Connect with us

NEWS

AMD Venice Sells Flash Memory Relief With 256 Cores

EPYC 9006 Venice ships on SP7 with 16 memory channels and MEXT flash-as-DRAM, a rack play aimed at soaring server memory bills.

Published

on

AMD began shipping Zen 6 EPYC 9006 Venice chips on November 10, 2026, into a DRAM market Counterpoint Research put at nearly $200 billion in the third quarter, up 369% on only about 4% more bits. The flagship EPYC 9996 lists at $14,904 for 256 cores and 512 threads, yet the features that change a rack bill are 16 memory channels, flash-backed MEXT, and new power controls, not the core count on the box.

Corporate VP of Server Product Madhu Rangarajan wrote that EPYC 9006 SP7 platforms are shipping from OEM and ODM partners, with Oracle Cloud Infrastructure E7 instances still to come and Microsoft Azure instances aimed at model work. SP8, Venice-X, and Verano stay on 2027 calendars. The silicon that left factories this month is the dense, 16-channel half of a four-family plan.

Venice Is Shipping Into a Memory Famine

Counterpoint’s blended DRAM selling price sat at $1.90 per gigabit in that third-quarter tally, which is why a CPU launch now reads like a memory product. AMD’s own IT note cites a September TrendForce forecast of another 10 to 15% quarterly DRAM increase for conventional contracts in the fourth quarter. Hasmukh Ranjan, AMD’s chief information officer, put the bind in one line: memory is crowding out other kit on the same budget.

THE DRAM BILL IN LATE 2026

  • Q3 revenue: Global DRAM sales reached nearly $200 billion, per Counterpoint Research.
  • The mix: Revenue jumped 369% year over year while bit shipments rose about 4%.
  • The unit price: The blended average selling price reached $1.90 per gigabit.
  • The next step: TrendForce’s September note pointed to a 10 to 15% contract rise in the fourth quarter.

That is the market the 9996 walks into at 600W default CPU power, configurable from 400W to 600W, with 1,024 MB of L3 and a 2.55 GHz base that will sit closer to real jobs than its 4.1 GHz boost. A dual-socket box holds 512 cores and 1,024 threads. The frequency partner in independent tests is the 96-core, 192-thread EPYC 9686F, a 500W part (400W to 600W) listed at $11,434 in AMD’s 1Ku sheet and boosting to 5 GHz. AMD’s shipping brief separately puts 96-core Venice parts at up to 5.3 GHz, a ceiling, not the SKU those tests used.

Thirty-one 9006 models span $700 for the eight-core EPYC 9016 to $14,904 for the 9996. Nine of them are SP7 parts. The other 22 wait on SP8 in the first half of 2027. The 192-core Turin EPYC 9965 launched at $14,813, so Venice’s top chip adds 64 cores for $91 and works out to about $58 per core.

Two Dies Buy 16 Channels and Spend PCIe Lanes

Zen 6 is not a clean-sheet core in the way AMD described Zen 5 in 2024. The frontend still uses dual 4-wide decode and 8-wide dispatch. AMD grew the op cache back to 6.75K entries (it can feed 18 micro-ops per cycle, up from 12), split the integer schedulers to 96 shared ALU/AGU plus 48 dedicated ALU entries, and enlarged the physical register file to 352 entries from 240. The vector path keeps a 512-bit data path, adds native AVX-512 FP16, and can take two 512-bit loads or four 256-bit loads. Store width is unchanged.

The CCD is the density jump. A standard Zen 6 CCD holds 12 cores and 48 MB of L3 (4 MB per core) with a 5.3 GHz ceiling on Venice. A Zen 6c CCD holds 32 cores and 128 MB of L3, double the cores and quadruple the cache of a Zen 5c CCD, with a 4.1 GHz ceiling. AMD fits up to eight CCDs on SP7, which is how a 256-core 6c part and a 96-core full-size part share one socket. Compute dies use TSMC’s 2nm N2P node. AMD called Venice the first HPC product to ramp on that process in a May 21, 2026 note from chair and CEO Lisa Su.

The platform change is the split I/O die. Two IODs sit in the middle of the package and give SP7 16 channels, MRDIMM-12800 at up to 1.6 TB/s, RDIMM-8000 at up to 1 TB/s, and 8 TB per socket. Independent tests of both Venice chips used 32 × 64 GB SK hynix DDR5-12800 ECC MRDIMMs, 2 TB, to fill those channels. SP7 is 76 × 80 mm, a step up from SP5 at about 72 × 75 mm.

