HPE Introduces 64 TB Memory Server

[erek]

“The server scales modularly to meet the rising demands of real-time analytics and high-velocity transactions. Available from a minimum of four sockets, up to 16 - and supporting up to 64 TB of DDR5 memory - the server enables customers to run financial services applications, databases, and systems of record. A dedicated external node controller in HPE Compute Scale-up Server 3250 offers a performance boost that is 100 times faster than Ethernet used in scale-out deployments. This is HPE's first scale-up server powered by Intel Xeon 6 processors and features a modular architecture well-suited for heavy compute and emerging agentic AI workloads.

"Our collaboration with HPE continues to blaze the trail for mission‑critical, in‑memory computing," said Kevork Kechichian, executive vice president and general manager, Data Center Group at Intel. "The HPE Compute Scale‑Up Server 3250, powered by Intel Xeon 6 processors, brings together exceptional compute density, scalability, and reliability to help enterprises run their most critical data‑intensive workloads."

Multi-layered security and resiliency
Built-in protection for HPE Compute Scale-up Server 3250 is available through HPE Integrated Lights Out (iLO) and spans every layer - from the chip to the cloud - as well as the entire lifecycle of the server. HPE iLO establishes a silicon root of trust through a dedicated security processor and validated firmware, which safeguards against future threats with post-quantum cryptography. Resiliency features of the HPE Compute Scale-up Server 3250 include advanced memory error detection and correction, as well as memory healing and deconfiguration. Additionally, the server is designed for fault-tolerant uptime, delivering the highest availability for critical business workloads.

HPE leadership for RISE with SAP-certified servers
HPE Compute Scale-up Server 3250 is the successor to the Intel-based HPE Scale-up Server 3200. HPE was recently named as a leader in the IDC MarketScape: Worldwide SAP HANA Certified Servers Appliances 2026 Vendor Assessment. HPE's long-standing collaboration with SAP and sustained engineering investment has positioned HPE as a leading provider of RISE with SAP-certified servers. HPE Compute Scale-up Server 3250 builds on that reputation with record-setting benchmark performance for large, in-memory configurations.

Availability
The HPE Compute Scale-up Server 3250 is available now.

This server can be acquired through HPE Financial Service's 90/9 Advantage program that offers no payments for 90 days and an additional 9 months at one percent.“

Source: https://www.techpowerup.com/348932/...r-for-critical-and-complex-business-workloads

 

[DanNeely]

HP prices this monstrosity at "Submit Your Request for a Custom Quote"; but on newegg 1TB of DDR5 in 4 RDIMMS is $23k, so this is probably around $1.5M in ram costs alone. and 9 months financing at 1% still probably grosses HP finance as much as I've spent on PCs in my lifetime.

 

[sleepeeg3]

Yes, but will it run Crysis?

This server can be acquired through HPE Financial Service's 90/9 Advantage program that offers no payments for 90 days

I guess I just have to try it for 90 days to find out!

 

[toast0]

This server can be acquired through HPE Financial Service's 90/9 Advantage program that offers no payments for 90 days and an additional 9 months at one percent.“

I hope they vet their customers... But then again, they can probably sell the ram for 1.5x when they repo the server.

On topic, if you have to use 16-socket servers, you've either done something terribly wrong or terribly right... or both. Horizontally scaling to 16 single socket servers with 4TB ram each is probably significantly less money, probably significantly less hassle, likely more compute throughput, easier to source, etc. But you have to do some work to partition your system.

It's nice to know there's options that go this big, though.

 

[uOpt]

On topic, if you have to use 16-socket servers, you've either done something terribly wrong or terribly right... or both. Horizontally scaling to 16 single socket servers with 4TB ram each is probably significantly less money, probably significantly less hassle, likely more compute throughput, easier to source, etc. But you have to do some work to partition your system.

It's nice to know there's options that go this big, though.

Well, if your application is strictly multithreaded and can't easily live with memory distributed across machines this is a good solution.

 

[T_A]

I hope they vet their customers... But then again, they can probably sell the ram for 1.5x when they repo the server.

