• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

Teamgroup announces worlds fastest SSD over PCIE 5.0 with read speeds at 13,000MB/S and write speeds at 12,000MB/S

CuNi

Member
Ya, I wouldn't want to play it off an HDD. It's still a bit less than I thought as it's 'only' twice more than Doom Eternal does in a similar amount of time.

The most IO intensive thing I've tried yet was actually the Matrix demo (edit: the open world part) which read ~100GB over the span of 10 minutes. That's also a pure streaming test while my R&C test above includes cutscenes etc.

Not intending to downplay the significance of consoles getting SSDs, but that example is terrible.
Console SSDs do the biggest work when they need to read lots of data in a short time. Accumulating reads over time undermines their work.
185GB in 30 minutes is a average of 103MB/s and 100GB in 10 Minutes is roughly 167MB/s.

If you google average HDD speeds, it spits out between 80MB/s to 160MB/s depending on HDD obviously. If you even settle in the middle, you have 120MB/s, which is enough to satisfy R&C demands (averaged out).
I don't know when most reads happen, for RC I'd assume it's when jumping rifts, loading a save etc. but those would only take longer on HDDs than SSD.
I also can't say how the Matrix demo handles reads, if it does constantly, in bursts etc. but as you can see it's only barely above average and probably within higher end HDD read speeds.
That's not including HDD caches for repeated read calls etc.

Until we ever get a game that was build with SSDs in mind that also releases on PC, we will never be able to tell how any game behaves when put on HDD.
RC could have micro stutter as it waits for data, or it could simple have pop in, or it would only load longer and have higher RAM usage to push constantly called data in there. Who knows.
 
MS promoted the Series SSD as sustained
Which was the point of the last paragraph. There's no pc ssd either pcie ver 3.0 or 4.0 that can sustain speeds like that without throttling due to heat and the cooling in a pc tends to be better than a console, which is why I'm doubtful the xbox can sustain a speed of 2.4 GB/s indefinitely when pc never could. Playstation in no way, shape or form is an exception either.
 
Last edited:

sinnergy

Member
Which was the point of the last paragraph. There's no pc ssd either pcie ver 3.0 or 4.0 that can sustain speeds like that without throttling due to heat and the cooling in a pc tends to be better than a console, which is why I'm doubtful the xbox can sustain a speed of 2.4 GB/s indefinitely when pc never could. Playstation in no way, shape or form is an exception either.
I don’t know , it’s what they advertised.. Sony talked not about sustained and MS did adjust cooling ..
 
I don’t know , it’s what they advertised.. Sony talked not about sustained and MS did adjust cooling ..
If neither demonstrated their ssds to prove they can hold onto those speeds indefinitely, especially Microsoft, who advertised it as sustained. I see no reason to believe that statement, since there's no solid ground to base that statement on. No benchmark, no test run, no stress test, nothing. Sony's can't run at that speed indefinitely that's for sure.
 
You are missing the point. If an SSD - like some people speculate in this thread - is to act as a slow RAM pool for asset streaming, you need to bring latency down with a huge margin (modern SSDs are good enough to act like slow RAM from a pure specification point of view). Direct Storage is an attempt to address this but will not in its current iteration completely succeed with this due to CPU overhead (still a very large improvement over today's situation).

You're only basing this on assumptions but where are the numbers to back up this belief that latency is still such an issue, even with NVMe 2.1 and PCIe 5.0, even with SAM/BAR, even with improving NAND modules increasing not just bandwidths but also lowering latency, even with improving DRAM caches and FMCs...that the latency figures won't be "good enough"? Even considering improvements in file I/O systems for OSes like Windows 11 and Linux distros, you just seem to be plainly ignoring all of this.

CPU speeds are only getting faster, with better integration of caches at last-level hierarchies. All of these things drive down the latency cost, let alone standards that can layer on top of PCIe like CXL which bring even lower latency alongside other benefits when talking about CXL 2.0 and especially 3.0. There's little more preventing a decently spec'd consumer PC (let alone higher-end PC systems or enterprise server clusters) from utilizing SSDs as a slow RAM pool for asset streaming than there is for the PS5. It's not the miracle wunderkid of I/O far out ahead of the class that you've been telling yourself it is. That doesn't mean it's bad, not by any means. But it's not in its own league, so to speak, when taking a look at the wider market of today.

To achieve this, the GPU would need to be allowed to call a file on the SSD and directly read it into VRAM without CPU involvement. This is technologically possible on a PC but will not be allowed for multiple security reasons in Windows.

It already exists with things like GPUDirectStorage on Nvidia GPUs. Also there are other OSes besides Windows :/

So to paraphrase your text: Yes, yes, yes - Direct Storage is a CPU run driver to control traffic faster to the VRAM pool but will still suffer from a CPU tax in terms of latency due to that every I/O step still being 100% under CPU control. This is the EXACT same reason why you need VRAM on your graphics card on a PC. Yes - you have all that super fast RAM on your motherboard and a very fast PCIe bus but that RAM and bus is under CPU control that introduces very high levels of latency that necessities VRAM on your card so you can cut the CPU out of the core graphics I/O cycle.

Again, if the CPU in question has enough overhead, this "tax" means virtually nothing, and CPUs for the PC space alone are only getting better and better. Once you get into the enterprise, cloud etc. space, then it's literally a non-issue; implementations in those spaces can handle I/O throughput magnitudes higher than the consoles simply on spare CPU cores and threads.

Also again, you're forgetting about SAM/BAR, which was introduced not that long ago to help streamline addressing schemes between the CPU and GPU.