Lane count is the bill. A 1P SP7 socket exposes 96 PCIe 6.0 lanes at 64 Gbps each, plus eight bonus PCIe 4.0 lanes; 2P drops the usable Gen 6 total to 128 because one 64G SerDes link ties the sockets. Turin offered 128 PCIe 5.0 lanes in 1P. Gen 6 doubles per-lane rate, so raw bytes can still rise while the number of devices you can hang off the CPU falls. SP8, the smaller 62.5 × 80 mm socket, flips the trade: eight channels, 819 GB/s on MRDIMM or 512 GB/s on RDIMM, and 128 PCIe 6.0 lanes in 1P (192 in 2P). Both sockets support CXL 3.1.

SP7, SP8 AND VERA AT A GLANCE

Processor Cores / threads (1P) Memory Peak bandwidth 1P PCIe 6.0 Default power
EPYC 9996 (SP7) 256 / 512 16-channel MRDIMM-12800 1.6 TB/s 96 600W
EPYC 9686F (SP7) 96 / 192 16-channel MRDIMM-12800 1.6 TB/s 96 500W
Nvidia Vera 88 / 176 8× SOCAMM2 LPDDR5X 1.2 TB/s 88 250W to 450W

Vera’s published architecture lists 88 Olympus cores and 1.2 TB/s of LPDDR5X on field-replaceable SOCAMM2, 1.5 TB per socket, a 164 MB system-level cache, and a monolithic mesh. Independent STREAM runs used 16 × 96 GB Samsung LPDDR5-9600, 1.5 TB. AMD will answer that package in the second half of 2027 with Verano, an LPDDR5X SOCAMM2 part, rather than on the SP7 boards shipping now.

What MEXT Does When DRAM Costs More Than the CPU

AMD bought MEXT in June 2026 and is using Venice as the commercial on-ramp, including HPE platforms marked MEXT ready. The software is a loadable Linux kernel module, an on-CPU AI engine, and a RAM View monitor. Cold pages move to NVMe flash. A predictor pulls pages back into DRAM before the application faults. The operating system still sees the full capacity. The engine runs on one core and does not need a GPU.

HOW MEXT TIERS A PAGE

  • Watch: The driver tracks which pages are hot and which have gone cold.
  • Offload: Cold pages leave DRAM for a dedicated NVMe pool that the OS still counts as memory.
  • Predict: The AI engine, a family of light heuristics plus neural nets, guesses which offloaded pages will be touched next.
  • Prefetch: Those pages return to DRAM before the request, so the app keeps running if the guess is early enough to hide flash latency.

Measured numbers sit in an October 2026 AMD white paper on Synopsys PrimeTime 2022.03-SP2, run on a 1P EPYC 9755 (Zen 5, 128 cores, SMT off) with a 512 GB DRAM-only baseline averaging 14,040 seconds. A 256 GB DRAM plus 256 GB MEXT pool averaged 14,516 seconds, 96.7% of baseline at half the DIMMs. 192 GB plus 315 GB MEXT averaged 15,182 seconds (92.5%). 128 GB plus 378 GB MEXT averaged 16,783 seconds (83.7%). AMD converts those to about 1.93×, 2.47×, and 3.35× performance per DRAM-dollar. A three-node Spark TPC-DS run on fifth-gen EPYC held 99% at a 50% DRAM split. AMD is clear the trick is workload-bound; jobs that thrash the working set will not give DIMMs back.

MEXT’s AI-driven approach gives us a viable way to address cost challenges by expanding usable memory without simply buying more DRAM.

Hasmukh Ranjan, Chief Information Officer, AMD

AMD IT started the production rollout in September 2026 on 86 engineering hosts and added about 58 TB of usable capacity. Buying that as DRAM would have cost about $3.3 million. The MEXT install cost about $601,000. MinQi Bao, senior director of systems design engineering, said the company counted nearly $2.78 million, or 82% of DRAM expansion costs. The same case study frames it as 20% more usable memory at 18% of equivalent DRAM spend. The internal plan is about 700 servers by March 2027 (around $21 million saved) and 2,000 by the end of 2027 (more than $60 million). That math, not the 256-core slide, is what a buyer facing $2,000-class 64 GB RDIMMs is being asked to underwrite.

