I suppose you could make a case for heat-pipes, but these heat-spreaders have no fluid from which to transfer heat and do not match the definition presented.
I have been thinking about this for quite some time, It will be interesting when Anandtech does a comparison between Optane and the best of SSD. But given some preliminary data, and other testing of RAM Disk, it seems we are now bottlenecked by the Single Thread CPU performance. Or CPU Clock Speed. Since clock speed ceiling is pretty much set with TDP and unless we change to another material, we are going to get ~4Ghz only.
Exactly what CPU are you running? My Haswell is clocked at 4.6 24/7. I would assume the latest Intel Core i7-7740X can do 5Ghz without a lot of trouble.
I think he was talking a rough estimate of stock clocks as 5GHz is not new territory to overclockers. Though, it does come at the expense of more TDP and less processor life. I think you missed his point, though. He's suggesting that these drives are limited by single threaded performance and we are quickly approaching the limit of what we can obtain on silicon. This may or may not be true, but Intel has given us no reason to believe otherwise. Clock speeds haven't increased all that much since the P4 era which took us from an initial maximum clock speed of 1.5GHz (stock) to 3.8GHz (stock) / 5GHz (OC). Afterwards, IPCs moved up pretty steadily until Sandybridge. Since then, we've had a slow down in generational IPC improvement.
@iwod: "But given some preliminary data, and other testing of RAM Disk, it seems we are now bottlenecked by the Single Thread CPU performance. Or CPU Clock Speed. "
If storage performance is in fact be Single Threaded performance limited under the current data transfer model, then we should start thinking about how to parallelize the workload. Flash based storage devices are using parallelism to obtain their throughput numbers to begin with. There is no reason to believe that this parallelism can not be extended to the processing elements. Flash devices already use controllers with multiple cores. It is conceivable that if a single CPU core isn't fast enough to issue commands, then multiple CPU cores could issue commands directly to individual controller cores. Think a CPU core per Controller core per flash module. This would take a rework of the current storage model, though.
However, I suspect that there are less disruptive solutions and I'm not entirely convinced that single threaded CPU performance is the limitation. In straight read scenarios it is entirely possible to saturate a 4xPCIe 3.0 bus with the fastest SSDs. Random read/write performance seems to be limited by the controller and flash itself. Intel's consumer Optane SSD only uses a 2xPCIe 3.0 bus and the Optane SSD DC P4800X doesn't quite max out its 4xPCIe 3.0 bus, though it does hit its max throughput at lower queue depths than traditional SSDs. In any case, the biggest advantage to Optane is the latency improvements.
While we've obviously hit a wall with CPU clock speeds, I don't think that means we've hit a wall with CPU performance. For example: I have two laptops, a 3 year old quad-core 2.5-3.5 GHz AMD A10-5750M and a brand new dual-core 2.9-3.5 GHz Core i7-7500U.
They basically have the same clock speed, but the AMD has double the cores. However, in a variety of benchmarks, I have found that the Intel gets about 1.5 to 2 times the performance - both single- and multi-threaded - than the AMD.
So that means that with the same clock speed, over a 3 year period, Intel has managed to get double the performance with half the cores, simply with greater efficiency. (Granted, I'm comparing two different architectures, but it's what I've got.)
Of course, this is hardly the same as we used to get. When I was a child, we were getting a doubling of clock speed every 18 months. My first computer was a 266 MHz Pentium II, and about 3 years later, the 1 GHz Pentium III was released.
So we're not getting those kinds of huge leaps in performance from pure clock speed anymore. But Intel and AMD are still obviously managing to find ways to achieve quite sizable gains from efficiency instead.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
9 Comments
Back to Article
DanNeely - Friday, July 14, 2017 - link
WTH did tacky heatsinks get promoted to radiators?No. It's not a radiator unless it's in a water loop.
thetuna - Friday, July 14, 2017 - link
Ehh, I mean, passive heat sinks are still radiators.BurntMyBacon - Monday, July 17, 2017 - link
Radiatora: any of various devices (such as a series of pipes or tubes) for transferring heat from a fluid within to an area or object outside
https://www.merriam-webster.com/dictionary/radiato...
