at 68-80 W across its ordinary ML and HPC workloads (arXiv 2606.19262), and the allowance takes the top of that range.
NVIDIA · node
- Accelerators
- 1 × GB10 Grace Blackwell Superchip (unified CPU+GPU package)
- Architecture
- Blackwell sm_120/sm_121
- Memory
- 128 GB
- Bandwidth
- 0.273 TB/s
- Power
- 0.2 kW: the whole system at full load
- Acquisition
- $4,699
- Deployment multiplier
- ×1
NVIDIA published price, September 2026. It opened at $3,999.
128 GB of unified LPDDR5x buys capacity no discrete desktop card offers, at 273 GB/s — about a seventh of an RTX PRO 6000. Decode is memory-bound, so this box fits models it cannot serve quickly, and the catalog holds no throughput measurement for it. Power is a measured full-load draw, recorded with its source in its own record.
Power: ServeTheHome measured the whole box at "just under 200W" with its CPU and GPU both loaded, and never reached 240 W; many LLM inference workloads ran at 60-90 W. 0.2 kW is that full-load figure, the basis the other systems here are charged at. It replaces 0.24 kW, the rating of the 240 W USB-PD supply, which bounds the draw rather than measuring it. The GB10 chip's own TDP is 140 W.
AMD · accelerator
- Accelerators
- 1 × Radeon AI PRO R9700 (Navi 48, RDNA 4)
- Architecture
- RDNA 4
- Memory
- 32 GB
- Bandwidth
- 0.64 TB/s
- Power
- 0.38 kW: 0.3 kW for the accelerator and a 0.08 kW allowance for the host
- Acquisition
- $1,299
- Deployment multiplier
- ×1
AMD retail price at launch; partner cards run $1,244 to $1,329.
32 GB GDDR6 at 640 GB/s and 300 W, from AMD's own product page. The price is press reporting of the launch figure, not a line on that page.
Power: 0.3 kW is the card alone, its 300 W total board power. The 0.08 kW host allowance is a measured desktop overhead: a study that metered an RTX 5080 workstation at the wall beside the GPU's own reported power found the difference at 68-80 W across its ordinary ML and HPC workloads (arXiv 2606.19262), and the allowance takes the top of that range.
Ryzen AI Halo (Ryzen AI Max+ 395, 128 GB)
AMD · node
- Accelerators
- 1 × Ryzen AI Max+ 395 (unified CPU+GPU+NPU, Radeon 8060S with 40 RDNA 3.5 CUs)
- Architecture
- RDNA 3.5
- Memory
- 128 GB
- Bandwidth
- 0.256 TB/s
- Power
- 0.15 kW: the whole system at full load
- Acquisition
- $3,999
- Deployment multiplier
- ×1
- Available from
- 2026-07-06
AMD's retail price, on its Ryzen AI Halo page on 2026-09-26: "Retail price for AMD Ryzen AI Halo is $3999", for the 128 GB box with a 2 TB SSD. Sold through Micro Center, which AMD calls its first global launch partner.
128 GB of LPDDR5x at 8,000 MT/s and 256 GB/s, from AMD's specification page, which names the GPU as a Radeon 8060S with 40 RDNA 3.5 compute units. The memory is unified, and AMD's developer software sets how much of it the GPU may use; AMD's pages give no ceiling for that on this processor, so the entry records all 128 GB, as the other unified-memory entries do. AMD promised pre-orders for June 2026 on 2026-05-20 and said on 2026-07-06 that the systems "are now available at Micro Center", the date recorded here. A Halo with the Ryzen AI Max+ PRO 495 and 192 GB is announced with no price or date and is not this entry. No throughput profile here was measured on this system, so it carries no curated rate and needs a manual one.
Power: AMD publishes no measured draw for this box. 0.15 kW is its own planning figure: the cost model on its Ryzen AI Halo page assumes a "150W sustained" draw for the whole system, which it calls the "nightmare case". The specification page gives a TDP of 120 W, which is the processor's alone, so it is not the figure charged. Graded an assumption because AMD states it as one, not as a measurement or a specification.
Apple · node
- Accelerators
- 1 × Apple M6 (unified CPU+GPU+NPU, 12-core GPU)
- Architecture
- Apple M6
- Memory
- 32 GB
- Bandwidth
- 0.17 TB/s
- Power
- 0.07 kW: the whole system at full load
- Acquisition
- $1,799
- Deployment multiplier
- ×1
- Available from
- 2026-09-22
Apple's own price for this configuration, with 32 GB and a 1 TB SSD, on its US store on 2026-09-25. Apple's $899 starting price buys 16 GB, whose memory runs at 153 GB/s rather than 170.
Apple states 153 GB/s for the M6 with 16 GB and 170 GB/s once it is configured to 24 or 32 GB, and 32 GB is the most an M6 takes. Orders opened on 2026-08-25 and deliveries began on 2026-09-22. No throughput profile here was measured on Metal, so this system carries no curated rate and needs a manual one.
Power: Apple measures Mac mini at the wall, power-supply losses included, while "a compute-intensive test application" maximizes processor usage. Its one M6 figure, 70 W, was measured with 16 GB and a 256 GB SSD. It publishes none for this 32 GB, 1 TB build, whose memory also runs faster, 170 GB/s against 153, so 0.07 kW carries that measurement across and errs low.