FAST Splits One Socket Into Two Clock Classes

Unified Power Provisioning treats CPU and DIMM draw as one budget. AMD’s example moves a rack from 830W across 16 servers to 706W across 18, a 15% cut in provisioned power while adding two boxes. Utilization Based Power Scaling then tries to close the gap between that provisioned number and what the job actually burns.

Frequency Adaptive Spectrum Technology is the odd one. AMD describes two commercial SKUs inside one socket. On a 96-core Venice CPU, 32 cores can sit in a premium class at a 3.65 GHz base and 5 GHz max while 64 batch cores run at 2 GHz with a 2.4 GHz max. Cloud operators who sell a fast VM SKU and a dense SKU from the same heap of silicon have been stranding whichever mix they guessed wrong. FAST makes that split a firmware policy. It also makes a 96-core, 500W to 600W part look like two products in a capacity planner, which is the point.

The 400W to 600W socket is the complaint that follows the SPEC charts around. A dual-9996 box is a kilowatt-class CPU draw before DIMMs, fans, and GPUs. UPP and UBPS exist because AMD is selling more cores into racks that already cannot feed them. The power features are not a side dish. They are how a 256-core, 16-channel part fits the same floor tile.

Integer Leads Flip When STREAM Uses Spread

Independent SPEC CPU 2026 runs (estimates, not submitted to SPEC) used Ubuntu Server 26.04, Linux 7.0, GCC 16 on Venice and GCC 15.2 elsewhere, plus jemalloc. Throughput copies matched thread count. A 2P 9996 posted a 93% lead over a 2P EPYC 9965 in integer throughput, on 1,024 threads versus 768, plus the new core, the denser CCD, and 12,800 MT/s MRDIMMs. The 9686F, short of threads against a 128-core EPYC 9755, still led that Turin proxy by 19% and led a 2P Vera by 17% in integer throughput. Float throughput is where the vector work shows: the 9686F led Vera by 40%, and both Venice chips more than doubled the Zen 5 scores. Integer speed, one job owning the socket, put the 9686F 21% ahead of Vera.

THE PUBLISHED TEST LEADS

Test 9686F vs 2P Vera 9996 vs 2P 9965
SPEC CPU 2026 integer throughput 17% ahead 93% ahead
SPEC float throughput 40% ahead More than doubled
SPEC integer speed 21% ahead Second to the 9686F
PyBench 32% ahead 24% ahead
Java SciMark composite 92% ahead 24% ahead

Python still leans on a fast thread, so the 9686F’s 32% PyBench lead over Vera (the 9996’s 14%) matches the clocks. Memory is the split. STREAM with OpenMP true binding put AMD 11% to 19% ahead of Vera. Spread binding, pinning jobs far apart across the die, leaves AMD flat and lifts Vera, whose mesh is built for that pattern. MBW with a 128 MB array favors the 9996 pair; larger arrays swing back to Vera. Anyone quoting “Venice wins memory” is quoting true binding on 16 channels of 12800 MRDIMM against eight channels of LPDDR5X. Change the policy and the ranking moves.

AMD’s own SPECrate 2026 integer estimates, published September 18, put a 2P 9996 at 2.24 times a 2P Vera platform and a 96-core high-frequency Venice at about 1.2 times Vera per core. Those are vendor projections. The independent integer-throughput gap of 17% on the 9686F is the number that survived someone else’s compiler flags. Elsewhere the 9996 led the 9965 by 30% to 40% in Blender (35% on Barbershop), Java SciMark rose 24% generation on generation, and the 9686F’s 92% SciMark lead over Vera is the loudest single composite. CloverLeaf had the 9686F more than twice as fast as Vera. Compression, 7-Zip, OSPRay, LLVM, and simdjson follow the same shape: Venice ahead unless the test is a cold-cache mesh walk, where Vera’s topology still jumps first.

Ramping Venice on TSMC 2nm process technology marks an important step forward in accelerating the next generation of AI infrastructure.

Lisa Su, Chair and CEO, AMD, May 21, 2026

AMD quotes a 27% average IPC lift over Zen 5 in enterprise, cloud, and AI. That figure is a fixed-frequency comparison of 16 Zen 6c cores in one CCD against 16 Zen 5 cores split across two CCDs, so some of the “IPC” is the missing hop. It is not a desktop number, and it is not the 93% throughput gap, which multiplies cores, clocks, cache, and 12,800 MT/s memory on top of the core.