On topic, if you have to use 16-socket servers, you've either done something terribly wrong or terribly right... or both. Horizontally scaling to 16 single socket servers with 4TB ram each is probably significantly less money, probably significantly less hassle, likely more compute throughput, easier to source, etc. But you have to do some work to partition your system.

It's nice to know there's options that go this big, though.

Not really , depends on the use case , we have 10 Lenovo SR930 V3 (8U) servers with 8 Proccessors and 16TB of ram each for SAP HANA DBs , you cant have a 6TB DB for example fit a 4TB server
the CPUs btw barely do anything load wise , its mostly a requirement to address so so much RAM

 

[cjcox]

 

[DanNeely]

I hope they vet their customers... But then again, they can probably sell the ram for 1.5x when they repo the server.

On topic, if you have to use 16-socket servers, you've either done something terribly wrong or terribly right... or both. Horizontally scaling to 16 single socket servers with 4TB ram each is probably significantly less money, probably significantly less hassle, likely more compute throughput, easier to source, etc. But you have to do some work to partition your system.

It's nice to know there's options that go this big, though.

I never got that high, but one project I was on scaled from a $150/mo AWS database server to a $800/mo one (I think 8 core, 32GB with the next two upgrade tiers being $2400 and $10k/mo (2x cores/ram on a better base server config, and another 50% but most of the cost going to licensing for a forced upgrade to SQL Server Enterprise not the far cheaper SQL Server Web ) because our customer refused to let us waste time on performance issues when she needed 6 days worth of new features added evrey week.

That happened because every time we ran into a performance issue (same two issues every time) we were told not to work on it and to just upgrade to the next larger size and ignore the two hot path queries that required most of the DB to be kept in ram one of which was doing a non-indexed search over the biggest table (we knew what the two biggest problems were, we were never able to ninja fixes into the code base although we were getting close on the bigger fix).

I'm not sure if it was the sticker shock, or realization that going *LA*LA*LA*LA*LA*LA*LA*LA*I*CANT*HEAR*YOU* every time we tried to bring up addressing the few places we badly missed performance wasn't going to be an option much longer; but the customer finally agreed to let us work on reducing DB load and create the load test setup that she somehow convinced herself we had from the start despite our never saying a word about it for the prior 4 or 5 years. I ended up transitioning off the project about that time so I never saw the details; but via gossip heard the first change (the missing index) got ~3x increased load, and a variety of other changes farther improved it to around 8 or 10x before I left.

 

[1_rick]

a non-indexed search over the biggest table

sigh.

 

[DanNeely]

sigh.

oh, it gets dumber.

Rant warning:

After looking at SQL profiling data it only took about 1 day to ID it as one of the 2 queries killing the DB and ~90-99% of the time spent in the web server call, generate an index, and micro-bench the change. Search went from something like 100ms at the time (and steadily getting worse as the table grew) to about 1ms. Insert was unchanged (I think only an ms). Quick easy win, you'd think. Nope!

The tech lead on the project had been on a prior project where some loon put indexes on every column and then a custom index for every single query in the application. That created a read only database that curled up and died at any insert, and a traumatized cow-orker who was fanatically opposed to ever having indexes on anything but the primary key.

He blocked the fix for about 2 years saying the microbenches couldn't be trusted, and was terrified under load the index inserts might magically take 100 or 1000x as long and make things worse. He'd only accept the PR if I wrote a full load test system to prove it was safe; knowing full well there was no chance that could happen.

Asking the client about deploying an "experimental" performance change and being ready to roll it back was also completely unacceptable for some reason.

Things remained stuck for 2 years until management decided to have the companies most senior devs give a few hours/month of 'consulting time' on all the other long term projects. He, being a smart and sane person looked at it and immediately agreed it was a good fix and my cow-orker was being irrationally paranoid.

By that point there was a code freeze because the client was about to do something very important (that would put peak load on the system). The table having gotten bigger, and the load being higher than before; resulted in the DB load going to 100%, faceplanting, and bringing the site down.

Sometimes I think not doing a "I told you so" dance and free form curseout rap that day may be one of my greatest professional achievements. /s but also not really.