F5F Stay Refreshed Hardware Desktop HBM vs DDR......conclusion?

HBM vs DDR......conclusion?

HBM vs DDR......conclusion?

A
ADIR_4444
Senior Member
417
09-17-2023, 12:12 AM
#1
After reviewing everything about HBM and DDR along with their advantages and disadvantages, I believe the main challenge keeping HBM from widespread adoption is indeed its latency problem. While cost is a factor, it’s expected to decrease as it becomes more common. The primary concern remains speed. Are you accurately pinpointing the current problem with HBM implementation?
A
ADIR_4444
09-17-2023, 12:12 AM #1

After reviewing everything about HBM and DDR along with their advantages and disadvantages, I believe the main challenge keeping HBM from widespread adoption is indeed its latency problem. While cost is a factor, it’s expected to decrease as it becomes more common. The primary concern remains speed. Are you accurately pinpointing the current problem with HBM implementation?

P
Person_
Member
126
09-17-2023, 04:09 AM
#2
No, it's not just about delay. HBM is built for wide buses (1024 bits wide, 2048 bits is planned) meaning a DDR5-style device using HBM would require around 3000-4000 contacts instead of the 288 contacts found in standard DDR5 modules. The electrical signals are very low power, so components need to be placed very near the memory controller. This is why we used HBM alongside a GPU on top of an interposer chip—this setup allows tight connections through multiple layers of copper traces inside the silicon die. A regular HBM module wouldn't fit well on a typical DDR5 stick because the distance would be too great, causing signal distortion. The interposer also adds expense; when Fury cards were produced, that basic interposer still cost about $30 to make. This price factor contributes to the overall cost, and HBM is created by layering several chips on top of a base die with controllers, drivers, and stacked dies to achieve higher capacities like 8, 16, or 32 GB. Often during stacking, some dies fail or don't work properly, so manufacturers sometimes replace entire stacks with fewer but functional ones, keeping prices up.
P
Person_
09-17-2023, 04:09 AM #2

No, it's not just about delay. HBM is built for wide buses (1024 bits wide, 2048 bits is planned) meaning a DDR5-style device using HBM would require around 3000-4000 contacts instead of the 288 contacts found in standard DDR5 modules. The electrical signals are very low power, so components need to be placed very near the memory controller. This is why we used HBM alongside a GPU on top of an interposer chip—this setup allows tight connections through multiple layers of copper traces inside the silicon die. A regular HBM module wouldn't fit well on a typical DDR5 stick because the distance would be too great, causing signal distortion. The interposer also adds expense; when Fury cards were produced, that basic interposer still cost about $30 to make. This price factor contributes to the overall cost, and HBM is created by layering several chips on top of a base die with controllers, drivers, and stacked dies to achieve higher capacities like 8, 16, or 32 GB. Often during stacking, some dies fail or don't work properly, so manufacturers sometimes replace entire stacks with fewer but functional ones, keeping prices up.

_
_digiboy
Member
196
09-18-2023, 11:03 AM
#3
Since you put this in the CPU, Motherboards, and Memory sub-forum, I'm going to assume you're talking about HBM for CPUs, which is something that has never been seen on the consumer market. In addition to the cost, which matters a lot for consumer electronics and should not be hand-waved away, HBM typically goes on the same package as the CPU. This reduces configurablity, which is very important for both DIY and OEM markets in the consumer space. Imagine if you had to choose how much RAM you want when you buy a CPU. The 14900K comes in a 128GB, 64GB, and 32GB version. And the cost of the memory also needs to be built into the cost of the CPU. So suddenly the top-of-the-line i9 and R9 CPUs cost $1200 because they include HBM memory. You now rely on Intel and AMD to give you reasonable memory configurations with their processors. If you need 128GB for video editing, but your software doesn't scale beyond 6 cores, you may just have to buy an i9 anyway because it's the only one that has 128GB on package. Now, in the server space, HBM can be used as a sort of caching layer, rather than as the only memory option, so it could be that you get both, but including an additional memory controller to handle the possibility of also wanting DDR5 memory on top will again add cost. The socket size is going to be giant, so motherboard prices would go up even more, and they have the engineering headache of figuring out how to fit the socket and RAM slots on a standard ATX board width. ITX would almost certainly only be able to use the on package HBM. And all of this is to get more bandwidth on a consumer platform where half the CPUs won't even benefit from it. You only see HBM included on CPUs with many cores, but Intel still even sells 2 core CPUs without HT. Imagine buying a Celeron with HBM. There's no way it will cost $50. So no, it's not just about latency.
_
_digiboy
09-18-2023, 11:03 AM #3