I suppose you could make a case for heat-pipes, but these heat-spreaders have no fluid from which to transfer heat and do not match the definition presented.
willis936 - Friday, July 14, 2017 - link
Well that's just wrong.iwod - Friday, July 14, 2017 - link
I have been thinking about this for quite some time, It will be interesting when Anandtech does a comparison between Optane and the best of SSD. But given some preliminary data, and other testing of RAM Disk, it seems we are now bottlenecked by the Single Thread CPU performance. Or CPU Clock Speed. Since clock speed ceiling is pretty much set with TDP and unless we change to another material, we are going to get ~4Ghz only.bigboxes - Friday, July 14, 2017 - link
Exactly what CPU are you running? My Haswell is clocked at 4.6 24/7. I would assume the latest Intel Core i7-7740X can do 5Ghz without a lot of trouble.BurntMyBacon - Monday, July 17, 2017 - link
I think he was talking a rough estimate of stock clocks as 5GHz is not new territory to overclockers. Though, it does come at the expense of more TDP and less processor life. I think you missed his point, though. He's suggesting that these drives are limited by single threaded performance and we are quickly approaching the limit of what we can obtain on silicon. This may or may not be true, but Intel has given us no reason to believe otherwise. Clock speeds haven't increased all that much since the P4 era which took us from an initial maximum clock speed of 1.5GHz (stock) to 3.8GHz (stock) / 5GHz (OC). Afterwards, IPCs moved up pretty steadily until Sandybridge. Since then, we've had a slow down in generational IPC improvement.BurntMyBacon - Monday, July 17, 2017 - link
@iwod: "But given some preliminary data, and other testing of RAM Disk, it seems we are now bottlenecked by the Single Thread CPU performance. Or CPU Clock Speed. "If storage performance is in fact be Single Threaded performance limited under the current data transfer model, then we should start thinking about how to parallelize the workload. Flash based storage devices are using parallelism to obtain their throughput numbers to begin with. There is no reason to believe that this parallelism can not be extended to the processing elements. Flash devices already use controllers with multiple cores. It is conceivable that if a single CPU core isn't fast enough to issue commands, then multiple CPU cores could issue commands directly to individual controller cores. Think a CPU core per Controller core per flash module. This would take a rework of the current storage model, though.
However, I suspect that there are less disruptive solutions and I'm not entirely convinced that single threaded CPU performance is the limitation. In straight read scenarios it is entirely possible to saturate a 4xPCIe 3.0 bus with the fastest SSDs. Random read/write performance seems to be limited by the controller and flash itself. Intel's consumer Optane SSD only uses a 2xPCIe 3.0 bus and the Optane SSD DC P4800X doesn't quite max out its 4xPCIe 3.0 bus, though it does hit its max throughput at lower queue depths than traditional SSDs. In any case, the biggest advantage to Optane is the latency improvements.
Mikewind Dale - Thursday, July 20, 2017 - link
While we've obviously hit a wall with CPU clock speeds, I don't think that means we've hit a wall with CPU performance. For example: I have two laptops, a 3 year old quad-core 2.5-3.5 GHz AMD A10-5750M and a brand new dual-core 2.9-3.5 GHz Core i7-7500U.They basically have the same clock speed, but the AMD has double the cores. However, in a variety of benchmarks, I have found that the Intel gets about 1.5 to 2 times the performance - both single- and multi-threaded - than the AMD.
So that means that with the same clock speed, over a 3 year period, Intel has managed to get double the performance with half the cores, simply with greater efficiency. (Granted, I'm comparing two different architectures, but it's what I've got.)
Of course, this is hardly the same as we used to get. When I was a child, we were getting a doubling of clock speed every 18 months. My first computer was a 266 MHz Pentium II, and about 3 years later, the 1 GHz Pentium III was released.
So we're not getting those kinds of huge leaps in performance from pure clock speed anymore. But Intel and AMD are still obviously managing to find ways to achieve quite sizable gains from efficiency instead.