F5F Stay Refreshed Hardware Desktop Not satisfied with Intel's speed, could adding more RAM improve it?

Not satisfied with Intel's speed, could adding more RAM improve it?

Not satisfied with Intel's speed, could adding more RAM improve it?

Pages (2): Previous 1 2
S
soccerblock11
Junior Member
21
08-02-2021, 05:34 AM
#11
I'm pretty sure HBM is better than DDR5, it's integrated memory after all. I would assume that just like any other type of workload, more RAM is needed if you're running out of it, otherwise it would mildly slow down the process. Found a Reddit post:
S
soccerblock11
08-02-2021, 05:34 AM #11

I'm pretty sure HBM is better than DDR5, it's integrated memory after all. I would assume that just like any other type of workload, more RAM is needed if you're running out of it, otherwise it would mildly slow down the process. Found a Reddit post:

T
tamemarco
Senior Member
482
08-24-2021, 05:22 AM
#12
The 9480s were likely surplus parts for the Aurora supercomputer, now available at under $1,000 each. Many collectors have already purchased them, driving prices up by 150-250% above standard values.
T
tamemarco
08-24-2021, 05:22 AM #12

The 9480s were likely surplus parts for the Aurora supercomputer, now available at under $1,000 each. Many collectors have already purchased them, driving prices up by 150-250% above standard values.

X
xttra
Junior Member
40
08-24-2021, 06:37 AM
#13
It offers very high bandwidth, yet many performance tests indicate extremely high latency.
X
xttra
08-24-2021, 06:37 AM #13

It offers very high bandwidth, yet many performance tests indicate extremely high latency.

S
soos50
Junior Member
22
08-24-2021, 05:19 PM
#14
Interesting, I hadn't realized how severe the problem was. It somewhat clarifies why it was included on consumer GPUs. How extensive is this project you're developing? If latency is indeed the main concern, you might consider using DDR5 and letting software decide between HBM and DDR5 pool in flat mode. From the video:
S
soos50
08-24-2021, 05:19 PM #14

Interesting, I hadn't realized how severe the problem was. It somewhat clarifies why it was included on consumer GPUs. How extensive is this project you're developing? If latency is indeed the main concern, you might consider using DDR5 and letting software decide between HBM and DDR5 pool in flat mode. From the video:

L
LoveKnights
Member
82
08-31-2021, 02:26 PM
#15
The system features around 128 channels using DDR4-3200MT/s (HBM2e is more aligned with DDR4 than DDR5) across the two CPUs. So far, I've tested CFD and CFD-EM benchmarks from L1Techs, which have memory sizes of about 10GB and 60GB respectively. These amounts seem limited for full parallel utilization, though I expected better performance compared to dual Epycs. My main workload is expected to reach around 120GB, which might yield noticeable gains with bigger tasks. This is the focus I'm aiming for, but I'm hesitant to invest heavily in memory and doubt a significant speed boost. Additionally, I'm uncertain how my application will manage NUMA nodes when no processors are connected to them, especially in flat mode.
L
LoveKnights
08-31-2021, 02:26 PM #15

The system features around 128 channels using DDR4-3200MT/s (HBM2e is more aligned with DDR4 than DDR5) across the two CPUs. So far, I've tested CFD and CFD-EM benchmarks from L1Techs, which have memory sizes of about 10GB and 60GB respectively. These amounts seem limited for full parallel utilization, though I expected better performance compared to dual Epycs. My main workload is expected to reach around 120GB, which might yield noticeable gains with bigger tasks. This is the focus I'm aiming for, but I'm hesitant to invest heavily in memory and doubt a significant speed boost. Additionally, I'm uncertain how my application will manage NUMA nodes when no processors are connected to them, especially in flat mode.

F
Fa837241
Member
100
09-22-2021, 03:03 AM
#16
No, the temperatures, power usage and clock speeds were not within normal ranges throughout the tests.
F
Fa837241
09-22-2021, 03:03 AM #16

No, the temperatures, power usage and clock speeds were not within normal ranges throughout the tests.

N
newcake
Junior Member
4
09-29-2021, 07:44 PM
#17
I was really worried about overheating since the HBM is quite temperature-sensitive. I chose dual 360 rads—two 180mm fans each—to keep things cool. I can maintain the CPU below 60°C even when running at full power. My software is configured to use only 64 cores at once, which maximizes performance. Clock speeds will swing between 3.2GHz and 3.5GHz based on workload demands, and power consumption will be quite inconsistent, fluctuating from 600 to 1100 watts during the session.
N
newcake
09-29-2021, 07:44 PM #17

I was really worried about overheating since the HBM is quite temperature-sensitive. I chose dual 360 rads—two 180mm fans each—to keep things cool. I can maintain the CPU below 60°C even when running at full power. My software is configured to use only 64 cores at once, which maximizes performance. Clock speeds will swing between 3.2GHz and 3.5GHz based on workload demands, and power consumption will be quite inconsistent, fluctuating from 600 to 1100 watts during the session.

E
Echo_Runner
Member
215
10-01-2021, 11:14 AM
#18
Great update, it looks like the math library issue was resolved by moving to the latest OneAPI version. Performance improved noticeably on the 60GB test, now matching the speed of the fastest dual socket EPYC system. It makes sense given the recent release of Sapphire Rapids in 2023. I'm interested in testing AMD's math library next, though I'm a bit concerned about HBM support and AMD's software development pace compared to Intel.
E
Echo_Runner
10-01-2021, 11:14 AM #18

Great update, it looks like the math library issue was resolved by moving to the latest OneAPI version. Performance improved noticeably on the 60GB test, now matching the speed of the fastest dual socket EPYC system. It makes sense given the recent release of Sapphire Rapids in 2023. I'm interested in testing AMD's math library next, though I'm a bit concerned about HBM support and AMD's software development pace compared to Intel.

B
Bigking
Member
154
10-02-2021, 12:05 AM
#19
I ended up purchasing the DDR5 RAM and it actually shows lower latency compared to HBM memory (~105ns vs ~120ns). ^^ Keep in mind the non-square matrix since NUMA nodes 8-15 are the HBM stacks and don’t have CPU connections to them. Also, my previous assumption was incorrect—the OneAPI update wasn’t responsible; the real reason for the performance boost was a different default damping factor in the solver.
B
Bigking
10-02-2021, 12:05 AM #19

I ended up purchasing the DDR5 RAM and it actually shows lower latency compared to HBM memory (~105ns vs ~120ns). ^^ Keep in mind the non-square matrix since NUMA nodes 8-15 are the HBM stacks and don’t have CPU connections to them. Also, my previous assumption was incorrect—the OneAPI update wasn’t responsible; the real reason for the performance boost was a different default damping factor in the solver.

Pages (2): Previous 1 2