This fundamental bottle-neck is not in place for next-generation consoles.

The consoles have their own bottleneck of sorts: shared usage of the bus enforced by bus arbitration. That's part of what comes with current hUMA memory architectures. Standards like CXL and OMI especially bring hUMA benefits to nUMA memory architectures with equivalent (or in case of OMI, superior) low latency than conventional hUMA memory designs, but I'll be fair and admit those things are not in the consumer PC space (gaming or otherwise), so don't bare as much mention here.

And for those that think this is console warring against PC: I play more games on my high-end PC than my console. This is simply about technology and how to evolve the PC platform.

Except the PC is already evolving and has addressed quite a few of the things (or are in the process of addressing quite a few of the things) you still seem to think are keeping it in the stone ages compared to consoles. And actually, I think the taste of some exaggeration that can maybe be picked up from your insistence on this console I/O vs. PC I/O discussion is what's more the issue, at least with me.

Truthfully speaking consoles have never had any technological advantage over PC gaming for more than a year or two, so the idea of consumer PC I/O (for systems geared towards gaming) taking the lead over console implementations by this point isn't far-fetched. I remember you saying in the past you felt that consoles (specifically PS5) would stay ahead on this for the generation, or at least for a good few years. That isn't happening.

PC always leading the charge of bleeding edge technology. I remember getting my first ssd and installing Witcher 2 on it and couldn’t believe how fast it loaded.

ms just needs to hurry up with direct storage.

It's supposed to be coming this year, but probably in H2 at this rate.
 

CamHostage

Member
Sounds cool, although personally, I want to see capacity/price improvements (and stability, if possible) over speed right now. $100/TB has me checking my wallet whenever I consider upgrading, I'd like to see that come down. Speed sounds great, but so far, the speeds we have are well enough and we'll see if any next-gen tech comes along that pushes read-speed requirements to breakneck delivery like this.
 
How much impact will DirectStorageAPI with SFS have on PCIE 5.0 NvME SSD's with 3-D Nand Controllers with read/write speeds of 10GBps-12.4GBps?
 

Hoddi

Member
How much impact will DirectStorageAPI with SFS have on PCIE 5.0 NvME SSD's with 3-D Nand Controllers with read/write speeds of 10GBps-12.4GBps?
Data rates scale with rendering resolution when using SFS. Running a game at 4k will typically read ~4x more data from disk than when running at 1080p, for example. But, even then, you're only looking at a few hundred megabytes vs maybe a gigabyte per second using the Intel SFS demo.

These SSD data rates are otherwise just marketing wank for the most part. Microsoft recently released the documentation which states that
The main design goal for DirectStorage is to allow a title to sustain 50K IOPS and only use between 5 percent and 10 percent of a single CPU core. This allows the title to achieve maximum bandwidth from the NVMe storage subsystem, while at the same time allowing the CPU to be used for other title requirements.
And 50K IOPS at 64kbyte per block equals 3.2GB/s.
 
Last edited:

Irobot82

Member
Data rates scale with rendering resolution when using SFS. Running a game at 4k will typically read ~4x more data from disk than when running at 1080p, for example. But, even then, you're only looking at a few hundred megabytes vs maybe a gigabyte per second using the Intel SFS demo.

These SSD data rates are otherwise just marketing wank for the most part. Microsoft recently released the documentation which states that

And 50K IOPS at 64kbyte per block equals 3.2GB/s.
What GDDR6 data rate?

Is this quick enough to keep textures loaded on the SDD and only pull them into RAM when needed? Or is latency I suppose still an issue.
 

Hoddi

Member
What GDDR6 data rate?

Is this quick enough to keep textures loaded on the SDD and only pull them into RAM when needed? Or is latency I suppose still an issue.
I wouldn't really compare this to graphics memory. SSDs are great at reading data but you can't really perform a lot of writes to them without killing the disk in a matter of days/weeks. Textures are only small part of the equation.
 
Last edited:

Hoddi

Member
What GDDR6 data rate?

Is this quick enough to keep textures loaded on the SDD and only pull them into RAM when needed? Or is latency I suppose still an issue.
It was late last night and I think I misread your post a bit. This demo doesn't really push VRAM bandwidth since SFS doesn't load whole textures at a time. Here's an Nsight capture showing how short the time period is.

I otherwise notice little pop-in unless I run it at 8k and flick the mouse around. This can exceed my 1.8GB/s SSD so some textures are delayed by a half second or so.
 

Kuranghi

Gold Member
Who's out here necro-reacting? Thank you good sir/madam.

What all dis bidness then? Are we approaching the singularity?
 
TEAMGROUP T-FORCE Debuts the Latest PCIe Gen5 SSD: Leaping Into a New Era of SSDs with Groundbreaking Technologies
TEAMGROUP is releasing its first PCIe Gen5 SSD under its gaming sub-brand T-FORCE as part of the CARDEA series. It is capable of maximum sequential read speeds of over 13,000MB/s and write speeds exceeding 12,000MB/s, and together with a maximum storage capacity of 4TB, it will be the highest performance PCIe Gen5 flagship SSD on the market when launched



Hot damn! I dont think the current gen consoles can benefit from these speeds. It is all for PC baby!
And now comes the most important question...whats the price??
 

poppabk

Member
Secondly general purpose CPUs in no way compete with custom designed hardware meant to perform a specific task. That's why specialized hardware exists. That's why we have GPUs for crying out loud, unless you want to still live in a world of software based rasterizers.
Yup that is why consoles are so sub par relative to PCs. No PPU for running physics. My Physx card is so awesome, you won't believe the physics it pumps out.
 
Top Bottom