Apple · node
- Accelerators
- 1 × Apple M5 Pro (unified CPU+GPU+NPU, 20-core GPU bin)
- Architecture
- Apple M5
- Memory
- 64 GB
- Bandwidth
- 0.307 TB/s
- Power
- 0.145 kW: the whole system at full load
- Acquisition
- $3,199
- Deployment multiplier
- ×1
- Available from
- 2026-09-22
Apple's own price for this configuration, with the 18-core CPU, the 20-core GPU, 64 GB and a 1 TB SSD, on its US store on 2026-09-25. The $1,699 headline buys the 15-core CPU and 16-core GPU with 24 GB.
Apple states 307 GB/s for the M5 Pro and lists no other figure for its 20-core GPU bin, and 64 GB is the most an M5 Pro takes. Orders opened on 2026-08-25 and deliveries began on 2026-09-22. No throughput profile here was measured on Metal, so this system carries no curated rate and needs a manual one.
Power: Apple measures Mac mini at the wall, power-supply losses included, while "a compute-intensive test application" maximizes processor usage. 145 W is its measurement of an M5 Pro with 64 GB, taken with an 8 TB SSD. The page does not say which of the two GPU bins it measured, so the confidence is medium rather than high. The figure sits under the 155 W that Apple's specifications give as every Mac mini's maximum continuous power.
Mac Studio (M5 Max, 128 GB)
Apple · node
- Accelerators
- 1 × Apple M5 Max (unified CPU+GPU+NPU, 40-core GPU bin)
- Architecture
- Apple M5
- Memory
- 128 GB
- Bandwidth
- 0.614 TB/s
- Power
- 0.2 kW: the whole system at full load
- Acquisition
- $5,099
- Deployment multiplier
- ×1
- Available from
- 2026-09-22
Apple's own price for this configuration, with the 18-core CPU, the 40-core GPU, 128 GB and a 512 GB SSD, on its US store on 2026-09-25. The $2,499 headline buys the 36 GB, 460 GB/s base machine, which is a different computer.
Apple publishes two M5 Max bandwidths, 460 GB/s for the 32-core GPU and 614 GB/s for the 40-core, and sells more than 36 GB only with the 40-core, so a 128 GB M5 Max is always the faster bin. Every throughput profile in this catalog was measured on a CUDA or ROCm serving stack and none of them transfers to Metal, so this system carries no curated rate and needs a manual one.
Power: Apple measures Mac Studio at the wall, power-supply losses included, while "a compute-intensive test application" maximizes processor usage: 200 W for the M5 Max with the 32-core GPU and 36 GB, and 385 W for the M5 Ultra with the 80-core GPU and 512 GB. It publishes no figure for this 40-core, 128 GB configuration, so 0.2 kW carries the 32-core measurement across; more GPU cores draw more at full load, so it errs low. It replaces 0.48 kW, the supply's maximum continuous rating, which bounds the draw rather than measuring it.
Mac Studio (M5 Ultra, 256 GB)
Apple · node
- Accelerators
- 1 × Apple M5 Ultra (unified CPU+GPU+NPU, 80-core GPU bin)
- Architecture
- Apple M5
- Memory
- 256 GB
- Bandwidth
- 1.2 TB/s
- Power
- 0.385 kW: the whole system at full load
- Acquisition
- $10,799
- Deployment multiplier
- ×1
- Available from
- 2026-09-22
Apple's own price for this configuration, with the 36-core CPU, the 80-core GPU, 256 GB and a 1 TB SSD, on its US store on 2026-09-25. The $5,499 headline buys the 64-core GPU with 96 GB, which cannot be configured to 256 GB.
Apple sells 256 GB and 512 GB only with the 80-core GPU, and states 1.2 TB/s for both M5 Ultra bins. Orders opened on 2026-08-25 and deliveries began on 2026-09-22. As on the M5 Max, no throughput profile here was measured on Metal, so this system carries no curated rate and needs a manual one.
Power: Apple measures Mac Studio at the wall, power-supply losses included, while "a compute-intensive test application" maximizes processor usage. Its one M5 Ultra figure, 385 W, was measured on the same 80-core GPU with 512 GB and a 16 TB SSD. It publishes none for this 256 GB, 1 TB build, so 0.385 kW carries that measurement across; half the memory and a smaller SSD draw less, so it errs high, against ownership.
Mac Studio (M5 Ultra, 512 GB)
announced Apple · node
- Accelerators
- 1 × Apple M5 Ultra (unified CPU+GPU+NPU, 80-core GPU bin)
- Architecture
- Apple M5
- Memory
- 512 GB
- Bandwidth
- 1.2 TB/s
- Power
- 0.385 kW: the whole system at full load
- Acquisition
- not published
- Deployment multiplier
- ×1
- Expected
- 2026-10-21 to 2026-10-31
Apple has published no price for this configuration: its store lists the 512 GB option with none beside it. The 256 GB build of the same chip is $10,799.
The largest unified memory Apple sells, on the same 80-core GPU and 1.2 TB/s as the 256 GB build; only the memory differs. It is an entry of its own because the machine is on sale and this configuration is not: one entry marked `announced` would say something false about the 256 GB build. As on every Mac Studio here, no throughput profile was measured on Metal, so it needs a manual rate.
Power: Apple measures Mac Studio at the wall, power-supply losses included, while "a compute-intensive test application" maximizes processor usage. 385 W is its measurement of this configuration, the 80-core GPU with 512 GB, taken with a 16 TB SSD.
Release: on 2026-08-25
Apple said “Mac Studio with 512GB of unified memory is coming in late October.”
“Late October” is read as the month's last third, the 21st to the 31st. Apple's store still said “512GB memory option for M5 Ultra coming late October” on 2026-09-25.