Since you put this in the CPU, Motherboards, and Memory sub-forum, I'm going to assume you're talking about HBM for CPUs, which is something that has never been seen on the consumer market. In addition to the cost, which matters a lot for consumer electronics and should not be hand-waved away, HBM typically goes on the same package as the CPU. This reduces configurablity, which is very important for both DIY and OEM markets in the consumer space. Imagine if you had to choose how much RAM you want when you buy a CPU. The 14900K comes in a 128GB, 64GB, and 32GB version. And the cost of the memory also needs to be built into the cost of the CPU. So suddenly the top-of-the-line i9 and R9 CPUs cost $1200 because they include HBM memory. You now rely on Intel and AMD to give you reasonable memory configurations with their processors. If you need 128GB for video editing, but your software doesn't scale beyond 6 cores, you may just have to buy an i9 anyway because it's the only one that has 128GB on package. Now, in the server space, HBM can be used as a sort of caching layer, rather than as the only memory option, so it could be that you get both, but including an additional memory controller to handle the possibility of also wanting DDR5 memory on top will again add cost. The socket size is going to be giant, so motherboard prices would go up even more, and they have the engineering headache of figuring out how to fit the socket and RAM slots on a standard ATX board width. ITX would almost certainly only be able to use the on package HBM. And all of this is to get more bandwidth on a consumer platform where half the CPUs won't even benefit from it. You only see HBM included on CPUs with many cores, but Intel still even sells 2 core CPUs without HT. Imagine buying a Celeron with HBM. There's no way it will cost $50. So no, it's not just about latency.

M
MikaEirik
Junior Member
22
09-20-2023, 12:31 PM
#4
Alternatively, using RAM on a chip could enable much more compact designs. This brings to mind the concept of integrated memory in computers—pretty cool, though most people believe it's still out of reach (though laptops already handle it perfectly!)
M
MikaEirik
09-20-2023, 12:31 PM #4

Alternatively, using RAM on a chip could enable much more compact designs. This brings to mind the concept of integrated memory in computers—pretty cool, though most people believe it's still out of reach (though laptops already handle it perfectly!)

N
Novaran
Member
62
09-20-2023, 02:19 PM
#5
It offers space efficiency benefits, which is why many laptops and certain mini PCs attach RAM directly to the board. It also provides lower latency than socketed RAM. But regarding flexibility, it would be very poor. Upgrading RAM would require purchasing an entirely new CPU or motherboard, a typical mid-life change for numerous devices.
N
Novaran
09-20-2023, 02:19 PM #5

It offers space efficiency benefits, which is why many laptops and certain mini PCs attach RAM directly to the board. It also provides lower latency than socketed RAM. But regarding flexibility, it would be very poor. Upgrading RAM would require purchasing an entirely new CPU or motherboard, a typical mid-life change for numerous devices.

Z
ZoloKu
Member
206
09-20-2023, 04:01 PM
#6
Sure thing... upgrading options, cost, and flexibility are all things that might not appeal to regular users, but for me... this is exactly what I need. A compact, low-latency setup that handles games without hassle—no complicated restrictions, no paywalls, and no tough modding limits. And if you're keen, we can add a GPU right away! In short, a simple console experience without all the usual baggage. *P.S.: Upgrades are still possible, but less flexible.*
Z
ZoloKu
09-20-2023, 04:01 PM #6

Sure thing... upgrading options, cost, and flexibility are all things that might not appeal to regular users, but for me... this is exactly what I need. A compact, low-latency setup that handles games without hassle—no complicated restrictions, no paywalls, and no tough modding limits. And if you're keen, we can add a GPU right away! In short, a simple console experience without all the usual baggage. *P.S.: Upgrades are still possible, but less flexible.*