Intel Launches Stratix 10 TX: Leveraging EMIB with 58G Transceivers | Hacker News Reader

Intel Launches Stratix 10 TX: Leveraging EMIB with 58G Transceivers

anandtech.com

45 pointsrbanffy21 comments

>Previously these were (and have been called) 56G, but Intel is using the term 58G because it believes that its solution will perform better than other 56G solutions on the market, enough to call it a 58G device.

G here should mean Gbps, it's a technical description not a marketing term. Either the transceivers can be configured for 29GHz PAM4, or they can't, I don't see how there's room for interpretation here.

"Ours goes up to eleven."

I had a server vendor selling "hot swap RAID drives", where you couldn't swap disks without a power down. They assured me "hot swap" is a marketing term. I don't trust any marketing material any more.

Where does anandtech get this from? The Intel press release very clearly states:

> up to 144 transceiver lanes with serial data rates of 1 to 58 Gbps

And the Altera website agrees: https://www.altera.com/solutions/technology/transceiver/over...

Hmm, I wonder if they initially designed it only for 56G but they found that it worked acceptably well up to 58G after tape-out.

Now, this gets interesting.

FPGA's usefulness as an accelerator in DCs was hugely hampered by IO bottlenecks. Even PCI-E can be said to be limiting the theoretical peak throughput of top FPGAs. With more IO being done by physical hardware, this should offset IO bottleneck somewhat.

I'm still waiting for FPGA on package Xeons that will have raw register to register linkage connection. They were once announced in 2016, some samples were shown, but then they went silent.

keep in mind that raw speed (in terms of say, ram access or PCI-E access (which is already faster than this implementation)) is rarely the obsticle these days, the issue is fetch/rtt latency, especially with trying to integrate FPGA acceleration into a codebase for offloading some operations, the penalty for for which can often be tens of thousands of cycles, or more.

Seems pretty desperate. It was known they were having a real hard time bringing up Altera's gigabit transceivers on Intel's processes. I guess they've given up and decided to just bond separate chips together.

Which isn't necessarily a bad solution....

Its a lot more expensive so I guess it depends on your definition of "bad".

I'm a neophyte to this area, but did I read this correctly, the stratix 10 SoC comes with an ARM cortex chip? Intel sells solutions with ARM chips?

ARM don't make chips. They design them. Then they license the designs to chip users, who in turn get chip manufacturers to make them. Intel make lots of ARM chips (as well as their own in-house designs).

I remember years ago (2010-ish?) Intel being claimed as the largest ARM customer at the time (as in, the entity paying the most license fees to ARM).

Intel only recently signed a manufacturing deal with ARM in 2016, I'm not sure how extensive it is though. I'm only aware of LG taking advantage of it until now.

Intel used to both design and manufacture their own line of ARM SOCs back in the 2000s under the XScale brand, but they sold it to Marvell in 2006.

Intel bought FPGA maker Altera a while back. That is how they acquired an FPGA line of products. Some Altera Stratix 10 do have ARM cores in them.

Yes, absolutely. AFAICT this is an Intel-branded design from the Altera acquisition. Pairing an FPGA with an ARM host CPU is a reasonably common thing.

It may be a surprise for you, but intel once had bridge chips with multiple arm cores working as audio codecs and ATA controllers back a decade ago.

They also just straight up sold ARM SoCs, with ARM cores as the main processor in their StrongARM days.

The Stratix 10 SX has an onboard AArch64 ARM core, correct. Previous systems used ARMv7. Most FPGAs with a built-in system-on-chip you'll see have ARM cores attached, as well, such as Xilinx boards. Intel isn't unfamiliar with selling ARM chips, but in this case -- Intel acquired Altera, who did this originally. You can find one-or-two FPGA SoCs from Intel with actual x86 chips instead, though.

But the idea is that the ARM cores normally just act as your "control plane" -- manage the FPGA, watch power usage, reload configurations, monitor & reporting metrics, etc etc. So they aren't the main attraction, of course.

In the long run, Intel will be moving to discrete PCIe FPGA cards for data center usage, as well as fusing the FPGAs directly on-chip, alongside the Xeon cores for extremely high-end stuff. That's where you'll find products like Stratix 10 (aside from a few other places.) So in datacenter-oriented offerings, you may not see the ARM cores floating around that much.

They did so for quite a while but somehow didn't really take off.

As I understood intel has the best FABs. About half the advantage over AMD lies purely in the production technology. ARM is said to be more energy efficient than x86. Yet, when intel makes ARM chips you don't hear them outperforming everything else. I wonder why.

This article made no sense at all. It could have made sense with a few strategically placed definitions.

Reply on news.ycombinator.com