Nine SP7 Chips Now, 22 SP8 Parts in 2027

SP7 is the half that exists. SP8 is the half most enterprise and storage racks will actually buy, cheaper, eight-channel, richer in PCIe, and absent until the first half of 2027. Venice-X (9006X) adds 3D V-Cache on SP7, up to 1,152 MB of L3 on 96 Zen 6 cores, in the second half of 2027. Verano, the 9006 LP LPDDR5X SOCAMM2 host node, lands in the same window. Intel’s Diamond Rapids is also on a 2027 clock. The 9686F is AMD’s Vera fighter for host-node and HPC work until Verano ships; the 9996 is the throughput brick. Neither is the volume SKU on SP8.

THE 9006 CALENDAR

  1. May 21, 2026: AMD says Venice has begun its production ramp on TSMC 2nm.
  2. June 2026: AMD acquires MEXT and its flash-tiering software.
  3. July 23, 2026: The four-family 9006 plan (SP7, SP8, Venice-X, Verano) is laid out.
  4. September 2026: 1Ku prices for all 31 SKUs go up, and AMD IT starts the 86-host MEXT rollout.
  5. November 10, 2026: SP7 platforms ship from OEM and ODM partners.
  6. First half of 2027: SP8 processors are due, 22 of the 31 names.
  7. Second half of 2027: Venice-X with 3D V-Cache and Verano on LPDDR5X SOCAMM2 are due.

Consumer Zen 6, Olympic Ridge, is also a 2027 item. Copying the 27% enterprise IPC line onto that chip would be a guess AMD has not published.

Frequently Asked Questions

Does AMD MEXT Run on Older EPYC Generations?

Yes. AMD IT’s production rollout in September 2026 used fourth-gen and fifth-gen EPYC engineering servers, not 9006 silicon, which is why the PrimeTime white paper is an EPYC 9755 result. Venice is when AMD began packaging MEXT as a licensed EPYC feature with HPE “MEXT ready” systems; the kernel module itself is not tied to Zen 6.

How Large Is the SP7 Socket Compared With SP5 and SP8?

SP7 measures 76 × 80 mm. SP5, the Genoa and Turin socket, is about 72 × 75 mm. SP8 is smaller still at 62.5 × 80 mm, which is how AMD cuts channel count to eight while raising 1P PCIe 6.0 to 128 lanes on the enterprise board.

How Many Bonus PCIe 4.0 Lanes Sit Beside PCIe 6.0?

SP7 adds eight PCIe 4.0 lanes in 1P and 16 in 2P, carried on a smaller 16G SerDes, on top of the 96/128 PCIe 6.0 counts. SP8 also carries eight bonus Gen 4 lanes in 1P and 16 in 2P beside its 128/192 Gen 6 lanes. Those bonus lanes are the leftover path for older devices after the 64G interfaces take PCIe 6.0 and the 2P link.

Is the 27% Zen 6 IPC Gain Meant for Desktop Chips?

No. AMD’s 27% average is an enterprise, cloud, and AI figure at fixed frequency, and the 16-core comparison places all Zen 6c cores in one CCD while Zen 5 splits 16 cores across two. Olympic Ridge, the consumer Zen 6 part expected in 2027, has no published client IPC number, and AMD warned against treating the server average as a broad claim.

The 9996 can fill a dual-socket box with 1,024 threads and a 93% integer-throughput gap over last year’s top Turin. The invoice still starts with 32 MRDIMMs, a 600W default, and a software license that turns NVMe into the other half of memory. SP8, Verano, and Venice-X are 2027 problems. The DIMM bill is this quarter’s.

Harry is the editor and lead writer of STUDIO ONE NETWORKS, an independent title he owns and runs himself. Ten years in journalism, reporting first and editing later, taught him that entertainment and business are one beat seen from two sides: a box office figure is a company number, a streaming deal is a contract, a casting rumour is not a story until someone puts their name to it. He works from the record, whether that is a distributor's statement, a licensing agreement, an interview transcript or a set of published ratings, and checks every number against it before publication. The same rule holds for the rest of the site, which covers news, technology, science, sports, lifestyle, travel, auto and gaming for an audience spread across the world. When he gets something wrong, the article is corrected and the change is noted and dated, under a corrections policy anyone can read. Reader mail is answered by him at support@studioonenetworks.com.

Click to comment

Leave a Reply

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

Trending